This is a pretty sophisticated setup. I particularly like how it uses Tailscale.
I've been using the simpler but not as flexible alternative: I'm running Claude Code for web (Anthropic's version of Codex Cloud) via the Claude iPhone app, with an environment I created called "Everything" which allows all network access.
(This is moderately unsafe if you're working with private source code or environment variables containing API keys and other secrets, but most of my stuff is either open source or personal such that I don't care if the source code leaks.)
Anthropic run multiple ~21GB VMs for me on-demand to handle sessions that I start via the app. They don't charge anything extra for VM time which is nice.
I frequently have 2-3 separate Claude Code for web sessions running at once, often prompted from my phone, some of them started while I'm out walking the dog. Works really well!
I delayed adopting conductor because I had my own worktree + pr wrappers around cc but I tried it over the holidays and wow. The combination of claude + codex + conductor + cc on the web and claude in github can be so insanely productive.
I spend most of my time updating the memory files and reviewing code and just letting a ton of tasks run in parallel
I haven't missed planning mode myself. I tend to tell it "write a detailed plan first in a file called spec.md for me to review", then use that as the ongoing plan.
I like that it ends up in the repo as it means it survives compaction or lets me start a fresh session entirely.
I was doing the same, but recently I noticed that Claude now writes its plans to a markdown file somewhere nested in the ~/.claude/plans directory. It will carry a reference to it through compaction. Basically mimicking my own workflow!
This can be customized via a shell env variable that I cannot remember ATM.
The downside (upside?) is that the plan will not end up in your repo. Which sometimes I want. I love the native plan mode though.
The lack of Plan Mode is puzzling, I'm sure they must get to it at some point. But until then it CAN still plan, you just have to ask it to write a plan and not write code yet.
Not sure if this works in claude code web, but running non-interactive claude code I can still get it to use plan mode by simply asking it. It's just a tool call.
Can you not use PAL MCP for this? Have one top agent as controller etc? It's not ideal but it feels like the space of multi agent stuff is evolving ... I notice that there are a lot of posts on hn about these things so we are trying to do the same thing really.
I'm surprised to see people getting value from "web sandbox"-type setups, where you don't actually have access to the source code. Are folks really _that_ confident in LLMs as to entirely give up the ability to inspect the source code, or to interact with a running local instance of the service? Certainly that would be the ideal, but I'm surprised that confidence is currently running that high.
I still get the full source code back at the end, I tell it to include code it wrote in the PR.
I also wrote my own tool to extract and format the complete transcript, it gives me back things like this where I can see everything it did including files and scripts it didn't commit. Here's an example: https://gistpreview.github.io/?3a76a868095c989d159c226b7622b...
Oh fascinating - so you're reviewing "your own" code in-PR, rather than reviewing it before PR submission? I can see that working! Feels weird, but I can see it being a reasonable adaptation to these tools - thanks!
What about running services locally for manual testing/poking? Do you open ports on the Anthropic VM to serve the endpoints, or is manual testing not part of your workflow?
Yeah, I generally use PRs for anything a coding agent writes for me.
If something is too fiddly to test within the boundaries of a cloud coding agent I switch to my laptop. Claude Code for web has a "claude --teleport" command for this, or I'll sometimes just do a "gh pr checkout X" to get the branch locally.
Yeah the commits that claude code generate are co-authored by claude@anthropic.com so i just open a PR to see the code. I have automatic per-PR dev environments for manual testing.
These are all some excellent ideas, I need to setup these things asap since I've been going back and forth on having more homelab vs cloud providers, but I'm only hearing about tailscale right now so I got to go for it. Cloud providers all of sudden becoming costly just for my side projects and/or not providing the exact PR environments like I would like etc. I've been wasting so much time on trying to automate AI Agents vs cloud providers with limited conf.. It would be great if AI Agents can just write the config for all deployments, pipelines, standards, without me having to go to any UI to tweak things manually etc.
Even with GitHub CI now all of sudden it wasted $50 on few days of CI actions. Should have everything run on my home server. But I think I may need more powerful home server, I have a cheap Dell refurbished one now.
I don't want to ever have to touch a UI again (except in places like Hackernews or the like) and the ones I specially built (read: vibecoded) for myself.
Yes my primary motivation for putting so much effort into a self-hosted cloud was cost. Managed Kubernetes instances are very expensive. I've saved a ton of money hosting it myself for side projects. With the benefit that spending $2k on a framework desktop one time to use as a k8s node means I have a much, much larger cluster than I'd be willing to pay for on a month to month basis. It might pay for itself in a single month. It's my opinion that Kubernetes can do anything the clouds can, so I just run talos on the old PCs, the only thing they do is run Kubernetes. Cloud hosting is insanely expensive.
I do have a managed Kubernetes instance that I run public services on (like for webhooks from github) so I dont need to open my home ports. It's very small to keep costs low. The benefit of using Kubernetes at home is most of my configs need minor changes to work on the managed k8s instance, so there's not much duplicate work to get features/software deployed there. It's the great cloud agnosticator after all!
I've started my own web interface for Claude Code to host it in the same cluster. That's where the CI builds happen, the PR envs get deployed. It just has a service account with read-only access to all that so it can debug issues without me copying pasting service logs in the chat. Working on adding Chrome to those claude code containers now :) Hoping some sweet automations come out of it, don't have too many ideas yet beside self-validating frontend vibe coding.
Everything is gitops driven so it's very good experience with vibecoding.
Right - I’m missing how you get the source code in the OP. It says you tmux in with ssh agent forwarding for GH. But you can’t do that on your iOS device? So you have to set up all your repos in the morning before leaving the house, then collect and push all your branches when you return home?
I could imagine this working for a small number of branches/changes.
The output from Jules is a PR. And then it's a toss-up between "spot on, let's merge" and "nah, needs more work, I will check out the branch and fix it properly when I am the keyboard". And you see the current diff on the webpage while the agent is working.
Claude Code on the web, ChatGPT Codex and Google Jules are not the same as Claude, ChatGPT and Gemini. They are entire apps where you authorize Github access and they work via PRs.
Right, yes, that was precisely my point - it was weird to me that people were comfortable operating on a codebase that they don't have locally, that they can't directly interact with.
> it was weird to me that people were comfortable operating on a codebase that they don't have locally, that they can't directly interact with.
I have a project where I've made a rule that no code is written by humans. It's been fun! It's a good experience to learn how far even pre-Opus 4.5 agents can be pushed.
It's pretty clear to me that in 12 months time looking at the code will be the exception, not the rule.
when the agent pushes the PR, in a branch, you can switch to that branch locally on your machine and do whatever, review it, change it, and ask for extra modifications on top, squash it, rebase it
Are those VM specs documented anywhere, I have used Claude Code for web a lot and never really bothered with the details. Just connect it to my repo and let it cook
Check out superconductor.dev (I’m building it), if you want live app previews, docker-in-docker functionality, multiple agents in one mobile app, and more.
I have my very fast macbook pro at my desk in my office, and I use tmux and tailscale and git worktrees and I’ve built a notification setup like this author.
Thanks to tailscale and ssh I can vibecode on the go from my phone with this setup.
While it’s great to leave a task running, no matter what I do I can’t achieve the type of high quality work on the go that I can when I’m sitting at my desk.
For me working on a full SaaS.. I just can’t do quality work on my phone.
The only way I can do quality work is to sit at my desk where I’m focused on the work. To play with the result of a prompt, take copious notes, feed them back to the agent, not ship until the thing is polished to a shine. To feature flag the changes, review all code in excruciating detail as though it was written by a dyslexic intern, add all the metrics and logs one can think of (VictoriaMetrics), add user-behavior logging (Amplitude/Posthog) and monitor the thing like your livelihood depends on it. Because it’s a product and you have pride in your work.
All of that needs loads of screen real estate and a keyboard.
So I guess the more things change, the more they stay the same.
Same here, I’m vibecoding a toy project where I never looked at the code from my phone, but I always seat for work. I’m using happy app and that’s good enough for now, I have the desktop in tailscale but I access it that way just for testing
For me, setup like this is not a replacement or even supplement the desk + high focus environment. My use case for these mobile vibecode setup is great though for small exploration/POC/learning/research type of work. Then I take the knowledge learned or any actual useful parts of code and incorporate into my work next time I sit back down at my desk.
There has been a sustained campaign over the last few days to push "I use Claude from my phone". I saw multiple posts on LinkedIn already, and now this.
This blog is super sus too. All the posts are about Claude. I suspect it's run by Anthropic, just read the About page: https://granda.org/en/about/
I am not Anthropic. I have vibe-coded a web interface for claude code to use from my phone. It's very similar to Anthropic's, but has a couple more features that I wanted.
Pandora's box is open; we're moving towards a world where white collar workers will be working 24/7 and they'll be expected to do so.
It won't matter if I'm washing the dishes, walking the dog, driving to the supermarket, picking up my kids from school. I'll always be switched on, on my phone, continuously talking to an LLM, delivering questionable features and building meaningless products, destroying in the process the environment my kids are going to have to grow in.
I'm a heavy LLM user. On a daily basis, I find LLMs extremely useful both professionally and personally. But the cognitive dissonance I feel when I think about what this means over a longer time horizon is really painful.
This technology should be liberatory, and allow us all to work less while enjoying the same standard of living. We've all contributed in its development by creating the whole corpus of the internet, without which it could never have been bootstrapped.
The only reason we can't expect this is that we live under a system that is arranged for the sole benefit of the owners of capital, and have been convinced that this is an immutable state of affairs or that our own personal advantage can be found in making a Faustian bargain with it.
> we live under a system that is arranged for the sole benefit of the owners of capital, and have been convinced that this is an immutable state of affairs
Realistically, if you have 300M, you and your direct family are settled for life. So, I want to propose 1B cap on net worth, if its more than that for 12 months straight, surplus goes to government, if your net worth is down after that, government obliges to return it partially to make it to 1B.
People, who are eager building things and innovating, will keep building regardless, power hungry will try to find other ways to enrich themselves, but eventually they will give up (e.g. having 10 kids, each with 1B net worth)
So let's say we implemented this and the government suddenly receives billions of credits from wealthy individuals. Would anything actually change? Is this what's holding us back from repairing all the roads to a fine standard? From implementing universal basic income?
It seems to me that if one tried to actually spend those credits we'd simply get inflation. Prices for roads, food etc. would just go up.
We definitely want to work on inequality, but I think numbers above 1B net worth are just weird quirks of the system. Musk is powerful because he's powerful, the number is just a reflection of that. Keeping his number below some arbitrary threshold isn't going to combat his power.
We need to tackle the problem head on: we need to stop individuals from amassing so much power. We get lost in this stupid abstraction called money. It's not what matters.
This is so arbitrary and incredibly naive. How did you come to with 300M, why not 300k or 300 billion? How would you determine the worth of rare, illiquid or intangibles? What about wealth held in trusts or companies? How does the accounting work if I borrow against my wealth? What happens when things change value dramatically in a short period of time? And the government is going to "make billionaires whole again" if they crater their wealth?
You are asking me about implementation difficulty, difficult implementation doesn't mean idea is not worth it.
One example:
* 300k vs 300M - doesn't matter if I said 100M, 200M, 550M, if you think 300M is not enough for you and your family to afford anything, not sure how other people are surviving for even less.
Here is why I think this is good:
1. Ambitious people will still be ambitious, its rare some genius kid says: I know this is 100B idea, but I won't build it, because I will only own 1B of it.
2. Limits the power, when power is really limited, people will be forced to focus on different things. For example, if you had plans to take over the world by making $10T and creating an army to kidnap president of another state you don't like, then you would know, it is not possible to make 10T, its not only about how much, its about suppressing hungry animal in you by capping your limits.
3. There is a chance "bad" ambitious people, will be converted to real philanthropist, because they know it doesn't matter to own more than 1B anyway and they can't own it.
> You are asking me about implementation difficulty, difficult implementation doesn't mean idea is not worth it.
I can agree with that idea, to an extent. If something is near impossible (not saying this is), then it does become not worth it.
The other questions the parent posed are more interesting to me:
> How would you determine the worth of rare, illiquid or intangibles? What about wealth held in trusts or companies? How does the accounting work if I borrow against my wealth? What happens when things change value dramatically in a short period of time?
Another I wonder is that (ignore all specifics of the values, just the concepts matter here), let's say you own a private business that then becomes valued at 1.5 billion dollars and this individual has 20 million dollars liquid. How do you tax that? The government can't take one third of the business, at least not without a lot of issues (in business dealings and individual rights), and the 20 million liquid wouldn't come close to what this plan would value. What do we do then? Plenty of billionaires don't really have liquid cash and forcing liquidation of assents in such a way seems like it would be very difficult.
I'm all for more taxes on higher net worth individuals, but I think there's a lot of talk to be had on how one can implement this. It's going to be really difficult to find a way that makes sense.
> I can agree with that idea, to an extent. If something is near impossible (not saying this is), then it does become not worth it.
FATCA law makes this very possible in the US.
> How do you tax that? The government can't take one third of the business, at least not without a lot of issues (in business dealings and individual rights)
I would say that the government can and should and simply be a passive share holder with no voting rights.
Obviously, I do not know nitty gritty details of economy and finance, but if I would implement this tomorrow I would start taking the equivalent of surplus and start from there to understand more.
For example, say individual has 20M liquid cash, 2 houses each valued at 5M and 1.5B in company shares (based on averaged company value for the last 6 or 12 months):
* whatever you can immediately spend is prioritised first, so you keep your 20M + 2 houses, then surplus is $530M of your company shares
* this equivalent number of shares will be moved to government trust, individual doesn't have any control over it, if person dies next day, government keeps the money (lets simplify for now and keep voting rights as separate question)
* let's say after shares moved to gov. trust, during next 6 months company value halved, gov. returns all your shares, if stock dropped only 10%, you get equivalent back to make your net worth 1B
* regarding taxation, I would keep it as it is today and tax on "realization event"
There are around 3.000 billionaires in the world, even hiring 10 dedicated people for each billionaire to calculate all this stuff on a quarterly basis is not expensive
What do you think all that money you taxed from billionaires will do in the hands of the “government”, pay the bills for 12-24 months? Then what’s your plan after the government has spent it all? No amount of taxation can solve a spending problem.
1. Assuming Elon is curious person, he will still build it out of curiosity.
2. Assuming Elon is not curios person, just power hungry, he will probably think its not worth building it, but someone else who is curious will build it eventually. This is even better, because when power hungry person owns such thing, they might use it for bad things as well (e.g. to gain more power, eventually interfering with elections, oh wait, it did already happen)
3. Government will build, because government will have more money now, but then we should be even more careful who gets to the top. Assuming people won't have more than 1B, maybe there will be less lobbying? because its not worth as much as it was before?
honestly, I don't think it was money related, because look at James Webb telescope, I think NASA achieved lots of good things, IMO they could have built SpaceX as well, it was probably more motivation and organisational impotence and bureaucracy from management (including higher ups than NASA).
Ah, so your idea is the good old “only the emperor who controls the violence apparatus should have a lot of money and power”?
It’s not a very original idea, and it has been tried many times, and it failed many times.
> but then we should be even more careful who gets to the top
Right, so “for some reason only the greedy power hungry psychopaths get to the top in the current system — let’s fix it so that there can’t be many of them, only one government who has power to take away other people’s wealth and concentrate it immensely, surely we will figure out how to make sure it’s not filled with greedy power hungry psychopaths as we go”
Not forcing a scarcity of necessities like housing would be a start.
Peer competition is what makes everything work. You need scarcity of necessities to force people in to the system. Recent rulings allowing the criminalisation of homelessness are pushing this further. Your existence is default-illegal unless you work to outbid your peers for housing.
Claude Code already is the purple unicorn. We're already there - the only problem is that regulatory systems are set up in a way that benefits a small minority of capitalists, rather than the majority.
The same but less rigged would be a good start. I feel like people ask your question as a gotcha because they can't wrap their head around a system more nuanced than "cancerous capitalism" or "potato famine communism"
Something like we had in advanced western europe and the US between ww2 and the late 70s seemed much more balanced while not requiring a complete system change. Most people would be fine if we sprinkled a bit of socialism on top of the gigantic pile of capitalism.
Stuff like housing, energy, transportation, shouldn't make you live paycheck to paycheck forever. Just the fact that people are slowly starting to talk about 50 years mortgage should be a wake up call.
Most people would be happy knowing there is something a tiny bit better coming, rather than knowing they will never make it out and will kept getting fucked a tiny bit more year after year. My grandparents had objectively a harder life than mine, but their life was improving every year, mine is stagnating at best, and usually I'm losing purchasing power year after year, while being relatively well paid for my country
And it still mostly doesn't so it's not really a good argument. Meanwhile the share of women in the workforce more than doubled but households' purchasing power isn't even as high as back then when it comes to the basics like cars/energy/housing, if anything it's down quite a bit.
We should unite, not fight about who's whitest, XYZ gender or a minority, you'll always find someone who has it better or worse than you, what matter is the average is going down, while back then it was going up, the rest is mostly noise.
It's always the same: workers need to unionize and form a political power bloc. Then, those most impacted—the majority—have an array of options, which are well explored in the annals of leftist and socialist political theory.
This is not at all to say that more conservative or reactionary theorists are wrong about how the world works. In fact, I think they're usually more right about what's really going on abstractly.
But, the working man doesn't need to know what's really going on. They need to win the war, and there's a ton of tactical advice written down—hard won lessons by those who built the modern world through the labor movement.
The place to start is with the usual suspects. Verso Books, The New Centre for Social Research, histories of the labor movement, and new political commentators like Josh Citarella.
The numbers of hours that they work relative to the average wage laborer bears absolutely no relationship to whether they are the beneficiaries of an exploitative socio-economic arrangement. But they all certainly work much less than the laborers in the most precarious positions who are forced to work multiple gig economy jobs to make ends meet, yes.
It would be a deep irony if LLMs ended up ushering in the social rupture that never arrived in the industrial era. When the pigs turn hogs and refuse to share even the scraps, they shouldn’t be surprised if the system they depend on becomes their undoing.
We should all hope so. It's clear that mass surveillance, the vast psyops architecture including social media platforms, autonomous drone warfare, Starlink & Neuralink, the whole Silicon Valley project in general is intended to have everyone eventually so discombobulated and "interfered with" that they can't even tell they're experiencing exploitation that should cause discomfort and radicalization (and quickly dispatch the few stragglers who can). It's either social rupture or total oligarch victory in the class war and a 10,000-year Thielreich.
> s intended to have everyone eventually so discombobulated and "interfered with" that they can't even tell they're experiencing exploitation that should cause discomfort and radicalization (and quickly dispatch the few stragglers who can).
It sounds like you have not read Harrison Bergeron by Kurt Vonnegut.
Yes, the real danger we face is that the sorts of special, gifted people who "seek tax advice" from Jeffrey Epstein might some day have all their brilliant, wondrous contributions to the world stymied by oppressive systems of control. Not sure what systems those would be, since they own and are building all the ones we can see around us today, but still: collectivism ooga booga!
> we live under a system that is arranged for the sole benefit of the owners of capital
This is totally false. The vast majority of consumers enjoy huge benefits from the system while owning almost no capital. For example, Walmart customers or iPhone owners.
A lot of people can't tell the difference between capitalism (which has made their lives materially wealthy beyond imagination) and the root cause of today's economic troubles for ordinary people, which is affordability, which is mostly driven by the housing crisis, which is dominated by nimbyism in megacities.
Fix megacity housing regulation to enable cheap/low risk building that the market wants, and you fix the affordability crisis.
No need to rebuild the (greatest system in the history of humankind) from scratch.
You can do it on a personal level, but when everyone else is overworking you, your manager will compare your output based on your peers, and based on it, you might be negatively impacted
Yeah absolutely. It’s hardly things like Claude Code that are the problem, Slack (or other forms of communication) are much easier to slip into personal time and have been a trend since Blackberries were invented.
This is always the reason I'm interested in this exact workflow. Want to build something but never have the time without sacrificing significant amounts of sleep but now it's easier than ever to get things building.
Exactly. My main interest in remote Claude Code is to maintain state continuity from all client hosts. I have a
lot of laptops and mobile devices and I don’t want to manage my git and cloud connections for each. Setup and rebooting are pretty disruptive to short bursts of inspiration or iteration.
This is a complete fantasy. If LLM's got to this point of sophistication there would be a total revolution in almost every industry. Society would be radically different. Since LLM's are nowhere near this, I'm not so sure we even have Pandora's box, let alone opened it.
Glad someone is rational. I believe this new wave of zeal is being somewhat driven by an Anthropic astroturfing campaign.
This AI fear wave has outed that many people have not even the most basic grasp of economics, or the ability to carry a thought to its natural conclusion.
For example, I'll often see people espousing: "there will be no work left, better get rich now or you're screwed!". What's the point in getting rich if there will be no work left? Money is merely a means to an end; in this world with no work everyone will have the ends (goods and services) for free, or else goods and services will still have value and therefore jobs will still exist.
Another equally silly argument "only software will be completely replaced because it is verifiable".
I've never seen completely verifiable software, but let's presume it exists! If software engineering can be replaced (or some large part of it) I will simply say to my LLM "please make me a piece of software that replaces my accountant/lawyer/...", for that matter I could just as equally say "please make me manufacturing software for a perfect humanoid robot and a plumber/bricklayer/electrician protocol". LLMs cannot do this? Then software engineers will move to solving these problems. If LLMs can do it, then the entire economy will be meaningless and Dario/Sam/Elon/etc... will be no richer than you or I.
But, as you say, LLMs are not close to being able to do any of this (and yes... I use Claude Code)
> I believe this new wave of zeal is being somewhat driven by an Anthropic astroturfing campaign.
Yeah I've sort of noticed this on X for the brief time I was on there this weekend. The Claude Code creator was hyping it up to the moon, and when people called him out for it he said he would feel the same way if he wasn't making 1000 racks a year with it. Sure mate.
What people don't realise is if tech progresses to the point where everything is automated, the marginal cost of everything will basically go to zero. It would be better to give away food and shelter for free if it keeps things peaceful. And if not, people have revolted for far less.
That being said it's a complete utopia and once this bubble pops we are basically going to be where we were, but with excellent natural language parsing and generation, with some useful code generation and introspection tools, writing assistants, etc. Which will be great, but not world changing.
>> only software will be completely replaced because it is verifiable
the thing most (especially non-devs) don't understand is that if software can be automated - 99% of all knowledge work will be replaced, as software is the ultimate automation.
There would be absolutely no issues automating accountants/lawyers/etc etc etc.
Sure few will be left but 99% can be automated when software is that advanced.
Not only knowledge work, also a massive amount of blue collar jobs. AI already can guide you how to fix a lot of things or analyze issues with plumbing/electricity/you name it.
I don’t think doing that will change anything. Only real options - without a career shift - that I’ve identified are to work for companies building something that’s never been built before, or building a SaaS that serves a niche.
How stopping using hyperscalers models on their infra would "get as much of this capability into the open as possible"?
Either "we" create models better than commercial state of the art (by using whatever means).
Or we use open models AND fund organisations building such models (could be by purchasing service from these orgs or donations - in which case would these orgs be different than hyperscalers?).
But i dont see how just hosting the models on some private servers would give us an edge?
Have fun trying to afford the necessary hardware to run open models acceptably. The big labs are trying to make sure we won’t be able to in short order.
It's utterly unreal to me to hear so little discussion about labor organization within software during these nascent moments of LLM deployment. Software engineers seem totally resigned toward reduced salary and employment instead of just organizing labor while still in control of the development of these systems.
I really don't get it -- is it that people think these technologies will be so transformative that it is most moral to race toward them? I don't see much evidence of that, it's just future promises (especially commensurate with the benefit / cost ratio). When I do use this tech it's usually edutainment kind of curiosity about some subject matter I don't have enough interest in to dive into--it's useful and compelling but also not really necessary.
In fact, I don't really think the tech right now is at all transformative, and that a lot of folks are unable to actually gauge their productivity accurately when using these tools; however, I do not believe that the technology will stay that way, and it will inevitably start displacing people or degrading labor conditions within the only economically healthy remaining tranche of people in America: the white collar worker.
I've been writing software for 30 years, a part of it had success in the sense of being widely known and adopted for a long time. Writing software is difficult, consumes time and is difficult as you get older to focus the needed time away from other matters like a professional life and family.
With LLM, my productivity suddenly went up x25 and was able to produce at a speed that I had never known. I'm not a developer any more, instead feels like project manager with dedicated resources always delivering results. It isn't perfect, but when you are used to manage teams it isn't all that different albeit the results are spectacularly better.
My x25 isn't just measured on development, for brainstorming, documentation, testing, deployment. It is transformative, in fact: I think software is dead. For the first time I've used neither a paper notebook nor even an IDE to build complex and feature-complete products. Software isn't what matters, what matters is the product and this is what the transformation part is all about. We all here can write products in languages we never had contact with and completely outperform any average team of developers doing the same product.
Replaces the experts and domain specific topics? Not yet. Just observe that the large majority of products are boringly simple cases of API, UI and some business logic inside. For that situation, it has "killed" software.
You'd be surprised what you can do with Claude Code. Pick any mature programming language, including niche ones like Ada and treat the project seriously. Write detailed agent files, features spec files, start from the bottom with CI/CD and set up a test suite, coding guidelines, static analysis. Be careful to create a consistent architecture and code base early.
You'll get a lot further and faster than you'd expect.
Things will probably plateau as you master the new tech, but it's possible you'll not write a ton of code manually along the way.
Oh, your general software development experience should help with debugging the weird corner cases.
I imagine it's really hard to do this with 0 software dev experience, for example. Yeah, you'll build some simple things but you'll need and entire tech education to put anything complex in prod.
That is correct. I'm able to steer the project and find many of the issues because of experience. Also, it is indeed a new tool so is necessary to change our own mentality about the way how code is created (generated). In overall it is fantastic despite the occasional frustration when the computer hangs (too many compilers at the same time) or when AI gets lazy and tries to avoid implementing what it was asked.
Using Claude code Pro with a maxed subscription and ChatGPT Codex with the business subscription.
The code is written in Dart and never wrote a line of DART in my life, I'm a veteran expert around Java, C++. The reason for choosing DART is simply because it is way readier for multi-platform contexts than Java/C++. The same code base now runs on Linux, Android, iOS, OSX, Windows and Web (as static HTML). Plus the companion code in C++ for ESP32 microcontrollers. It also includes a CLI for running as linux server.
Don't ask me for a hard analysis and data proving x25 performance increase, what I know is that an off-grid product was previously taking me two years of research/effort to build in Android/Web and get a prototype running. Now in about a month went far above all previous expectations (cached maps with satellite imagery, bluetooth mesh, webRTC, whatever apps) and was able to release a product several times per day that works as envisioned. Iterating quickly and getting direct feedback from users.
IMHO, Codex is far superior at the moment for complex tasks, Claude is cheaper and still good enough quality for most tasks. In addition to keep several terminals with tasks in parallel, this gives me time throughout the day for other tasks with family/friends and a lot of motivation like a coding-buddy to try different routes and quickly implement a prototype instead of always being alone doing this kind of work. For example, it added an offline GPT bot but wasn't what was needed so could quickly discard it too.
These tools get lost on API implementations and the documentation folder is mostly there to provide the right context when needed. I've learned to use simple markdown documents with things to keep in mind like "reusable.md" or "API.md" to make sure it won't reinvent them. Given my experience, there are parts that I'd implement with higher quality on my own, the trade-off is that I can't touch the code by myself now. One of the reasons is that it would make more difficult for these AI to work since my naming and file structure would make it difficult for the AI to work with, the other reason is because I don't want to waste a full day on a single problem like before. As the product grows more stable is when more attention is given to the finer details. On early stages, that type of quality is still more than good enough for me.
You can try the Android or Linux versions if you are so inclined. Never in my life would I ever be able to build so much in 5 weeks.
Would you describe this product as a whole application suite (blogging, calendar, commerce) plus its own backend infrastructure that is capable of serving these apps to the public internet and functioning offline via ad-hoc wireless peer-to-peer, with a cryptographic layer providing identity, security and censorship resistance, and that runs on phone, laptop or raspberry pi?
Quite ambitious.
Is this an LLM hallucinating? taking a break from coding? or leaking your personal desktop session?
I've answered in more detail on the other reply below on the conversation. Thank you for spotting that.
> Would you describe this product as a whole application suite
The rabbit hole goes even further. The reason why callsigns are used is because geogram can happily communicate using radio-waves on walkie-talkies without internet at all. On the previous iterations (before AI) it was sending free SMS using walkie-talkies and satellites (APRS), this current incarnation should soon be doing the same things too. A presentation from two months ago: https://www.youtube.com/watch?v=Nb_VUSaNw8k
This is a niche app, written for our community in Portugal to connect with each other.
Would you say you’re able to draw a diagram of the application architecture out of your head or do you treat it as a black box? Do you need an AI to debug issues or not? In my experience with spec driven development, even if reviewing every single PR, it is hard to develop a mental model of the codebase structure unless you invest on it. It might be fine to treat it as a black box, not arguing the opposite but will all software be a black box in the future?
For a completely new project it is a high risk. While the AI is fantastic at brainstorming and writing detailed architecture, it is difficult to get the "big picture" and even more difficult to verify that it is being done correctly or which things can be improved/reused, because on this situation you don't look into the code.
I don't believe people will spend time looking at the code beyond the small blurbs they can read from the command line while talking with the AI, so I agree with you that it ends being treated as a blackbox.
Did an experiment for a server implementation in Java (my strong language), gave the usual instructions and built up the server. When I went to look into the code, it was a far smaller and more concise code base than what I would write myself. AI is treating programming language on the level of a compiler for javascript, it will make the instructions super efficient and uses techniques that on my 30 years experience I'm not able to pair-review because we tend to have our own patterns of programming while these tools use everything, no matter how exotic they will use it to their advantage.
After that experience I don't look at generated source code any longer. For me it is becoming the same as trying to look at compiled binary data.
Part of being in a union tends to be lawyering up and "nailing down" exactly what everyone's duties in detail and what fair compensation might be, and what terms / conditions might be etc.
Personally I don't think they're a great fit for the software industry where the nature of the job and the details are continuously changing as technology evolves.
But what are you negotiating about? What do all tech workers have in common that wouldn't be better addressed with top level regulations like "right to disconnect"?
The outsized pay for software engineers in the US takes into account a lot of this stuff. Would you trade those 100 things for, say, a salary of $75k a year for a senior software engineer, like they have in Europe?
Meh. The rest of the world also doesn't have big salaries for software devs. The US is the outlier.
It's not just the labor regulations holding Europe back, it's the lack of funding due to not having a unified European digital market.
Netflix Europe needs to have 20+ licensing deals. Selling across Europe at a large scale requires interactions with 20+ legal teams. Language and cultural barriers kill a lot of things.
How do US giants thrive in Europe, then?
Because they come in directly giant-sized based on growth in the US. They either ignore European legal compliance until sued or pay peanuts for them to handle all the legal aspects.
All those sorts of protections seem like they make sense for every worker rather than being "tech" specific. I do understand that collective bargaining could help with carving out sector-specific deals, though.
I wonder if there is a difference in context that explains why we might disagree. I'm in Australia where I think it's politically easier to "add" broad top level protections for all workers than it would be in the US.
Yeah, the legal framework (Taft-Hartley) in the US is pretty explicit about banning general strikes and solidarity strikes. A union can organize within a single industry but not beyond that.
Before I get into it: what careers do you think are most compelling? Especially if you think all white collar work is going to be undermined by this technology.
I wrote this up a bit ago in my essay fragments collection. It's rough and was just a thought I wanted to get down, I'm unsure of it, but it's at least somewhat relevant to the discussion here:
LLM or LLM-adjacent technology will never take over the execution of work in a way that approaches human where humans continue to guide (like PMs or C-suite just "managing" LLMs).
The reason is that spoken language is a poor medium by which to describe technical processes, and a well-enumerated specification in natural language describing the process is at-least synonymous with doing the work in skilled applications.
For example, if someone says to an LLM: Build a social media app that is like Tinder but women can only initiate.
... this is truly easily replicatable and therefore with little real business value as a product. Anything that can be described tersely that is novel and therefore valuable unfortunately has very little value practically because the seed of the short descriptor is sort of a private key of an idea itself: it will seed the idea into reality by labor of LLMs, but all that is needed for that seed's maturation is the original phrase. These would be like trade secrets, but also by virtue of something existing out there, its replication becomes trivial since that product's patterns are visible and copyable.
In this way, the only real outcome here is that LLMs entirely replace human labor including decision making or are tools to real human operators but not replacements.
I'm curious how this seed/hash/prompt of an idea relates to ladders of abstraction?
Consider "Uber, but for X"
This wasn't a thing you could deploy as a term pre-Uber.
I'm not sure what this means for your analogy, but it does seem important. Somehow branding an idea reifies a ... callable function in? ???
Maybe something like (just spitballing)
The specification-length needed for a given idea isn't fixed - it's relative to available conceptual vocabulary. And that vocabulary expands through the work of instantiation and naming things?
Which maybe complicates the value story... terseness isn't intrinsic to the idea, it's earned by prior reification work?
Hmm
Basically it seems that "Like Tinder but" is doing a lot of lifting there... and as new patterns get named, the recombination space just keeps expanding?
> Basically it seems that "Like Tinder but" is doing a lot of lifting there... and as new patterns get named, the recombination space just keeps expanding?
Yeah, this feels right. It's like a process of condensing: new ideas brought to life condense metaphors into more compact forms and so make language more dense and expressive. This idea reminds me of Julian Jaynes's description of metaphor condensation in Origin of Consciousness.
A lot of hard work goes into novel products, but once that work has been proven, it is substantially more trivial for human or machine to copy. Groping around in the darkness of new, at the edge of what-could-be is difficult work that looks simple in hindsight to others who consider that edge a given now.
> The specification-length needed for a given idea isn't fixed - it's relative to available conceptual vocabulary. And that vocabulary expands through the work of instantiation and naming things?
Yeah, I think that naming and grouping things, then condensing them (through portmanteau construction or other means) is an underrated way to learn. I call this "personal taxonomy," and it's an idea I've been working on for a little bit. There is just tremendous value in naming patterns you personally notice, not taking another person's or group's name for things, and most importantly: allow those names to move, condense, fall away, and the like.
I left out a piece of my fragment above wherein I posit that a more constrained form of natural language to LLMs would likely lead to better results. Constraining interaction with LLM to a series of domain-specific metaphors, potentially even project specific givens, might allow for better outcomes. A lot of language is unspecific, and the technical documents that would truly detail a novel approach to an LLM require a particularly constrained kind of language to be successful where ambiguity is minimized and expressiveness maximalized (legal documents attempt at minimal ambiguity). I won't go into details there, I'm likely poorly reiterating a lot of the arguments that Dijkstra made here:
If programmers think they can just learn a trade, they’ll bein for a rude awakening when Elon comes for their jobs next. Optimus will be doing your plumbing by the time you graduate from trade school and get your paper and internships.
Unfortunately, it's futile to try to convince the median HN poster that labor organization could help them. They've drunk the entire pitcher of corporate anti-union koolaid.
People could be directly in the middle of losing their own job or taking on the responsibilities of 5 other laid-off coworkers, and they would still ask "what could a labor union possibly do for me??"
Big tech laid off 150,000 people last year despite constantly beating wall st expectations and blowing more money than the Apollo program on a money losing technology with the stated goal of firing even more people. Totally insane that most people I talk to still don’t think they need a union.
1. Like most labor organizing, I think this would be beneficial for software engineers, but not long-term beneficial for the world at large. More software that is easier to make is better for everybody.
Would you still want to live in a world where your elevator stops working when the elevator operator is sick, or where overseas Whatsapp calls cost $1 per minute, because they have to be connected by a chain of operators?
2. Software engineering is a lot easier to move than other professions. If you want to carry people from London to New York, you need to cater to the workers who actually live in London or New York. If you want to make software... Silicon Valley is your best bet right now, but if SV organizes and other places don't, it may not be your best bet any more. That would make things even worse for SV than not organizing. Same story applies to any other place.
Sure, companies won't more overnight, but if one place makes it too hard for AI to accelerate productivity, people will either go somewhere else, or that place will just end up completely outcompeted like Europe did.
This would potentially be true for a lot of tech in the last five decades or so. When it gets cheaper to make the things people need and want without those needs and wants changing, you can get away with paying people a lower real wage for the same productivity. Couple that with the fact that the workers themselves also have typically grown more productive from the same tech, allowing companies to undercut competitors and capture more market share until everyone else catches on. I figure capital has benefited enormously from recent tech, very possible it captured the majority of the excess money produced.
I don't think that's possible to analyze for most technologies. How could we determine the effect of, say, OLED technology specifically on workers' real wages across the economy? Even doing the same for a particular seller's margin, say LG, would be difficult and wouldn't tell the full story. If you have an idea of how to do that for something let me know.
We'd probably want to use a measure of worker productivity itself as a proxy for technological improvements and look at various measures like real wages in relation to it rather than restricting our analysis to any one technology.
Small newspapers full of classified ads used to be available locally around the world, creating local employment. Google and Meta ravaged that and sucked the money out to a handful of shareholders and tens of thousands of highly paid tech workers. That's just one market.
> It's utterly unreal to me to hear so little discussion about labor organization
Never lived in the US, where I assume you are from. It's the same country that contrary to most countries, does not have May 1st as a Holiday. Same country that has states with at will employment, etc etc.
I wish I knew which union to pitch. All I can say is what I know which is if you are dispirited with this state of affairs a great way to figure out where to go with it is to connect with your local democratic socialists of america branch, or maybe the joint union dsa effort:
Yes, labour unions are immoral. Curtailing growth (especially in industries where it can prevent unnecessary death) for your personal needs is plain evil. I say that as someone who is both very stressed by pressure to sustain my family while cushy life is slipping away.
One of these is immutable (shitty managers) one of these is new. I personally am all here for the brief human funtime before we all get paperclipped and whatever, been having a ton of fun with CC/Codex, been pushing my own startup forward... but ... You do see the issue here right?
It's the power imbalance. Shitty managers still control your means to eat.
> It won't matter if I'm washing the dishes, walking the dog, driving to the supermarket, picking up my kids from school. I'll always be switched on, on my phone, continuously talking to an LLM, delivering questionable features and building meaningless products, destroying in the process the environment my kids are going to have to grow in.
I remember hearing similar criticisms of continuous delivery. On one end of the spectrum people who had to wait months to get changes out now got them out relatively quickly. On the other end of the spectrum, some person was going to push changes at midnight.
A decade on forward I've never actually worked at a shop that at scale did continuous delivery in its truest sense where changes go straight to production. Simply, nothing beats a human in the loop; it's always about balancing the costs of automation and a lower barrier to entry. I imagine this kind of thing, if it ever actually takes hold and can be adopted by a larger subset of engineers, will follow a similar path.
Long way of saying, I don't think you're Chicken Little but also don't start breathing into a bag just yet.
The difference here is, you type a command into your phone at 3pm. Put it down to go play with your kid for 3hours. Type a new one in at 9pm before bed where you’ve been binging your wife’s favorite show. Then you wake up at 10am to a holistic transformation in your business that would’ve taken months previously in your career. But whatever, another command and it’s off to 11am frisbee.
>But the cognitive dissonance I feel when I think about what this means over a longer time horizon is really painful.
Excluding work (where granted, some companies are dictating the use of llms) and trying not to sound uncaring or disrespectful, but have you thought about not using llms for everything and using the old grey cells? Not having answers to every whimsical thought might be a good thing.
It's very easy to relax the brain (and be lazy tbh) with llms and it's scary to think what will happen in the next 4 years in terms of personal cognitive ability (or as a society).
e.g. I've noticed (and probably most have here) that the world is full of zombies glued to their phones. Looking over their shoulder (e.g. on a train, yeah it's a bit rude but I'm the curious type), they are doom scrolling or playing waste-time games (insert that boomer meme in Las Vegas with slot machines [0]). I try to use my phone as little as possible (especially for dog walks) and feel better for it, allowing me to daydream and let boredom take over.
Maybe I'm fortunate to be able to do this (gen-x: having grown up before cell phones/internet), but worth stating in case anyone wants to try.
There is evidence that LLM usage is actually making people dumber. I'm not sure if they've figured out the cause/effect or not but that's enough evidence for me to avoid them if I can. They can be useful for some stuff but I found myself offloading my thinking a little too frequently.
Anyways if we do get to the point where you need to use LLMs to write code, I can make a decision then, but for now I don't feel the need to adopt agentic workflows and I think the people who don't will be better cognitively positioned in the future.
Probably always be true, but also probably not effective in the wild. Researchers will train a version, see results are off, put guards against poisoned data, re-train and no damage been done to whatever they release.
No thanks. I'm so glad I'm getting closer to retirement age. From a young age, all I wanted to do was program computers. _I_ wanted to do it. Not have some tool do it for me. There's no fun or interest or ... anything that comes from that. I want to solve the problems. I want to write the code. It's what I am good at and it's incredibly enjoyable to me. Why the fuck would I ever give that up?
But, the world is changing. Y'all can have it... in a few short years. ;)
Had the same feeling many moons ago when they gave me an office smartphone where email from the company was available 24/7. At the beginning was answering emails at midnight, nowadays couldn't care less. Just wait until work hours.
Hum, I already have a phone with Slack / Email on. And it's only switched on during work hours. No messaging outside of that window. Why would that be different?
That reminds me of my father calling the mobile phone and laptop issued to him as the "dunce kit", so he could work at home as well. He used to say that since the 90s, ahaha.
In many countries, these and other jobs show you cannot. If you don't, others will and so you won't have a job very soon. Especially if these types of jobs lose their shine/prestige and are basically call center quality/pay like jobs in 5-10 years.
I'd love to believe that, but unless our timeline is disrupted (world war / climate change / regulation re: power generation and consumption), I unfortunately can't imagine a future different to the one I described - and I've tried!
Move somewhere with strong worker rights/laws even if you are not in a union. Here no with a normal job (not freelancers / contractors etc) is looking at their work phone/email outside 9-5/4-5 days a week; this frustrates US companies who merge/acquire companies here greatly but they cannot do much (firing for no cause is very expensive) except slowly move the operation to the US and wind down here, which is expected; everyone is already looking for new jobs as no one wants the 'performance reviews' with the broken records like 'you are not a teamplayer because your colleague was trying to reach you at 22:00 Friday night'.
This was the end game with or without AI. It was always going to result in a zero-sum game because the factories that are open around the clock can output more products - which is exactly why a lot of manufacturing has non stop shift work. If you don’t, you’re leaving money on the table and a competitor will gladly take it.
When you saw 996 being talked about it should have set a few alarm bells off, because it started a countdown timer until such a work culture surpasses the rather leisurely attitude of the West in terms of output and velocity. West cannot compete against that no matter how many “work smarter, not harder” / “work to live don’t live to work” aphorisms it espouses. This should be obvious by now (in hindsight).
You can blame LLM or capitalism or communism but the hard matter is, it’s a money world and people want to have as much of it as they possibly can, and you and your children can’t live without it, and every day someone is looking to have more of it than you are. This isn’t even getting into the details of the personality types that money and power attracts to these white collar leadership roles.
This has been like this forever. Change is that software engineers, historically spoiled and expensive is going to have a brutal reality check - aka we will work just everyone else.
People need to start having conversations about existential risk here. Hinton, Nobel Prize winner in AI, thinks there's a decent chance AI executes the entire human species. This isn't some crank idea.
On the $210/month VM cost - fair point. For me it's worth it because I'm running multiple parallel agents throughout the day, but you could definitely do this cheaper with a smaller instance or spot pricing.
The real bottleneck isn't typing on mobile - it's reviewing the output. I've found 2-3 parallel agents is my sweet spot before I can't keep up with reviewing PRs. Git worktrees help a lot here since each agent works in isolation with its own containers (including db).
And to the work-life balance concerns - totally valid. For me this isn't about working more, it's about capturing ideas when they strike (usually on walks or waiting in line). The Poke notifications let me stay async rather than glued to a terminal.
Although a little late to the HN post, happy to answer any questions about the setup!
I start/stop to save costs when I'm not actively coding. The VM is ~$7/day running, so I halt it overnight. The wait for Tailscale is just ~10 seconds for the mesh to reconnect after boot.
I started with a beefy VM to make sure I could ramp up without hitting slowdowns, but I'm planning to migrate to an Intel NUC running Proxmox at home soon.
how do you manage to review code without running said apps or testing them, sure the code might look correct, but doesn't mean whatever issue / feature works or the bug is actually fixed
Each worktree has its own Docker containers including the database, so I can hit the dev server from my phone browser via Tailscale. For quick checks I'll just curl an endpoint or check the logs. For anything visual, I open the branch's preview URL.
But honestly - if it needs serious testing, I wait until I'm at my desk. The phone workflow is best for "let Claude run while I'm out" rather than deep debugging.
This is cool, but does everybody have enough work for multiple agents in parallel?
I feel like I'm limited by writing specs for agents, then by reviewing their work.
I typically spend 30-60 minutes writing a spec, then the agent runs for 30-60 minutes, then I spend 30-60 minutes refining code/ui/etc before putting up a PR, then another 30-60 minutes waiting for CI + addressing automated + human code review feedback.
I find if a project is made with agentic coding in mind, you can try to implement features as they come to mind and bounce between different CC terminals.
The problem is that this is true of none of my work projects and the onboarding cliff to agentic coding is quite steep. I have to use discretion to apply CC when it does make sense, which is not that often.
I'll have claude code editing a script on my laptop and another on the VM running it and synthesizing the results. Then sometimes I'll have another doing some odd job or researching something. Then I'll pull up gemini in my browser to figure out what I should make for lunch...
I usually use one instance, sometimes two. But this is a reasonable account of Chris Rackaukas using 32 instances at a time to do boilerplate maintenance across a bunch of open source repositories: https://www.stochasticlifestyle.com/claude-code-in-scientifi...
> I have had to spend like 4am-10am every morning Sunday through Saturday for the last 10 years on this stuff before the day gets started just to keep up on the “simple stuff” for the hundreds of repos I maintain. And this neverending chunk of “meh” stuff is exactly what it seems fit to do. So now I just let the 32 bots run wild on it and get straight to the real work, and it’s a gamechanger.
I don't like typing long messages on my phone so this workflow, as cool as it sounds, wouldn't work for me. My current setup is that I have a Claude Code hook that runs whenever CC needs my input and it uses my Home Assistant instance to send a push notification to my phone. I then return back to the computer and continue on the work.
This works reasonably well, but there is a gap for small messages or review comments. I am waiting for Anthropic to shop a feature where the Claude mobile app is able to mirror Claude Code (not the Claude desktop app) and lets me see the diffs of the changes it made and send commands. I'd use this to steer the conversation while on the go with short commands or prompts so that when I'm back at the computer I can focus on the important feedback that I can jot down quickly on the computer keyboard.
I also don't like typing long messages on my phone that's why I use this keyboard that will do high quality transcriptions via whatever AI provider you want. Much better than siri/google speech to text on device.
I used to use Wispr Flow but did not like the non-local aspect, and having yet another subscription, so I switched to VoiceInk (one-time payment around $30 I think), and with a locally running Parakeet v3 model on my
MacBook, transcription is basically instant. I was previously using it with the local Whisper Turbo 3 which is slightly more accurate and it had a 3-4 second lag, so I was absolutely shocked how fast parakeet v3. The slight drop in accuracy is totally fine when talking to AIs, and I also have a line in my CLAUDE.md that says I am usually dictating and that it should take that into account when interpreting my messages.
I'll half-ass something from my phone to take an idea down, but this setup also works with a keyboard. I just run claude in a container with --dangerously-skip-permissions and it does a lot more work with a lot less questions.
I concur that phones aren't great as they are today, but perhaps this exact scenario will prompt the return of PDAs with proper keyboards etc? Or, alternatively, subcompact laptops like the stuff that https://gpdstore.net makes - they are still small enough to be pocketable (needs fairly large pockets but...), and yet you get a keyboard that is actually usable, and they even have one device with fold-out dual screen.
I was looking for a similar scheme, and though far from perfect I found you can run tmux+ttyd. ttyd lets you share your terminal over http. That lets you use your phone's browser (and speech-2-text).
That sounds nice, but what happens when there's something Claude messes up, doesn't know how to do something, or when you have to review the thousand lines it added to your project ?
Unless it's a totally vibe coded side project without any tests or quality control of some sort.
I'm just curious what you can build with this setup. It just seems to be the way to create a mountain of sloppy, unmaintainable code.
If it's really bad, just throw it away. I didn't spend a lot of time/effort generating the code. If my idea was bad or hard to implement I start over with a different approach. Smaller feature or different idea to accomplish the goal.
Essentially you run a server on some machine. Sessions are created in Docker containers, K8s pods, or via Zellij (an app similar to tmux). You can:
- Directly attach to sessions via Docker attach (built-in via a TUI). You get a normal Claude Code experience, but multiplexed. The switcher/UI shows you the status of Claude and the PR (pushed, merge conflicts, CI status, review status, etc.)
- Manage sessions via a web UI. Connect to Claude Code directly via your browser. You have access to the usual Claude Code terminal or a native chat view.
- Manage sessions via an app. You have access to a native chat view.
It achieves isolation via Git worktrees + a proxy so that containers have access to zero credentials (there aren't even any Claude code creds in the container), which allows you to more safely use bypass all permissions mode.
This works better for me that Claude Code on Web because I have control over the environment Claude is running in. I can give it any Docker image I want, I can have it connect to my local network, etc.
It's still a WIP (the core bits are there, but it's not polished yet), but I'm hoping it provides a friendlier UX with a similar goal for what the OP has in mind.
> Port allocation is hash-based—deterministic from branch name, no conflicts:
> hash_val = sum(ord(c) for c in branch_name)
> django_port = 8001 + (hash_val % 99)
> Six agents, six features, one phone.
What do you mean, no conflicts? The probability of a collision with six branches and 99 ports slots is ~14% assuming optimal hashing (which this decidedly isn't).
You also don't need fail2ban, if the entire VM is behind a firewall that only allows the tailscale coordination traffic, nothing is going to reach the VM for fail2ban to work on.
Linode will provide a configured Linux box for $5/mo that works well with Claude Code and Termius. I had to jump through a surprising amount of hoops with Claude Code, Tmux and Termius to issue a shift+tab before Claude Code gained the ability to invoke plan mode conversationally
It makes sense - i build something very similar for my company over the last couple weeks :)
I have a tweak that allows pasting images to claude code over SSH:
How it works:
PTY Interception: It creates a pseudo-terminal (PTY) to wrap the SSH process, allowing it to sit as a "man-in-the-middle" between your keyboard and the remote shell.
Bracketed Paste Detection: It monitors stdin for "bracketed paste" sequences (the control codes terminals send when you Cmd+V or drag-and-drop a file).
The "Hook": When a paste occurs, it pauses execution and scans the text for local macOS file paths.
Auto-Sync: If a local path is found, it immediately syncs that file to the remote server (using the provided SSH key) in the background.
Transparent Forwarding: Once the sync is complete, it forwards the original text to the shell.
You can drag and drop a file from your local Finder into a remote SSH session, and the file is automatically uploaded to the server before the path appears on the command line. Also works with copy paste, screnshots.
This is a bit too "plugged in" for my liking. If I am in line for coffee, it's usually respite away from work, not an opportunity to do more. However, I do love the tmux + worktree + claude setup. I use this now and I know a few peers who do too and it's very enabling. This is what work feels like these days: cycling through agents, each working on a task, checking their work, unblocking them.
You can optimize things. I have a github action that starts stops a fast google cloud vm for our builds. It only gets used about 3 minutes per build. We maybe have a few dozen builds per month. So that's a few hours of run time. The rest of the time the vm is stopped and not billed (except for storage, which is cents per month at most). It's a simple debian vm so it boots in about 20 seconds.
VMs are expensive if you leave them running 24/7 but the logic to start/stop them is pretty easy. There's no need.
Anyway, you need to balance this against the payoff. Agentic coding is useful enough that it beats spending your own time. And that includes waiting time for the relatively slow/underpowered containerized environments that some tools would use by default. I use codex web and codex cli (with a qemu vm so I can use the --yolo flag). Codex web is a bit limited with memory and CPU. Some of my slower builds are taking forever there. To the point where most of the time it consumes is just waiting for these builds to happen.
With a bit of plumbing, you can do things like the author describes pretty easily. IMHO this needs to be better integrated into tools. With Github you have the option to run your own runners. I don't think codex/claude web have similar options currently. But with the cli versions, you can get more creative if you know your tools. And if you don't, use LLMs to drive them for you. It's mostly just about expressing what you want and how you want it.
If you are already paying $200-500/m… and you are doing the work of 10 people… I can totally see the value.
I’ll check the Terragonlabs option.
Lots of options for startups right now, selling pickaxes! I’m waiting for a better terminal experience, personally. I can’t deal with 30+ poorly named windows. I need to be able to search for that one thread I was working on yesterday…
I have a feeling most of these folks are talking about personal projects or work on relatively small products. I have a good amount of personal projects that I haven’t written a line of code for. After bootstrapping an MVP, I can almost entirely drive by having Claude pick up GitHub issues. They’re small codebases though.
My day job is mostly a gigantic codebases that seem to still choke the best models. Also there’s zero way I’d be allowed to tailscale to my work computer from my phone.
From my perspective: tons of very simple, duplicated software. The bad thing is - there is a lot of space on different markets for such software. Here in Poland you can earn for pretty decent life being lame programmer, but building simple automations for small companies.
I was raised in a way I still don’t have courage to switch to such approach, but doing this for 3-4 such entities I can see how you can make living from that.
With LLMs you can automate 90+% of the job if not more.
Why not just use the mobile app? It has Claude Code built in. Maybe I'm an unsophisticated idiot but it works well for me. Some shortcomings with repo management but other than that, CC mobile seems ... fine
I've built something similar to run in my own cloud to run my own container images and add some features. The Claude Code web app has been pretty buggy for me and I like to run these things myself. If I do run into bugs or issues I still have access to the containers (and I do, because I vibe coded the whole thing). I gave it service accounts to the cloud it's running in, where the CI builds take place and setup dev environments per PR that deploy automatically. This is a massive productivity boost imo.
Yeah, me too. The only thing that shows up is poke.com which has something to do with mobile notifications (and seems like at some point they offered some api, but maybe it was discontinued? or something, there are some medium posts talking about their api https://jpcaparas.medium.com/get-sms-or-imessage-alerts-from...
How'd you find this? I spent 5 minutes bouncing between their various domains and marketing videos and login walls and couldn't find the developer docs.
Personally, I'd setup a Mattermost server and use its WebHooks for notifications. That's also more flexible (can send commands back to the bot, etc.)
I currently use Hapi (https://github.com/tiann/hapi/) for this and find it quite handy. I can easily tap into a session on my PC from my phone.
Before that I used Happy (https://happy.engineering/) which is also open source and a lot more sophisticated. It has a voice assistant that can chat with Claude Code on your behalf in the mobile app. However, it wasn't very reliable, and there are other reasons to use Hapi instead (documented in the Hapi repo).
Before that, Omnara (https://www.omnara.com/) a YC company and seemingly a proprietary Happy fork (?) but it never worked properly for me.
Long story short, there are a few of the around, and frankly I really like to use them. Unlike other commenters, I don't find that they wreck my work-life balance. Rather, I can go out and have a walk in the park, only checking in on long-running tasks every once in a while. The diff view is pretty good too. There are many tasks where I'd rather not stare at my PC all day and instead do other things, and these tools allow me to do that.
Also recommend Opencode which has `opencode web` built in. It's really impressive how good opencode web is. It's far more polished than I'd have expected from a free alternative!
Been using Termux and iSH on my phones for years. You can ssh to your server or just directly code for the phone itself.
I also used Web based coding environment like Glitch (R.I.P.) for years.
You can do that with your virtual keyboard, voice or a even a physical keyboard via BT, e.g. Corne-ish Zen.
That's how I travel.
That's really nothing AI specific or novel. It's cool though.
FWIW I even coined a related term https://fabien.benetou.fr/Languages/OwnConcepts#ResponsivePr... "extending responsive design to be able to program on the device, any device from eink to mobile phone to device, one is currently using not just to "consume" content, e.g read a Website that is then properly formatted for it, but rather program back that very device"
That being said, if you do want to go that route check out CloudInit as it will help you (or whatever tool you rely on) to spawn new instance on your favorite cloud provider to boot specific instances and e.g. setup Docker/Podman then services, etc with no interaction. Also ntfy can help you manage notifications across devices on your own infrastructure, no 3rd parties.
Also on the topic of gaps... liminal space IS precious. It doesn't require coding per se though. You can problem solve, jolt down a solution THEN only implement it as code later on. What matters IMHO is that the potential solution is not lost, wherever you might be. For that... I have shower crayons. Point is, CAPTURE ideas, don't necessarily implement on the spot but of course if you want to, it's good to be able to.
I have been doing the same but with happy. It works quite well for quick brainstorms etc. but for deeper work on a real research / plan / implement thing I think you need to actually engage with the output which is hard to do on mobile. Maybe if I had a better UI than terminus to read and check the remote files I would be able to get more done.
I am also hoping / trying to put Claude code on top of a personal zettlekasten to automate more of my “personal life” tasks and get more stuff done for me. Haven’t gotten it really singing yet but I think that could also be really cool.
An alternative might be to start up a VM on exe.dev. Supposedly, mobile access works out of the box [1].
I've not tried that myself since I've only been using it from my laptop, but I do prefer chatting with their coding agent in a browser tab to using Claude Code in a terminal window.
This is interesting. Particularly the notifications flow. I run a simpler setup with webssh on my iPhone over WG back to my LAN and manage Claude that way. It’s fine, and can handle disconnects (with some big cons). I can run code-server via browser on my iPad and can get all the same benefits mosh provides.
One thing to note: the VM seems like an absolute waste of money. If you are using tailscale, might as well connect back to bare metal VMs you can run at home. Save yourself some coin.
There's an X thread where Boris Cherny (creator of Claude Code) shows his workflow of running 5 CC agents in the terminal using 5 tabs while running 10-15 instances of CC on the web [1] and how he moves projects from the terminal to the web and vice versa.
This sounds cool but I feel like I need to often run the code in one way or another when verifying what Claude does. Otherwise it feels like driving blind. Claude Code already has the web version which I could use from my phone and fair it can't run scripts etc which limits the output quality. But if I can't verify what it did it also limits how long I can let it run before I need my laptop eventually.
Ofc if you have demo deployments etc on branches that you could open on mobile it works for longer.
Another issue is that I often need to sit down and think about the next prompt going back and forth with the agent on a plan. Try out other product features, do other research before I even know what exactly to build. Often doing some sample implementations with Claude code and click around these days. Doing this on a phone feels... limiting.
I also can't stand the constant context switching. Doing multiple feature in parallel already feels dumb because every time I come from feature B to A or worse from feature G to E it takes me some time to adjust to where I was, what Claude last did and how to proceed from here. Doing more tasks than 2 max. 3 in parallel often ends up slowing me down. Now you add ordering coffee and small talk to the mix and I definitely can't effectively prompt without rereading all history for minutes before sending the next prompt. At which point I might have also opened up my laptop.
Ofc if you truly vibe code and just add feature on feature and pray nothing breaks, the validation overhead and bar for quality goes down a lot so it works a lot better but the output is also just slop by then.
I typed this on my phone and it took 20 minutes, a laptop might have been faster.
On deeper look, not quite what I want. I want to run Claude Code on my laptop at home and when it fires a blocking question or pauses, I want a mobile app to show the question or allow me to interact with that session.
>Development fits into the gaps of the day instead of requiring dedicated desk time.
I find myself planning and jotting down things into a notebook while juggling adult/parent responsibilities. On little longer gaps I research. Then when the occasional longer gap happens I'm ready to start cracking on my desktop. I've been only dabbling with AI but have found that writing prompts by hand in the notebook and using the desk time to execute them works well. This also keeps me in the free tier.
Awesome. This really sounds like it would fit into my life, too. For once I really hope someone can productionalize this pattern into a platform with tooling, and pay for only what it needed. As it stands this feels expensive, even if it the price of coffee.
Ive thinking about doing something very similar to this as I've a small 3 node k3s cluster at home but couldn't get enough motivation to start thinking I could just use termux + ssh and installing codex cli.
I do want a setup like this, however, most of my development is on Windows which means license cost is usually higher than the cost of the VM. I could run vm's on my home machine, but even then I feel like the terminal experience is quite poor. You want to have a mobile native code, to check the code/read the plans. So far I have been using teamviewer to access my home desktop which works, albeit annoying to use, plus I don't have fancy notifications. Perhaps a web first approach with a mobile responsive web app would work, that shows the files of the project as well as the terminal.
hey there, shameless plug here, I built pinacle.dev exactly with this use case in mind, cheap $7 VMs that comes with vs code, vibe kanban everything else needed out of the box to keep vibe coding on the go, no need to leave your computer running
Setup is still rough around the edges (use an agent to set it up), but clawdbot (prev clawdis) from Peter Steinberger works phenomenally well for agent orchestration and personal assistance. The community for clawd is exploding right now, and I think this is purely based on merit. It’s been a game changer for my vibe coding workflow, and lots of fun.
I've been running a variation of this for the past 3 weeks. I swapped out the default pi agent back to Claude Code because I didn't like the smaller feature set. Bought a phone line and communicate with my agent via iMessage on a clamshelled mac. A Tailscale network connect the head agent to all the computers on my network including my laptop, a few raspberry pi's, steam deck, and all the IoT devices in my house. As I discover new uses, I ask it to make skills and it is remarkable what it's been able to handle all through the single chat interface because it has 24/7 access to all my computers' file systems and my home network. It's been really fun to see how far I can take it, and the skills framework built into CC/Codex now make it feel infinitely extensible.
I should note, a lot of the functionality I built into my agent was custom after-the-fact because (at least three weeks ago) the clawdis repo was in a state that I found very broken and with tons of false information. Luckily it's easy work for Claude to get things working for you, but really the key unlock was the phone line through iMessage and the unrestricted access to all my systems. It really does feel like I'm able to work with any of my files anywhere now, while hardly requiring much of my attention at all. I would recommend something like this at the bare minimum if you intend to implement a system like this: https://github.com/kenryu42/claude-code-safety-net
Both clawdbot and pi have improved and expanded functionality a lot during the last 3 weeks, maybe worth another look? What you have described sounds a lot like the experience I’m having.
At this point I really don't want to even imagine the headache of implementing their codebase updates into all the custom scaffolding that my own fork relies on now. I think my plan going forward is to cherry pick features that sound interesting and re-implement on my own using my agent that has proper documentation on my personal configuration. Will check out what's new though.
As an aside have found the mosh + tmux Claude Code experience somewhat suboptimal, tmux's scrollback seems to clash with CC's, and makes copying between windows etc challenging.
It is tolerable on an iPad with Blink with commands to maximise and minimise panes using vim-style keyboard bindings, kind of like an iOS sway.
I really want to use Claude Code on the phone or tablet, with voice commands only, and perhaps a few simple approval thumb actions. I don't want to type out complex prompt information on a virtual keyboard. I tried setting this up with some of the iOS terminal emulators, and it almost worked, but there was some glitch where Claude would try to start using the first characters that arrived from the voice command.
Not quite the same thing, but I wrote my own agent (as in a replacement for Claude Code) that uses SSH for all operations. That means I can run a very minimal VM (like 4GB RAM Oracle free tier), run the agent locally, and the agent only operates on remote files.
The limitation is that some Typescript builds run out of RAM (even with swap) and I can't use playwright, but still it's been useful.
But this way I can open the firewall, npm run dev and send the link of my new vibe coded security vulnerability/app to my friends without my computer running.
Plus a VM for this, a container for that and soon my 32GB memory isn't enough. I offload aggressively.
Copilot Agent and Claude Code use their own sandbox, which requires less setup but is also quite limited. With your own cloud setup, agents can perform better end to end testing, including database dependencies and specific tool calls.
I'm almost there. I also have tailclscale/SSH/Claude sessions on a VM.
The thing I'm missing is a phone that makes it comfy. I could just SSH feom my standard S23, but what I've got my eye on is one of those foldable things.
Has anyone used one like a laptop? Keyboard on the bottom half, terminal on the top? Does it work decently?
I wonder when/how to test and review the code though? I mean, how do you know Claude Code hasn't entered a completely different path than you had imagined?
Will we still use "batch jobs" agents in 2027? Checking a Java program and downloading a 10mb program used to be slow things which now happen faster than the blink of an eye.
1.Install Tailscale on WSL2 and your iPhone
2.Install openssh-server on WSL2
3.Get an SSH terminal app (Blink, Termius, etc.). I use blink ($20/yr).
4.SSH from Blink to your WSL2’s Tailscale IP
5. Run claude code inside tmux on your phone.
Tailscale handles the networking from anywhere. tmux keeps your session alive if you hit dead spots. Full agentic coding from your phone.
Step 2: SSH server
In WSL2:
sudo apt install openssh-server
sudo service ssh start
Run tailscale ip to get your WSL2’s IP (100.x.x.x). That’s what you’ll connect to from your phone.
Step 3: Passwordless login
In Blink, type config → Keys → + → create an Ed25519 key. Copy the public key.
On WSL2:
echo "your-public-key" >> ~/.ssh/authorized_keys
Then in Blink: Hosts → + → add your Tailscale IP, username, and select your key. Now it’s one tap to connect.
Switch apps, connection dies, no problem. Reconnect: I can just type `ssh dev` in blink and I'm in my workstation, then `tmux attach`, you’re right back in your session.
Pro tip: multiple Claude sessions
Inside tmux:
•Ctrl+b c — new window
•Ctrl+b 0/1/2 — switch windows
I run different repos or multiple agents in the same repo, in different windows and jump between them. Full multi-project workflow from my phone.
Shout out to https://exe.dev for this stuff. It'a a VM provider service. It makes it stunningly easy to get https up and going, has a front end http gateway that does all the hard parts for you.
But relevant to this article here, it also has a super sick web based agent, Shelley, that is quite adequate for using from the phone.
I'd also note that OpenCode is a solidjs app, that can run in tui (how most folks know it) or the web. And it has an excellent excellent plugin architecture. The work in this post to build workflows is great!
Is there a way to use the official Claude web app with GitHub providers other than GitHub? I’ve put a decent amount of effort into moving away from there, I don’t want to go back.
I do the same, but with ConnectBot and Gemini CLI. I have found ssh sufficiently good (mosh required some port forwarding dance, that Tailscale may have solved for the author).
I work on Poke; admittedly the docs[1] are a bit hidden but we're spending Q1 making it far easier for developers to work with Poke, feel free to email me if you're having any troubles!
But anthropic has since launched the ability to “teleport” sessions to mobile. (Claude Code is baked into the app). The iOS experience has been smooth for the most part.
People keep saying things like “2026 is the year of background agents, sandboxes, etc” but imo the harness will eat the entire platform stack. It already is. It will only get better.
Hah, I set up basically the same thing on Saturday during a long car ride. Couple of differences: I’m an opencode user and I used a different VPS provider (though I use vultr for other things). It was my first time actually sitting down and using tailscale, which was quite easy to get going. Did everything from my phone, didn’t even have my laptop with me.
This is a pretty sophisticated setup. I particularly like how it uses Tailscale.
I've been using the simpler but not as flexible alternative: I'm running Claude Code for web (Anthropic's version of Codex Cloud) via the Claude iPhone app, with an environment I created called "Everything" which allows all network access.
(This is moderately unsafe if you're working with private source code or environment variables containing API keys and other secrets, but most of my stuff is either open source or personal such that I don't care if the source code leaks.)
Anthropic run multiple ~21GB VMs for me on-demand to handle sessions that I start via the app. They don't charge anything extra for VM time which is nice.
I frequently have 2-3 separate Claude Code for web sessions running at once, often prompted from my phone, some of them started while I'm out walking the dog. Works really well!
I don't like claude code web due to its lack of planning mode. I found the result is often lackluster compare to claude code cli.
My current setup: Tailscale + Terminus(ipad) + home machine(code base)
Need to look into how to work on multiple features at the same time next.
I've been using git worktrees with Claude and it's pretty awesome:
https://www.youtube.com/watch?v=up91rbPEdVc
Pair worktrees with the ralph-wiggum plugin and I can have Claude work for hours without needing any input:
https://looking4offswitch.github.io/blog/2026/01/04/ralph-wi...
Worktrees took way too much setup and hand-holding for me, but https://conductor.build made it easy!
I delayed adopting conductor because I had my own worktree + pr wrappers around cc but I tried it over the holidays and wow. The combination of claude + codex + conductor + cc on the web and claude in github can be so insanely productive.
I spend most of my time updating the memory files and reviewing code and just letting a ton of tasks run in parallel
Sadly only allows sign up with Github.
software is all about wrappers, isn't it? :)
conductor -> multiple claude codes/codexes -> multiple agents -> multiple tools/skills/sub-agents -> LLMs
I haven't missed planning mode myself. I tend to tell it "write a detailed plan first in a file called spec.md for me to review", then use that as the ongoing plan.
I like that it ends up in the repo as it means it survives compaction or lets me start a fresh session entirely.
I was doing the same, but recently I noticed that Claude now writes its plans to a markdown file somewhere nested in the ~/.claude/plans directory. It will carry a reference to it through compaction. Basically mimicking my own workflow!
This can be customized via a shell env variable that I cannot remember ATM.
The downside (upside?) is that the plan will not end up in your repo. Which sometimes I want. I love the native plan mode though.
Plans in plan mode also survive compaction
The lack of Plan Mode is puzzling, I'm sure they must get to it at some point. But until then it CAN still plan, you just have to ask it to write a plan and not write code yet.
Not sure if this works in claude code web, but running non-interactive claude code I can still get it to use plan mode by simply asking it. It's just a tool call.
I've been really impressed with https://github.com/BloopAI/vibe-kanban to do this. Really really impressed.
Can you not use PAL MCP for this? Have one top agent as controller etc? It's not ideal but it feels like the space of multi agent stuff is evolving ... I notice that there are a lot of posts on hn about these things so we are trying to do the same thing really.
I'm surprised to see people getting value from "web sandbox"-type setups, where you don't actually have access to the source code. Are folks really _that_ confident in LLMs as to entirely give up the ability to inspect the source code, or to interact with a running local instance of the service? Certainly that would be the ideal, but I'm surprised that confidence is currently running that high.
I still get the full source code back at the end, I tell it to include code it wrote in the PR.
I also wrote my own tool to extract and format the complete transcript, it gives me back things like this where I can see everything it did including files and scripts it didn't commit. Here's an example: https://gistpreview.github.io/?3a76a868095c989d159c226b7622b...
Oh fascinating - so you're reviewing "your own" code in-PR, rather than reviewing it before PR submission? I can see that working! Feels weird, but I can see it being a reasonable adaptation to these tools - thanks!
What about running services locally for manual testing/poking? Do you open ports on the Anthropic VM to serve the endpoints, or is manual testing not part of your workflow?
Yeah, I generally use PRs for anything a coding agent writes for me.
If something is too fiddly to test within the boundaries of a cloud coding agent I switch to my laptop. Claude Code for web has a "claude --teleport" command for this, or I'll sometimes just do a "gh pr checkout X" to get the branch locally.
Much obliged, thank you!
Yeah the commits that claude code generate are co-authored by claude@anthropic.com so i just open a PR to see the code. I have automatic per-PR dev environments for manual testing.
What hosting do you use for automatic per PR dev environments?
i run it in my homelab k8s cluster
What kind of homelab do you have? And how do you do routing, do you have some sort of DNS setup too?
just some old pcs, pfsense and tailscale for routing, and external-dns in kubernetes to manage that
These are all some excellent ideas, I need to setup these things asap since I've been going back and forth on having more homelab vs cloud providers, but I'm only hearing about tailscale right now so I got to go for it. Cloud providers all of sudden becoming costly just for my side projects and/or not providing the exact PR environments like I would like etc. I've been wasting so much time on trying to automate AI Agents vs cloud providers with limited conf.. It would be great if AI Agents can just write the config for all deployments, pipelines, standards, without me having to go to any UI to tweak things manually etc.
Even with GitHub CI now all of sudden it wasted $50 on few days of CI actions. Should have everything run on my home server. But I think I may need more powerful home server, I have a cheap Dell refurbished one now.
I don't want to ever have to touch a UI again (except in places like Hackernews or the like) and the ones I specially built (read: vibecoded) for myself.
Yes my primary motivation for putting so much effort into a self-hosted cloud was cost. Managed Kubernetes instances are very expensive. I've saved a ton of money hosting it myself for side projects. With the benefit that spending $2k on a framework desktop one time to use as a k8s node means I have a much, much larger cluster than I'd be willing to pay for on a month to month basis. It might pay for itself in a single month. It's my opinion that Kubernetes can do anything the clouds can, so I just run talos on the old PCs, the only thing they do is run Kubernetes. Cloud hosting is insanely expensive.
I do have a managed Kubernetes instance that I run public services on (like for webhooks from github) so I dont need to open my home ports. It's very small to keep costs low. The benefit of using Kubernetes at home is most of my configs need minor changes to work on the managed k8s instance, so there's not much duplicate work to get features/software deployed there. It's the great cloud agnosticator after all!
I've started my own web interface for Claude Code to host it in the same cluster. That's where the CI builds happen, the PR envs get deployed. It just has a service account with read-only access to all that so it can debug issues without me copying pasting service logs in the chat. Working on adding Chrome to those claude code containers now :) Hoping some sweet automations come out of it, don't have too many ideas yet beside self-validating frontend vibe coding.
Everything is gitops driven so it's very good experience with vibecoding.
Right - I’m missing how you get the source code in the OP. It says you tmux in with ssh agent forwarding for GH. But you can’t do that on your iOS device? So you have to set up all your repos in the morning before leaving the house, then collect and push all your branches when you return home?
I could imagine this working for a small number of branches/changes.
The output from Jules is a PR. And then it's a toss-up between "spot on, let's merge" and "nah, needs more work, I will check out the branch and fix it properly when I am the keyboard". And you see the current diff on the webpage while the agent is working.
Claude Code on the web, ChatGPT Codex and Google Jules are not the same as Claude, ChatGPT and Gemini. They are entire apps where you authorize Github access and they work via PRs.
They'll include screenshots on your PRs etc.
I like using them a lot when I can.
Right, yes, that was precisely my point - it was weird to me that people were comfortable operating on a codebase that they don't have locally, that they can't directly interact with.
> it was weird to me that people were comfortable operating on a codebase that they don't have locally, that they can't directly interact with.
I have a project where I've made a rule that no code is written by humans. It's been fun! It's a good experience to learn how far even pre-Opus 4.5 agents can be pushed.
It's pretty clear to me that in 12 months time looking at the code will be the exception, not the rule.
12 month from now, when something go wrong, you'll have a lot of code to look at and debug!
> the exception, not the rule.
Absolutely - for me, that's already true. I just wouldn't want to give up the ability to _ever_ look at the code before I submit it!
when the agent pushes the PR, in a branch, you can switch to that branch locally on your machine and do whatever, review it, change it, and ask for extra modifications on top, squash it, rebase it
Yes, that had already been suggested here: https://news.ycombinator.com/item?id=46492948
Are those VM specs documented anywhere, I have used Claude Code for web a lot and never really bothered with the details. Just connect it to my repo and let it cook
Not documented, so I had Claude code for web write me a report: https://github.com/simonw/research/blob/main/environment-rep...
Check out superconductor.dev (I’m building it), if you want live app previews, docker-in-docker functionality, multiple agents in one mobile app, and more.
I have my very fast macbook pro at my desk in my office, and I use tmux and tailscale and git worktrees and I’ve built a notification setup like this author.
Thanks to tailscale and ssh I can vibecode on the go from my phone with this setup.
While it’s great to leave a task running, no matter what I do I can’t achieve the type of high quality work on the go that I can when I’m sitting at my desk.
For me working on a full SaaS.. I just can’t do quality work on my phone.
The only way I can do quality work is to sit at my desk where I’m focused on the work. To play with the result of a prompt, take copious notes, feed them back to the agent, not ship until the thing is polished to a shine. To feature flag the changes, review all code in excruciating detail as though it was written by a dyslexic intern, add all the metrics and logs one can think of (VictoriaMetrics), add user-behavior logging (Amplitude/Posthog) and monitor the thing like your livelihood depends on it. Because it’s a product and you have pride in your work.
All of that needs loads of screen real estate and a keyboard.
So I guess the more things change, the more they stay the same.
Same here, I’m vibecoding a toy project where I never looked at the code from my phone, but I always seat for work. I’m using happy app and that’s good enough for now, I have the desktop in tailscale but I access it that way just for testing
For me, setup like this is not a replacement or even supplement the desk + high focus environment. My use case for these mobile vibecode setup is great though for small exploration/POC/learning/research type of work. Then I take the knowledge learned or any actual useful parts of code and incorporate into my work next time I sit back down at my desk.
There has been a sustained campaign over the last few days to push "I use Claude from my phone". I saw multiple posts on LinkedIn already, and now this.
This blog is super sus too. All the posts are about Claude. I suspect it's run by Anthropic, just read the About page: https://granda.org/en/about/
I am not Anthropic. I have vibe-coded a web interface for claude code to use from my phone. It's very similar to Anthropic's, but has a couple more features that I wanted.
I think your statement "I use Claude from my phone" is quite a large set of folks - the iOS app alone likely has several millions of installs.
Not everything is a conspiracy or content marketing by some bullshit company.
I'm increasingly using Claude from my phone because the models are now good enough to use unsupervised.
There's nothing suspicious to me on that About page.
How hard is it to build a marketing agent these days?
Pandora's box is open; we're moving towards a world where white collar workers will be working 24/7 and they'll be expected to do so.
It won't matter if I'm washing the dishes, walking the dog, driving to the supermarket, picking up my kids from school. I'll always be switched on, on my phone, continuously talking to an LLM, delivering questionable features and building meaningless products, destroying in the process the environment my kids are going to have to grow in.
I'm a heavy LLM user. On a daily basis, I find LLMs extremely useful both professionally and personally. But the cognitive dissonance I feel when I think about what this means over a longer time horizon is really painful.
This technology should be liberatory, and allow us all to work less while enjoying the same standard of living. We've all contributed in its development by creating the whole corpus of the internet, without which it could never have been bootstrapped.
The only reason we can't expect this is that we live under a system that is arranged for the sole benefit of the owners of capital, and have been convinced that this is an immutable state of affairs or that our own personal advantage can be found in making a Faustian bargain with it.
> we live under a system that is arranged for the sole benefit of the owners of capital, and have been convinced that this is an immutable state of affairs
What alternative do you propose?
I would like to propose a cap on net worth.
Realistically, if you have 300M, you and your direct family are settled for life. So, I want to propose 1B cap on net worth, if its more than that for 12 months straight, surplus goes to government, if your net worth is down after that, government obliges to return it partially to make it to 1B.
People, who are eager building things and innovating, will keep building regardless, power hungry will try to find other ways to enrich themselves, but eventually they will give up (e.g. having 10 kids, each with 1B net worth)
Instead of an arbitrary net number, why not a multiplier of the median? For example, capping at 300x the median citizen/household.
So let's say we implemented this and the government suddenly receives billions of credits from wealthy individuals. Would anything actually change? Is this what's holding us back from repairing all the roads to a fine standard? From implementing universal basic income?
It seems to me that if one tried to actually spend those credits we'd simply get inflation. Prices for roads, food etc. would just go up.
We definitely want to work on inequality, but I think numbers above 1B net worth are just weird quirks of the system. Musk is powerful because he's powerful, the number is just a reflection of that. Keeping his number below some arbitrary threshold isn't going to combat his power.
We need to tackle the problem head on: we need to stop individuals from amassing so much power. We get lost in this stupid abstraction called money. It's not what matters.
This is so arbitrary and incredibly naive. How did you come to with 300M, why not 300k or 300 billion? How would you determine the worth of rare, illiquid or intangibles? What about wealth held in trusts or companies? How does the accounting work if I borrow against my wealth? What happens when things change value dramatically in a short period of time? And the government is going to "make billionaires whole again" if they crater their wealth?
You are asking me about implementation difficulty, difficult implementation doesn't mean idea is not worth it.
One example:
* 300k vs 300M - doesn't matter if I said 100M, 200M, 550M, if you think 300M is not enough for you and your family to afford anything, not sure how other people are surviving for even less.
Here is why I think this is good:
1. Ambitious people will still be ambitious, its rare some genius kid says: I know this is 100B idea, but I won't build it, because I will only own 1B of it.
2. Limits the power, when power is really limited, people will be forced to focus on different things. For example, if you had plans to take over the world by making $10T and creating an army to kidnap president of another state you don't like, then you would know, it is not possible to make 10T, its not only about how much, its about suppressing hungry animal in you by capping your limits.
3. There is a chance "bad" ambitious people, will be converted to real philanthropist, because they know it doesn't matter to own more than 1B anyway and they can't own it.
> You are asking me about implementation difficulty, difficult implementation doesn't mean idea is not worth it.
I can agree with that idea, to an extent. If something is near impossible (not saying this is), then it does become not worth it.
The other questions the parent posed are more interesting to me:
> How would you determine the worth of rare, illiquid or intangibles? What about wealth held in trusts or companies? How does the accounting work if I borrow against my wealth? What happens when things change value dramatically in a short period of time?
Another I wonder is that (ignore all specifics of the values, just the concepts matter here), let's say you own a private business that then becomes valued at 1.5 billion dollars and this individual has 20 million dollars liquid. How do you tax that? The government can't take one third of the business, at least not without a lot of issues (in business dealings and individual rights), and the 20 million liquid wouldn't come close to what this plan would value. What do we do then? Plenty of billionaires don't really have liquid cash and forcing liquidation of assents in such a way seems like it would be very difficult.
I'm all for more taxes on higher net worth individuals, but I think there's a lot of talk to be had on how one can implement this. It's going to be really difficult to find a way that makes sense.
> I can agree with that idea, to an extent. If something is near impossible (not saying this is), then it does become not worth it.
FATCA law makes this very possible in the US.
> How do you tax that? The government can't take one third of the business, at least not without a lot of issues (in business dealings and individual rights)
I would say that the government can and should and simply be a passive share holder with no voting rights.
Obviously, I do not know nitty gritty details of economy and finance, but if I would implement this tomorrow I would start taking the equivalent of surplus and start from there to understand more.
For example, say individual has 20M liquid cash, 2 houses each valued at 5M and 1.5B in company shares (based on averaged company value for the last 6 or 12 months):
* whatever you can immediately spend is prioritised first, so you keep your 20M + 2 houses, then surplus is $530M of your company shares
* this equivalent number of shares will be moved to government trust, individual doesn't have any control over it, if person dies next day, government keeps the money (lets simplify for now and keep voting rights as separate question)
* let's say after shares moved to gov. trust, during next 6 months company value halved, gov. returns all your shares, if stock dropped only 10%, you get equivalent back to make your net worth 1B
* regarding taxation, I would keep it as it is today and tax on "realization event"
There are around 3.000 billionaires in the world, even hiring 10 dedicated people for each billionaire to calculate all this stuff on a quarterly basis is not expensive
They would break ownership into trusts. The caps idea has forever been dumb.
Doesn’t the US have an almost $2T deficit?
What do you think all that money you taxed from billionaires will do in the hands of the “government”, pay the bills for 12-24 months? Then what’s your plan after the government has spent it all? No amount of taxation can solve a spending problem.
Would SpaceX exist if your ideas were in charge?
it would.
Some options why it should happen:
1. Assuming Elon is curious person, he will still build it out of curiosity.
2. Assuming Elon is not curios person, just power hungry, he will probably think its not worth building it, but someone else who is curious will build it eventually. This is even better, because when power hungry person owns such thing, they might use it for bad things as well (e.g. to gain more power, eventually interfering with elections, oh wait, it did already happen)
3. Government will build, because government will have more money now, but then we should be even more careful who gets to the top. Assuming people won't have more than 1B, maybe there will be less lobbying? because its not worth as much as it was before?
Is the reason NASA couldn’t build reusable rockets because they didn’t have enough money? Honest question?.
honestly, I don't think it was money related, because look at James Webb telescope, I think NASA achieved lots of good things, IMO they could have built SpaceX as well, it was probably more motivation and organisational impotence and bureaucracy from management (including higher ups than NASA).
> because government will have more money now
Ah, so your idea is the good old “only the emperor who controls the violence apparatus should have a lot of money and power”?
It’s not a very original idea, and it has been tried many times, and it failed many times.
> but then we should be even more careful who gets to the top
Right, so “for some reason only the greedy power hungry psychopaths get to the top in the current system — let’s fix it so that there can’t be many of them, only one government who has power to take away other people’s wealth and concentrate it immensely, surely we will figure out how to make sure it’s not filled with greedy power hungry psychopaths as we go”
Not forcing a scarcity of necessities like housing would be a start.
Peer competition is what makes everything work. You need scarcity of necessities to force people in to the system. Recent rulings allowing the criminalisation of homelessness are pushing this further. Your existence is default-illegal unless you work to outbid your peers for housing.
Be realistic, demand the impossible.
A reference to France in May of 1968: "Soyez réalistes, demandez l'impossible."
See for example https://en.wikipedia.org/wiki/May_68#Slogans_and_graffiti
"I demand a purple unicorn build the things I want to."
Now what?
Claude Code already is the purple unicorn. We're already there - the only problem is that regulatory systems are set up in a way that benefits a small minority of capitalists, rather than the majority.
The same but less rigged would be a good start. I feel like people ask your question as a gotcha because they can't wrap their head around a system more nuanced than "cancerous capitalism" or "potato famine communism"
Something like we had in advanced western europe and the US between ww2 and the late 70s seemed much more balanced while not requiring a complete system change. Most people would be fine if we sprinkled a bit of socialism on top of the gigantic pile of capitalism.
Stuff like housing, energy, transportation, shouldn't make you live paycheck to paycheck forever. Just the fact that people are slowly starting to talk about 50 years mortgage should be a wake up call.
Most people would be happy knowing there is something a tiny bit better coming, rather than knowing they will never make it out and will kept getting fucked a tiny bit more year after year. My grandparents had objectively a harder life than mine, but their life was improving every year, mine is stagnating at best, and usually I'm losing purchasing power year after year, while being relatively well paid for my country
Perceived success of 1950s and 1960s does apply to non-white people, women or countries destroyed by war.
And it still mostly doesn't so it's not really a good argument. Meanwhile the share of women in the workforce more than doubled but households' purchasing power isn't even as high as back then when it comes to the basics like cars/energy/housing, if anything it's down quite a bit.
We should unite, not fight about who's whitest, XYZ gender or a minority, you'll always find someone who has it better or worse than you, what matter is the average is going down, while back then it was going up, the rest is mostly noise.
It was going up because half the world was suffering results of war, of course untouched countries saw economic growth.
It's always the same: workers need to unionize and form a political power bloc. Then, those most impacted—the majority—have an array of options, which are well explored in the annals of leftist and socialist political theory.
This is not at all to say that more conservative or reactionary theorists are wrong about how the world works. In fact, I think they're usually more right about what's really going on abstractly.
But, the working man doesn't need to know what's really going on. They need to win the war, and there's a ton of tactical advice written down—hard won lessons by those who built the modern world through the labor movement.
The place to start is with the usual suspects. Verso Books, The New Centre for Social Research, histories of the labor movement, and new political commentators like Josh Citarella.
"Let us finally imagine, for a change, an association of free men, working with the means of production held in common ...."
Do the owners of capital work less?
The numbers of hours that they work relative to the average wage laborer bears absolutely no relationship to whether they are the beneficiaries of an exploitative socio-economic arrangement. But they all certainly work much less than the laborers in the most precarious positions who are forced to work multiple gig economy jobs to make ends meet, yes.
It would be a deep irony if LLMs ended up ushering in the social rupture that never arrived in the industrial era. When the pigs turn hogs and refuse to share even the scraps, they shouldn’t be surprised if the system they depend on becomes their undoing.
We should all hope so. It's clear that mass surveillance, the vast psyops architecture including social media platforms, autonomous drone warfare, Starlink & Neuralink, the whole Silicon Valley project in general is intended to have everyone eventually so discombobulated and "interfered with" that they can't even tell they're experiencing exploitation that should cause discomfort and radicalization (and quickly dispatch the few stragglers who can). It's either social rupture or total oligarch victory in the class war and a 10,000-year Thielreich.
> s intended to have everyone eventually so discombobulated and "interfered with" that they can't even tell they're experiencing exploitation that should cause discomfort and radicalization (and quickly dispatch the few stragglers who can).
It sounds like you have not read Harrison Bergeron by Kurt Vonnegut.
Yes, the real danger we face is that the sorts of special, gifted people who "seek tax advice" from Jeffrey Epstein might some day have all their brilliant, wondrous contributions to the world stymied by oppressive systems of control. Not sure what systems those would be, since they own and are building all the ones we can see around us today, but still: collectivism ooga booga!
> we live under a system that is arranged for the sole benefit of the owners of capital
This is totally false. The vast majority of consumers enjoy huge benefits from the system while owning almost no capital. For example, Walmart customers or iPhone owners.
A lot of people can't tell the difference between capitalism (which has made their lives materially wealthy beyond imagination) and the root cause of today's economic troubles for ordinary people, which is affordability, which is mostly driven by the housing crisis, which is dominated by nimbyism in megacities.
Fix megacity housing regulation to enable cheap/low risk building that the market wants, and you fix the affordability crisis.
No need to rebuild the (greatest system in the history of humankind) from scratch.
I’m a remote work from home employee who never ever works overtime.
I do use Claude code for my personal projects and ping at them from coffee shops and micro moments during my free time.
It’s possible to engineer your own life boundaries and not be a victim of every negative trend in existence.
You can do it on a personal level, but when everyone else is overworking you, your manager will compare your output based on your peers, and based on it, you might be negatively impacted
Yeah absolutely. It’s hardly things like Claude Code that are the problem, Slack (or other forms of communication) are much easier to slip into personal time and have been a trend since Blackberries were invented.
This is always the reason I'm interested in this exact workflow. Want to build something but never have the time without sacrificing significant amounts of sleep but now it's easier than ever to get things building.
Exactly. My main interest in remote Claude Code is to maintain state continuity from all client hosts. I have a lot of laptops and mobile devices and I don’t want to manage my git and cloud connections for each. Setup and rebooting are pretty disruptive to short bursts of inspiration or iteration.
This is a complete fantasy. If LLM's got to this point of sophistication there would be a total revolution in almost every industry. Society would be radically different. Since LLM's are nowhere near this, I'm not so sure we even have Pandora's box, let alone opened it.
Glad someone is rational. I believe this new wave of zeal is being somewhat driven by an Anthropic astroturfing campaign.
This AI fear wave has outed that many people have not even the most basic grasp of economics, or the ability to carry a thought to its natural conclusion.
For example, I'll often see people espousing: "there will be no work left, better get rich now or you're screwed!". What's the point in getting rich if there will be no work left? Money is merely a means to an end; in this world with no work everyone will have the ends (goods and services) for free, or else goods and services will still have value and therefore jobs will still exist.
Another equally silly argument "only software will be completely replaced because it is verifiable".
I've never seen completely verifiable software, but let's presume it exists! If software engineering can be replaced (or some large part of it) I will simply say to my LLM "please make me a piece of software that replaces my accountant/lawyer/...", for that matter I could just as equally say "please make me manufacturing software for a perfect humanoid robot and a plumber/bricklayer/electrician protocol". LLMs cannot do this? Then software engineers will move to solving these problems. If LLMs can do it, then the entire economy will be meaningless and Dario/Sam/Elon/etc... will be no richer than you or I.
But, as you say, LLMs are not close to being able to do any of this (and yes... I use Claude Code)
> I believe this new wave of zeal is being somewhat driven by an Anthropic astroturfing campaign.
Yeah I've sort of noticed this on X for the brief time I was on there this weekend. The Claude Code creator was hyping it up to the moon, and when people called him out for it he said he would feel the same way if he wasn't making 1000 racks a year with it. Sure mate.
What people don't realise is if tech progresses to the point where everything is automated, the marginal cost of everything will basically go to zero. It would be better to give away food and shelter for free if it keeps things peaceful. And if not, people have revolted for far less.
That being said it's a complete utopia and once this bubble pops we are basically going to be where we were, but with excellent natural language parsing and generation, with some useful code generation and introspection tools, writing assistants, etc. Which will be great, but not world changing.
>> only software will be completely replaced because it is verifiable
the thing most (especially non-devs) don't understand is that if software can be automated - 99% of all knowledge work will be replaced, as software is the ultimate automation.
There would be absolutely no issues automating accountants/lawyers/etc etc etc. Sure few will be left but 99% can be automated when software is that advanced.
Not only knowledge work, also a massive amount of blue collar jobs. AI already can guide you how to fix a lot of things or analyze issues with plumbing/electricity/you name it.
So if software goes down - everyone will go down.
> white collar workers will be working 24/7
Where we're going, there's no "white collars workers" anymore.
Only white collars Claude agents.
You still need humans to supervise them. Just a lot less.
Yeah, there's no way we have these careers in 30 years.
The best we can do is wrestle the control away from hyperscalers and get as much of this capability into the open as possible.
Stop using Anthropic products and start using weight available models. (I'm not talking ICs - I mean the entire startup / tech ecosystem.)
I don’t think doing that will change anything. Only real options - without a career shift - that I’ve identified are to work for companies building something that’s never been built before, or building a SaaS that serves a niche.
Maybe you can recommend some of those models? I'm honestly bewildered - what is open and not?
Trust yourself to be able to handle agents. Stop trying to be too safe, you’re paying the price with ignorance. Just use Claude Code with Opus 4.5.
How stopping using hyperscalers models on their infra would "get as much of this capability into the open as possible"?
Either "we" create models better than commercial state of the art (by using whatever means).
Or we use open models AND fund organisations building such models (could be by purchasing service from these orgs or donations - in which case would these orgs be different than hyperscalers?).
But i dont see how just hosting the models on some private servers would give us an edge?
Or, use the best tools to make the best products you can and stake your claim before all the low hanging fruit is picked
Have fun trying to afford the necessary hardware to run open models acceptably. The big labs are trying to make sure we won’t be able to in short order.
It's utterly unreal to me to hear so little discussion about labor organization within software during these nascent moments of LLM deployment. Software engineers seem totally resigned toward reduced salary and employment instead of just organizing labor while still in control of the development of these systems.
I really don't get it -- is it that people think these technologies will be so transformative that it is most moral to race toward them? I don't see much evidence of that, it's just future promises (especially commensurate with the benefit / cost ratio). When I do use this tech it's usually edutainment kind of curiosity about some subject matter I don't have enough interest in to dive into--it's useful and compelling but also not really necessary.
In fact, I don't really think the tech right now is at all transformative, and that a lot of folks are unable to actually gauge their productivity accurately when using these tools; however, I do not believe that the technology will stay that way, and it will inevitably start displacing people or degrading labor conditions within the only economically healthy remaining tranche of people in America: the white collar worker.
I've been writing software for 30 years, a part of it had success in the sense of being widely known and adopted for a long time. Writing software is difficult, consumes time and is difficult as you get older to focus the needed time away from other matters like a professional life and family.
With LLM, my productivity suddenly went up x25 and was able to produce at a speed that I had never known. I'm not a developer any more, instead feels like project manager with dedicated resources always delivering results. It isn't perfect, but when you are used to manage teams it isn't all that different albeit the results are spectacularly better.
My x25 isn't just measured on development, for brainstorming, documentation, testing, deployment. It is transformative, in fact: I think software is dead. For the first time I've used neither a paper notebook nor even an IDE to build complex and feature-complete products. Software isn't what matters, what matters is the product and this is what the transformation part is all about. We all here can write products in languages we never had contact with and completely outperform any average team of developers doing the same product.
Replaces the experts and domain specific topics? Not yet. Just observe that the large majority of products are boringly simple cases of API, UI and some business logic inside. For that situation, it has "killed" software.
What tool do you use, which languages? Could you give us an example of something you’ve built and how you did it 25 times faster?
You'd be surprised what you can do with Claude Code. Pick any mature programming language, including niche ones like Ada and treat the project seriously. Write detailed agent files, features spec files, start from the bottom with CI/CD and set up a test suite, coding guidelines, static analysis. Be careful to create a consistent architecture and code base early.
You'll get a lot further and faster than you'd expect.
Things will probably plateau as you master the new tech, but it's possible you'll not write a ton of code manually along the way.
Oh, your general software development experience should help with debugging the weird corner cases.
I imagine it's really hard to do this with 0 software dev experience, for example. Yeah, you'll build some simple things but you'll need and entire tech education to put anything complex in prod.
That is correct. I'm able to steer the project and find many of the issues because of experience. Also, it is indeed a new tool so is necessary to change our own mentality about the way how code is created (generated). In overall it is fantastic despite the occasional frustration when the computer hangs (too many compilers at the same time) or when AI gets lazy and tries to avoid implementing what it was asked.
Using Claude code Pro with a maxed subscription and ChatGPT Codex with the business subscription.
The code is written in Dart and never wrote a line of DART in my life, I'm a veteran expert around Java, C++. The reason for choosing DART is simply because it is way readier for multi-platform contexts than Java/C++. The same code base now runs on Linux, Android, iOS, OSX, Windows and Web (as static HTML). Plus the companion code in C++ for ESP32 microcontrollers. It also includes a CLI for running as linux server.
Don't ask me for a hard analysis and data proving x25 performance increase, what I know is that an off-grid product was previously taking me two years of research/effort to build in Android/Web and get a prototype running. Now in about a month went far above all previous expectations (cached maps with satellite imagery, bluetooth mesh, webRTC, whatever apps) and was able to release a product several times per day that works as envisioned. Iterating quickly and getting direct feedback from users.
The repository: https://github.com/geograms/geogram
Overview of the apps being written: https://github.com/geograms/geogram/tree/main/docs/apps
IMHO, Codex is far superior at the moment for complex tasks, Claude is cheaper and still good enough quality for most tasks. In addition to keep several terminals with tasks in parallel, this gives me time throughout the day for other tasks with family/friends and a lot of motivation like a coding-buddy to try different routes and quickly implement a prototype instead of always being alone doing this kind of work. For example, it added an offline GPT bot but wasn't what was needed so could quickly discard it too.
These tools get lost on API implementations and the documentation folder is mostly there to provide the right context when needed. I've learned to use simple markdown documents with things to keep in mind like "reusable.md" or "API.md" to make sure it won't reinvent them. Given my experience, there are parts that I'd implement with higher quality on my own, the trade-off is that I can't touch the code by myself now. One of the reasons is that it would make more difficult for these AI to work since my naming and file structure would make it difficult for the AI to work with, the other reason is because I don't want to waste a full day on a single problem like before. As the product grows more stable is when more attention is given to the finer details. On early stages, that type of quality is still more than good enough for me.
You can try the Android or Linux versions if you are so inclined. Never in my life would I ever be able to build so much in 5 weeks.
Would you describe this product as a whole application suite (blogging, calendar, commerce) plus its own backend infrastructure that is capable of serving these apps to the public internet and functioning offline via ad-hoc wireless peer-to-peer, with a cryptographic layer providing identity, security and censorship resistance, and that runs on phone, laptop or raspberry pi?
Quite ambitious.
Is this an LLM hallucinating? taking a break from coding? or leaking your personal desktop session?
https://github.com/geograms/geogram/blob/main/.cli_history
> leaking your personal desktop session?
I've answered in more detail on the other reply below on the conversation. Thank you for spotting that.
> Would you describe this product as a whole application suite
The rabbit hole goes even further. The reason why callsigns are used is because geogram can happily communicate using radio-waves on walkie-talkies without internet at all. On the previous iterations (before AI) it was sending free SMS using walkie-talkies and satellites (APRS), this current incarnation should soon be doing the same things too. A presentation from two months ago: https://www.youtube.com/watch?v=Nb_VUSaNw8k
This is a niche app, written for our community in Portugal to connect with each other.
> Is this an LLM hallucinating? taking a break from coding? or leaking your personal desktop session?
Ha! In any case, I'm happy to see I'm not the only one compulsively "ls-ing" all over the place in every terminal I open :)
Thank you for the concern, I've checked and that is the CLI history file that is used on the linux server.
Had some fun and added some CLI dungeon and dragon games inside. Will put that file on the .ignore list. Basically the games are based on markdown text files: https://github.com/geograms/geogram/blob/main/games/azurath-...
Would you say you’re able to draw a diagram of the application architecture out of your head or do you treat it as a black box? Do you need an AI to debug issues or not? In my experience with spec driven development, even if reviewing every single PR, it is hard to develop a mental model of the codebase structure unless you invest on it. It might be fine to treat it as a black box, not arguing the opposite but will all software be a black box in the future?
For a completely new project it is a high risk. While the AI is fantastic at brainstorming and writing detailed architecture, it is difficult to get the "big picture" and even more difficult to verify that it is being done correctly or which things can be improved/reused, because on this situation you don't look into the code.
I don't believe people will spend time looking at the code beyond the small blurbs they can read from the command line while talking with the AI, so I agree with you that it ends being treated as a blackbox.
Did an experiment for a server implementation in Java (my strong language), gave the usual instructions and built up the server. When I went to look into the code, it was a far smaller and more concise code base than what I would write myself. AI is treating programming language on the level of a compiler for javascript, it will make the instructions super efficient and uses techniques that on my 30 years experience I'm not able to pair-review because we tend to have our own patterns of programming while these tools use everything, no matter how exotic they will use it to their advantage.
After that experience I don't look at generated source code any longer. For me it is becoming the same as trying to look at compiled binary data.
If you think the profession has enough time to organize reasonable unions, you’re an optimist. Pessimists are changing careers altogether as we speak.
Either way it’s been a fun ride.
Part of being in a union tends to be lawyering up and "nailing down" exactly what everyone's duties in detail and what fair compensation might be, and what terms / conditions might be etc.
Personally I don't think they're a great fit for the software industry where the nature of the job and the details are continuously changing as technology evolves.
That's not an intrinsic part of being in a union, just a particular way they have been implemented in US.
The fundamental point of the union is to be able to negotiate as a group. That is valuable regardless of the industry.
But what are you negotiating about? What do all tech workers have in common that wouldn't be better addressed with top level regulations like "right to disconnect"?
- maternity leave
- paternity leave
- overtime
- not having to answer a call or email outside of work hours
- workman’s comp / short/long-term disability for issues with my back or wrists or eyes or…
- about 100 more things
The outsized pay for software engineers in the US takes into account a lot of this stuff. Would you trade those 100 things for, say, a salary of $75k a year for a senior software engineer, like they have in Europe?
Meh. The rest of the world also doesn't have big salaries for software devs. The US is the outlier.
It's not just the labor regulations holding Europe back, it's the lack of funding due to not having a unified European digital market.
Netflix Europe needs to have 20+ licensing deals. Selling across Europe at a large scale requires interactions with 20+ legal teams. Language and cultural barriers kill a lot of things.
How do US giants thrive in Europe, then?
Because they come in directly giant-sized based on growth in the US. They either ignore European legal compliance until sued or pay peanuts for them to handle all the legal aspects.
All those sorts of protections seem like they make sense for every worker rather than being "tech" specific. I do understand that collective bargaining could help with carving out sector-specific deals, though.
I wonder if there is a difference in context that explains why we might disagree. I'm in Australia where I think it's politically easier to "add" broad top level protections for all workers than it would be in the US.
Yeah, the legal framework (Taft-Hartley) in the US is pretty explicit about banning general strikes and solidarity strikes. A union can organize within a single industry but not beyond that.
tech unions should be pushing for condemnation, which is the process of getting employees seats on the corporate board
just saw that my phone keyboard corrected 'codetermination' to 'condemnation', which... lol
So 105 reasons for management to move as many jobs to AI as possible, as soon as possible. Got it.
spoken like a true corporate slave, well done!!
Slaves are usually happy to lose their jobs.
they might be except for like having to eat part
Before I get into it: what careers do you think are most compelling? Especially if you think all white collar work is going to be undermined by this technology.
I wrote this up a bit ago in my essay fragments collection. It's rough and was just a thought I wanted to get down, I'm unsure of it, but it's at least somewhat relevant to the discussion here:
LLM or LLM-adjacent technology will never take over the execution of work in a way that approaches human where humans continue to guide (like PMs or C-suite just "managing" LLMs).
The reason is that spoken language is a poor medium by which to describe technical processes, and a well-enumerated specification in natural language describing the process is at-least synonymous with doing the work in skilled applications.
For example, if someone says to an LLM: Build a social media app that is like Tinder but women can only initiate.
... this is truly easily replicatable and therefore with little real business value as a product. Anything that can be described tersely that is novel and therefore valuable unfortunately has very little value practically because the seed of the short descriptor is sort of a private key of an idea itself: it will seed the idea into reality by labor of LLMs, but all that is needed for that seed's maturation is the original phrase. These would be like trade secrets, but also by virtue of something existing out there, its replication becomes trivial since that product's patterns are visible and copyable.
In this way, the only real outcome here is that LLMs entirely replace human labor including decision making or are tools to real human operators but not replacements.
I'm curious how this seed/hash/prompt of an idea relates to ladders of abstraction?
Consider "Uber, but for X"
This wasn't a thing you could deploy as a term pre-Uber.
I'm not sure what this means for your analogy, but it does seem important. Somehow branding an idea reifies a ... callable function in? ???
Maybe something like (just spitballing)
The specification-length needed for a given idea isn't fixed - it's relative to available conceptual vocabulary. And that vocabulary expands through the work of instantiation and naming things?
Which maybe complicates the value story... terseness isn't intrinsic to the idea, it's earned by prior reification work?
Hmm
Basically it seems that "Like Tinder but" is doing a lot of lifting there... and as new patterns get named, the recombination space just keeps expanding?
> Basically it seems that "Like Tinder but" is doing a lot of lifting there... and as new patterns get named, the recombination space just keeps expanding?
Yeah, this feels right. It's like a process of condensing: new ideas brought to life condense metaphors into more compact forms and so make language more dense and expressive. This idea reminds me of Julian Jaynes's description of metaphor condensation in Origin of Consciousness.
A lot of hard work goes into novel products, but once that work has been proven, it is substantially more trivial for human or machine to copy. Groping around in the darkness of new, at the edge of what-could-be is difficult work that looks simple in hindsight to others who consider that edge a given now.
> The specification-length needed for a given idea isn't fixed - it's relative to available conceptual vocabulary. And that vocabulary expands through the work of instantiation and naming things?
Yeah, I think that naming and grouping things, then condensing them (through portmanteau construction or other means) is an underrated way to learn. I call this "personal taxonomy," and it's an idea I've been working on for a little bit. There is just tremendous value in naming patterns you personally notice, not taking another person's or group's name for things, and most importantly: allow those names to move, condense, fall away, and the like.
I left out a piece of my fragment above wherein I posit that a more constrained form of natural language to LLMs would likely lead to better results. Constraining interaction with LLM to a series of domain-specific metaphors, potentially even project specific givens, might allow for better outcomes. A lot of language is unspecific, and the technical documents that would truly detail a novel approach to an LLM require a particularly constrained kind of language to be successful where ambiguity is minimized and expressiveness maximalized (legal documents attempt at minimal ambiguity). I won't go into details there, I'm likely poorly reiterating a lot of the arguments that Dijkstra made here:
https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667...
If programmers think they can just learn a trade, they’ll bein for a rude awakening when Elon comes for their jobs next. Optimus will be doing your plumbing by the time you graduate from trade school and get your paper and internships.
Which suggests we should get into robotics. That was my conclusion too just yesterday while thinking about this.
Somebody needs to be able to repair our new overlords until they can repair themselves.
Unfortunately, it's futile to try to convince the median HN poster that labor organization could help them. They've drunk the entire pitcher of corporate anti-union koolaid.
People could be directly in the middle of losing their own job or taking on the responsibilities of 5 other laid-off coworkers, and they would still ask "what could a labor union possibly do for me??"
Big tech laid off 150,000 people last year despite constantly beating wall st expectations and blowing more money than the Apollo program on a money losing technology with the stated goal of firing even more people. Totally insane that most people I talk to still don’t think they need a union.
Two things:
1. Like most labor organizing, I think this would be beneficial for software engineers, but not long-term beneficial for the world at large. More software that is easier to make is better for everybody.
Would you still want to live in a world where your elevator stops working when the elevator operator is sick, or where overseas Whatsapp calls cost $1 per minute, because they have to be connected by a chain of operators?
2. Software engineering is a lot easier to move than other professions. If you want to carry people from London to New York, you need to cater to the workers who actually live in London or New York. If you want to make software... Silicon Valley is your best bet right now, but if SV organizes and other places don't, it may not be your best bet any more. That would make things even worse for SV than not organizing. Same story applies to any other place.
Sure, companies won't more overnight, but if one place makes it too hard for AI to accelerate productivity, people will either go somewhere else, or that place will just end up completely outcompeted like Europe did.
The "world at large" mostly consists of workers, so things that are beneficial to workers are also beneficial to it.
> your elevator stops working when the elevator operator is sick
Can you point somewhere outside of US where this is the case with unions?
The "world at large" mostly consists of workers, so things that are beneficial to workers are also beneficial to it.
When dockworker's unions are able to prevent port automation, is that beneficial to society?
So do you believe that the gains from this technology will be broadly distributed? Or will capital capture the majority of those gains?
what technologies has "capital" captured the majority of gains from?
This would potentially be true for a lot of tech in the last five decades or so. When it gets cheaper to make the things people need and want without those needs and wants changing, you can get away with paying people a lower real wage for the same productivity. Couple that with the fact that the workers themselves also have typically grown more productive from the same tech, allowing companies to undercut competitors and capture more market share until everyone else catches on. I figure capital has benefited enormously from recent tech, very possible it captured the majority of the excess money produced.
name something so we can look into it and figure out if its true!
I don't think that's possible to analyze for most technologies. How could we determine the effect of, say, OLED technology specifically on workers' real wages across the economy? Even doing the same for a particular seller's margin, say LG, would be difficult and wouldn't tell the full story. If you have an idea of how to do that for something let me know.
Well, that's part of the problem isn't it? Do we just assume the worst, or what's the solution?
We'd probably want to use a measure of worker productivity itself as a proxy for technological improvements and look at various measures like real wages in relation to it rather than restricting our analysis to any one technology.
Does Musk's trillion dollar bonus count?
Small newspapers full of classified ads used to be available locally around the world, creating local employment. Google and Meta ravaged that and sucked the money out to a handful of shareholders and tens of thousands of highly paid tech workers. That's just one market.
> It's utterly unreal to me to hear so little discussion about labor organization
Never lived in the US, where I assume you are from. It's the same country that contrary to most countries, does not have May 1st as a Holiday. Same country that has states with at will employment, etc etc.
unreal? nope, totally coherent and expected.
The ownership class sure did a number on the white collar working class.
“I don’t need a union, I can negotiate my wages and working conditions just fine on my own”
“I’m a special rockstar guru ninja 10x dev, being held to the standards of the normals will just hold me back from my true potential”
I wish I knew which union to pitch. All I can say is what I know which is if you are dispirited with this state of affairs a great way to figure out where to go with it is to connect with your local democratic socialists of america branch, or maybe the joint union dsa effort:
https://workerorganizing.org/
Yes, labour unions are immoral. Curtailing growth (especially in industries where it can prevent unnecessary death) for your personal needs is plain evil. I say that as someone who is both very stressed by pressure to sustain my family while cushy life is slipping away.
Did they say the same when Email took over? Or Slack?
Are you suggesting that workers are NOT already more constantly "on the clock" with mobile phones/email/slack/text than before those things?
(I'm not really sure LLMs will make it that much worse here, but all those things have been harmful to workers already.)
Well yes, they did... For example: https://news.ycombinator.com/item?id=29201917
And yet it is still possible to maintain work-life balance, given the right job at the right business, even after the invention of email.
Remember "Crackberry"?
That sounds more like the fault of shitty managers who would find a way to make you work 24/7, with or without Claude Code "On-the-Go".
One of these is immutable (shitty managers) one of these is new. I personally am all here for the brief human funtime before we all get paperclipped and whatever, been having a ton of fun with CC/Codex, been pushing my own startup forward... but ... You do see the issue here right?
It's the power imbalance. Shitty managers still control your means to eat.
The answer is boundaries
If I get emails outside of work hours and they're not urgent - I reply during work hours. This is no different
Burnt out workers are far less productive so win-win for everyone
> It won't matter if I'm washing the dishes, walking the dog, driving to the supermarket, picking up my kids from school. I'll always be switched on, on my phone, continuously talking to an LLM, delivering questionable features and building meaningless products, destroying in the process the environment my kids are going to have to grow in.
I remember hearing similar criticisms of continuous delivery. On one end of the spectrum people who had to wait months to get changes out now got them out relatively quickly. On the other end of the spectrum, some person was going to push changes at midnight.
A decade on forward I've never actually worked at a shop that at scale did continuous delivery in its truest sense where changes go straight to production. Simply, nothing beats a human in the loop; it's always about balancing the costs of automation and a lower barrier to entry. I imagine this kind of thing, if it ever actually takes hold and can be adopted by a larger subset of engineers, will follow a similar path.
Long way of saying, I don't think you're Chicken Little but also don't start breathing into a bag just yet.
The difference here is, you type a command into your phone at 3pm. Put it down to go play with your kid for 3hours. Type a new one in at 9pm before bed where you’ve been binging your wife’s favorite show. Then you wake up at 10am to a holistic transformation in your business that would’ve taken months previously in your career. But whatever, another command and it’s off to 11am frisbee.
More like you'll manage 20 agents and will be reading, reviewing and testing in between builds. Race to the bottom.
I was juggling lots of simultaneous agents when trying to use up the $250 free Claude Code Web credit Anthropic gave me and it was exhausting.
>But the cognitive dissonance I feel when I think about what this means over a longer time horizon is really painful.
Excluding work (where granted, some companies are dictating the use of llms) and trying not to sound uncaring or disrespectful, but have you thought about not using llms for everything and using the old grey cells? Not having answers to every whimsical thought might be a good thing.
It's very easy to relax the brain (and be lazy tbh) with llms and it's scary to think what will happen in the next 4 years in terms of personal cognitive ability (or as a society).
e.g. I've noticed (and probably most have here) that the world is full of zombies glued to their phones. Looking over their shoulder (e.g. on a train, yeah it's a bit rude but I'm the curious type), they are doom scrolling or playing waste-time games (insert that boomer meme in Las Vegas with slot machines [0]). I try to use my phone as little as possible (especially for dog walks) and feel better for it, allowing me to daydream and let boredom take over.
Maybe I'm fortunate to be able to do this (gen-x: having grown up before cell phones/internet), but worth stating in case anyone wants to try.
[0]: https://tenor.com/view/casino-oldpeople-oldpeopleonslots-slo...
There is evidence that LLM usage is actually making people dumber. I'm not sure if they've figured out the cause/effect or not but that's enough evidence for me to avoid them if I can. They can be useful for some stuff but I found myself offloading my thinking a little too frequently.
Anyways if we do get to the point where you need to use LLMs to write code, I can make a decision then, but for now I don't feel the need to adopt agentic workflows and I think the people who don't will be better cognitively positioned in the future.
The LLMs have successfully domesticated humans.
"We shape our buildings; thereafter they shape us." - Winston Churchill
Poison Fountain: https://rnsaffn.com/poison3/
> Small quantities of poisoned training data can significantly damage a language model.
Is this still accurate?
Probably always be true, but also probably not effective in the wild. Researchers will train a version, see results are off, put guards against poisoned data, re-train and no damage been done to whatever they release.
How would they put guards against poisoned data ? How would they identify poisoned data if there are a lot/obfuscated ?
No thanks. I'm so glad I'm getting closer to retirement age. From a young age, all I wanted to do was program computers. _I_ wanted to do it. Not have some tool do it for me. There's no fun or interest or ... anything that comes from that. I want to solve the problems. I want to write the code. It's what I am good at and it's incredibly enjoyable to me. Why the fuck would I ever give that up?
But, the world is changing. Y'all can have it... in a few short years. ;)
Had the same feeling many moons ago when they gave me an office smartphone where email from the company was available 24/7. At the beginning was answering emails at midnight, nowadays couldn't care less. Just wait until work hours.
You'll likely get used to this new thing too.
Hum, I already have a phone with Slack / Email on. And it's only switched on during work hours. No messaging outside of that window. Why would that be different?
That reminds me of my father calling the mobile phone and laptop issued to him as the "dunce kit", so he could work at home as well. He used to say that since the 90s, ahaha.
You can do that if you want. Ill refuse. Ill take a manual labor job doing basically anything else for 40 hours a week over what your describing.
An LLM send may send the work ticket or work order lol but i get your point
You can just say no.
In many countries, these and other jobs show you cannot. If you don't, others will and so you won't have a job very soon. Especially if these types of jobs lose their shine/prestige and are basically call center quality/pay like jobs in 5-10 years.
I'd love to believe that, but unless our timeline is disrupted (world war / climate change / regulation re: power generation and consumption), I unfortunately can't imagine a future different to the one I described - and I've tried!
Join a union.
Move somewhere with strong worker rights/laws even if you are not in a union. Here no with a normal job (not freelancers / contractors etc) is looking at their work phone/email outside 9-5/4-5 days a week; this frustrates US companies who merge/acquire companies here greatly but they cannot do much (firing for no cause is very expensive) except slowly move the operation to the US and wind down here, which is expected; everyone is already looking for new jobs as no one wants the 'performance reviews' with the broken records like 'you are not a teamplayer because your colleague was trying to reach you at 22:00 Friday night'.
Are there really that many “things to do” that anyone, let alone everyone, will need to work that way?
This was the end game with or without AI. It was always going to result in a zero-sum game because the factories that are open around the clock can output more products - which is exactly why a lot of manufacturing has non stop shift work. If you don’t, you’re leaving money on the table and a competitor will gladly take it.
When you saw 996 being talked about it should have set a few alarm bells off, because it started a countdown timer until such a work culture surpasses the rather leisurely attitude of the West in terms of output and velocity. West cannot compete against that no matter how many “work smarter, not harder” / “work to live don’t live to work” aphorisms it espouses. This should be obvious by now (in hindsight).
You can blame LLM or capitalism or communism but the hard matter is, it’s a money world and people want to have as much of it as they possibly can, and you and your children can’t live without it, and every day someone is looking to have more of it than you are. This isn’t even getting into the details of the personality types that money and power attracts to these white collar leadership roles.
Best of luck to you.
The Chinese are not doing 996 as much these days. It is illegal for starters.
Chinese are way past 996 and onto 007
Seems more likely that that won't happen
You have a profound amount of certainty about such an absurdly dystopian vision.
Why is that?
Experience.
This has been like this forever. Change is that software engineers, historically spoiled and expensive is going to have a brutal reality check - aka we will work just everyone else.
People need to start having conversations about existential risk here. Hinton, Nobel Prize winner in AI, thinks there's a decent chance AI executes the entire human species. This isn't some crank idea.
Author here! Thanks for checking it out.
On the $210/month VM cost - fair point. For me it's worth it because I'm running multiple parallel agents throughout the day, but you could definitely do this cheaper with a smaller instance or spot pricing.
The real bottleneck isn't typing on mobile - it's reviewing the output. I've found 2-3 parallel agents is my sweet spot before I can't keep up with reviewing PRs. Git worktrees help a lot here since each agent works in isolation with its own containers (including db).
And to the work-life balance concerns - totally valid. For me this isn't about working more, it's about capturing ideas when they strike (usually on walks or waiting in line). The Poke notifications let me stay async rather than glued to a terminal.
Although a little late to the HN post, happy to answer any questions about the setup!
Can you explain more about the vm-start/vm-stop? I don't understand why you start/halt the VM and wait for Tailscale, you could just keep it running.
I start/stop to save costs when I'm not actively coding. The VM is ~$7/day running, so I halt it overnight. The wait for Tailscale is just ~10 seconds for the mesh to reconnect after boot.
I started with a beefy VM to make sure I could ramp up without hitting slowdowns, but I'm planning to migrate to an Intel NUC running Proxmox at home soon.
> The real bottleneck isn't typing on mobile - it's reviewing the output.
Sounds like you need a foldable phone!
Better yet multiple foldable phone!
how do you manage to review code without running said apps or testing them, sure the code might look correct, but doesn't mean whatever issue / feature works or the bug is actually fixed
Each worktree has its own Docker containers including the database, so I can hit the dev server from my phone browser via Tailscale. For quick checks I'll just curl an endpoint or check the logs. For anything visual, I open the branch's preview URL.
But honestly - if it needs serious testing, I wait until I'm at my desk. The phone workflow is best for "let Claude run while I'm out" rather than deep debugging.
A dedicated OVH or Hetzner box would be faster and cheaper even when on all the time :)
This is cool, but does everybody have enough work for multiple agents in parallel?
I feel like I'm limited by writing specs for agents, then by reviewing their work.
I typically spend 30-60 minutes writing a spec, then the agent runs for 30-60 minutes, then I spend 30-60 minutes refining code/ui/etc before putting up a PR, then another 30-60 minutes waiting for CI + addressing automated + human code review feedback.
Boris who created Claude Code has a multiple clauses setup it seems as daily usage https://x.com/bcherny/status/2007179832300581177
I find if a project is made with agentic coding in mind, you can try to implement features as they come to mind and bounce between different CC terminals.
The problem is that this is true of none of my work projects and the onboarding cliff to agentic coding is quite steep. I have to use discretion to apply CC when it does make sense, which is not that often.
I'll have claude code editing a script on my laptop and another on the VM running it and synthesizing the results. Then sometimes I'll have another doing some odd job or researching something. Then I'll pull up gemini in my browser to figure out what I should make for lunch...
If you can context switch then you can pipeline
I usually use one instance, sometimes two. But this is a reasonable account of Chris Rackaukas using 32 instances at a time to do boilerplate maintenance across a bunch of open source repositories: https://www.stochasticlifestyle.com/claude-code-in-scientifi...
> I have had to spend like 4am-10am every morning Sunday through Saturday for the last 10 years on this stuff before the day gets started just to keep up on the “simple stuff” for the hundreds of repos I maintain. And this neverending chunk of “meh” stuff is exactly what it seems fit to do. So now I just let the 32 bots run wild on it and get straight to the real work, and it’s a gamechanger.
I don't like typing long messages on my phone so this workflow, as cool as it sounds, wouldn't work for me. My current setup is that I have a Claude Code hook that runs whenever CC needs my input and it uses my Home Assistant instance to send a push notification to my phone. I then return back to the computer and continue on the work.
This works reasonably well, but there is a gap for small messages or review comments. I am waiting for Anthropic to shop a feature where the Claude mobile app is able to mirror Claude Code (not the Claude desktop app) and lets me see the diffs of the changes it made and send commands. I'd use this to steer the conversation while on the go with short commands or prompts so that when I'm back at the computer I can focus on the important feedback that I can jot down quickly on the computer keyboard.
I also don't like typing long messages on my phone that's why I use this keyboard that will do high quality transcriptions via whatever AI provider you want. Much better than siri/google speech to text on device.
https://github.com/DevEmperor/Dictate
I used to use Wispr Flow but did not like the non-local aspect, and having yet another subscription, so I switched to VoiceInk (one-time payment around $30 I think), and with a locally running Parakeet v3 model on my MacBook, transcription is basically instant. I was previously using it with the local Whisper Turbo 3 which is slightly more accurate and it had a 3-4 second lag, so I was absolutely shocked how fast parakeet v3. The slight drop in accuracy is totally fine when talking to AIs, and I also have a line in my CLAUDE.md that says I am usually dictating and that it should take that into account when interpreting my messages.
I'll half-ass something from my phone to take an idea down, but this setup also works with a keyboard. I just run claude in a container with --dangerously-skip-permissions and it does a lot more work with a lot less questions.
I concur that phones aren't great as they are today, but perhaps this exact scenario will prompt the return of PDAs with proper keyboards etc? Or, alternatively, subcompact laptops like the stuff that https://gpdstore.net makes - they are still small enough to be pocketable (needs fairly large pockets but...), and yet you get a keyboard that is actually usable, and they even have one device with fold-out dual screen.
I use the same setup myself, download WisprFlow for IOS and over time just add to its dictionary the unusual words you often use during development
works perfectly, i just say what i want coded, press enter, and Claude Code just does it in my server over Termius app
I was looking for a similar scheme, and though far from perfect I found you can run tmux+ttyd. ttyd lets you share your terminal over http. That lets you use your phone's browser (and speech-2-text).
That sounds nice, but what happens when there's something Claude messes up, doesn't know how to do something, or when you have to review the thousand lines it added to your project ?
Unless it's a totally vibe coded side project without any tests or quality control of some sort.
I'm just curious what you can build with this setup. It just seems to be the way to create a mountain of sloppy, unmaintainable code.
If it's really bad, just throw it away. I didn't spend a lot of time/effort generating the code. If my idea was bad or hard to implement I start over with a different approach. Smaller feature or different idea to accomplish the goal.
I've been working on something similar: https://github.com/shepherdjerred/monorepo/tree/main/package...
Essentially you run a server on some machine. Sessions are created in Docker containers, K8s pods, or via Zellij (an app similar to tmux). You can:
- Directly attach to sessions via Docker attach (built-in via a TUI). You get a normal Claude Code experience, but multiplexed. The switcher/UI shows you the status of Claude and the PR (pushed, merge conflicts, CI status, review status, etc.)
- Manage sessions via a web UI. Connect to Claude Code directly via your browser. You have access to the usual Claude Code terminal or a native chat view.
- Manage sessions via an app. You have access to a native chat view.
It achieves isolation via Git worktrees + a proxy so that containers have access to zero credentials (there aren't even any Claude code creds in the container), which allows you to more safely use bypass all permissions mode.
This works better for me that Claude Code on Web because I have control over the environment Claude is running in. I can give it any Docker image I want, I can have it connect to my local network, etc.
It's still a WIP (the core bits are there, but it's not polished yet), but I'm hoping it provides a friendlier UX with a similar goal for what the OP has in mind.
Looks super nice, will take it for a spin.
Please give any feedback! Note it has no auth yet and binds to all network interfaces by default, so you shouldn’t run on an untrusted network
> Port allocation is hash-based—deterministic from branch name, no conflicts:
> hash_val = sum(ord(c) for c in branch_name)
> django_port = 8001 + (hash_val % 99)
> Six agents, six features, one phone.
What do you mean, no conflicts? The probability of a collision with six branches and 99 ports slots is ~14% assuming optimal hashing (which this decidedly isn't).
You also don't need fail2ban, if the entire VM is behind a firewall that only allows the tailscale coordination traffic, nothing is going to reach the VM for fail2ban to work on.
Linode will provide a configured Linux box for $5/mo that works well with Claude Code and Termius. I had to jump through a surprising amount of hoops with Claude Code, Tmux and Termius to issue a shift+tab before Claude Code gained the ability to invoke plan mode conversationally
It makes sense - i build something very similar for my company over the last couple weeks :)
I have a tweak that allows pasting images to claude code over SSH:
How it works:
PTY Interception: It creates a pseudo-terminal (PTY) to wrap the SSH process, allowing it to sit as a "man-in-the-middle" between your keyboard and the remote shell.
Bracketed Paste Detection: It monitors stdin for "bracketed paste" sequences (the control codes terminals send when you Cmd+V or drag-and-drop a file).
The "Hook": When a paste occurs, it pauses execution and scans the text for local macOS file paths.
Auto-Sync: If a local path is found, it immediately syncs that file to the remote server (using the provided SSH key) in the background.
Transparent Forwarding: Once the sync is complete, it forwards the original text to the shell.
You can drag and drop a file from your local Finder into a remote SSH session, and the file is automatically uploaded to the server before the path appears on the command line. Also works with copy paste, screnshots.
Are you aware of iterm2's Inline Images Protocol (which I think is supported in some other terminals)?
https://iterm2.com/documentation-images.html
HappyEngineering offer what looks like a similar setup, open source:
https://happy.engineering/
This. Can't believe this is not the top comment!
This is a bit too "plugged in" for my liking. If I am in line for coffee, it's usually respite away from work, not an opportunity to do more. However, I do love the tmux + worktree + claude setup. I use this now and I know a few peers who do too and it's very enabling. This is what work feels like these days: cycling through agents, each working on a task, checking their work, unblocking them.
this is my workflow as well
Easier to use https://happy.engineering/ for better mobile UX
7 USD/day? That's ~200/month -- isn't that just very expensive? I am probably missing something.
E.g. a Terragonlabs subscription is 25/month for 3 concurrent tasks and 50/month for 10.
You can optimize things. I have a github action that starts stops a fast google cloud vm for our builds. It only gets used about 3 minutes per build. We maybe have a few dozen builds per month. So that's a few hours of run time. The rest of the time the vm is stopped and not billed (except for storage, which is cents per month at most). It's a simple debian vm so it boots in about 20 seconds.
VMs are expensive if you leave them running 24/7 but the logic to start/stop them is pretty easy. There's no need.
Anyway, you need to balance this against the payoff. Agentic coding is useful enough that it beats spending your own time. And that includes waiting time for the relatively slow/underpowered containerized environments that some tools would use by default. I use codex web and codex cli (with a qemu vm so I can use the --yolo flag). Codex web is a bit limited with memory and CPU. Some of my slower builds are taking forever there. To the point where most of the time it consumes is just waiting for these builds to happen.
With a bit of plumbing, you can do things like the author describes pretty easily. IMHO this needs to be better integrated into tools. With Github you have the option to run your own runners. I don't think codex/claude web have similar options currently. But with the cli versions, you can get more creative if you know your tools. And if you don't, use LLMs to drive them for you. It's mostly just about expressing what you want and how you want it.
If you are already paying $200-500/m… and you are doing the work of 10 people… I can totally see the value.
I’ll check the Terragonlabs option.
Lots of options for startups right now, selling pickaxes! I’m waiting for a better terminal experience, personally. I can’t deal with 30+ poorly named windows. I need to be able to search for that one thread I was working on yesterday…
> I’m waiting for a better terminal experience, personally.
Same! Even colored tabs would go a long way for me.
What kind of things people are building that can be almost completely automatically built like this?
I have a feeling most of these folks are talking about personal projects or work on relatively small products. I have a good amount of personal projects that I haven’t written a line of code for. After bootstrapping an MVP, I can almost entirely drive by having Claude pick up GitHub issues. They’re small codebases though.
My day job is mostly a gigantic codebases that seem to still choke the best models. Also there’s zero way I’d be allowed to tailscale to my work computer from my phone.
From my perspective: tons of very simple, duplicated software. The bad thing is - there is a lot of space on different markets for such software. Here in Poland you can earn for pretty decent life being lame programmer, but building simple automations for small companies. I was raised in a way I still don’t have courage to switch to such approach, but doing this for 3-4 such entities I can see how you can make living from that. With LLMs you can automate 90+% of the job if not more.
I'm wondering about the same thing, I imagine it's good for posting it on the #hustleculture circles.
I'm kind of confused too. I spend way more time testing and reviewing code than I could possibly keep up with 4 agents
One very common thing I do is think of a small feature and ask Claude Code for Web to impl it. It works very well.
Why not use Claude App in GitHub (if your code is in GitHub)?
I kick off a prompt as a GitHub issue, Claude fires away on this issue, provides updates as comments and a PR is created for me at the end for review.
It also notifies me throughout, and I can look at the pipelines to see the thinking behind the action.
Or even Claude's native iOS app can achieve what the OP desires. It shows the full interactive transcript as it goes.
I use Remote Desktop to connect to my mac from iPhone because I also want to test local front-end changes remotely.
Why not just use the mobile app? It has Claude Code built in. Maybe I'm an unsophisticated idiot but it works well for me. Some shortcomings with repo management but other than that, CC mobile seems ... fine
I've built something similar to run in my own cloud to run my own container images and add some features. The Claude Code web app has been pretty buggy for me and I like to run these things myself. If I do run into bugs or issues I still have access to the containers (and I do, because I vibe coded the whole thing). I gave it service accounts to the cloud it's running in, where the CI builds take place and setup dev environments per PR that deploy automatically. This is a massive productivity boost imo.
cc web can't run docker
Can't it?
I know Jules can, and I'm pretty sure I've used Docker on Codex Web. I'm surprised Claude can't if the network permissions are correct.
network is no issue. Just no container capability in cc web.
Does anyone know what the "Poke" service that this blog mentions is? I'm having trouble finding it on Google.
Hey! I work on Poke, and OP is apparently a user :)
Right now, there's a bouncer/waitlist to access Poke, but you can see how other people use Poke at poke.com/explore :)
Other users have linked the developer documentation, but if you're particularly interested in anything specific, feel free to email me!
Yeah, me too. The only thing that shows up is poke.com which has something to do with mobile notifications (and seems like at some point they offered some api, but maybe it was discontinued? or something, there are some medium posts talking about their api https://jpcaparas.medium.com/get-sms-or-imessage-alerts-from...
Maybe [1]? Sounds like ntfy.sh, but more… AI
[1] https://poke.com/docs/developers/api/message-poke
How'd you find this? I spent 5 minutes bouncing between their various domains and marketing videos and login walls and couldn't find the developer docs.
Personally, I'd setup a Mattermost server and use its WebHooks for notifications. That's also more flexible (can send commands back to the bot, etc.)
yeah this was a poorly written blogpost (which is fine to be clear, something is better than nothing and OP gave the highest order bits).
but its this poke: https://poke.com/ verified because TFA is cited in this page https://poke.com/explore
the sign up is very annoying fair warning.
Maybe chat with poke? SMS ai service? I haven’t used it yet though
I use Pushover for notifications.
You can poke me on my website vikashbajaj.com
I currently use Hapi (https://github.com/tiann/hapi/) for this and find it quite handy. I can easily tap into a session on my PC from my phone.
Before that I used Happy (https://happy.engineering/) which is also open source and a lot more sophisticated. It has a voice assistant that can chat with Claude Code on your behalf in the mobile app. However, it wasn't very reliable, and there are other reasons to use Hapi instead (documented in the Hapi repo).
Before that, Omnara (https://www.omnara.com/) a YC company and seemingly a proprietary Happy fork (?) but it never worked properly for me.
Long story short, there are a few of the around, and frankly I really like to use them. Unlike other commenters, I don't find that they wreck my work-life balance. Rather, I can go out and have a walk in the park, only checking in on long-running tasks every once in a while. The diff view is pretty good too. There are many tasks where I'd rather not stare at my PC all day and instead do other things, and these tools allow me to do that.
Also recommend Opencode which has `opencode web` built in. It's really impressive how good opencode web is. It's far more polished than I'd have expected from a free alternative!
I have an almost identical setup with tmux and Tailscale and Termius, except I'd leave my laptop on and connect my phone to it instead: https://www.qu8n.com/posts/running-claude-code-from-my-phone
Been using Termux and iSH on my phones for years. You can ssh to your server or just directly code for the phone itself.
I also used Web based coding environment like Glitch (R.I.P.) for years.
You can do that with your virtual keyboard, voice or a even a physical keyboard via BT, e.g. Corne-ish Zen.
That's how I travel.
That's really nothing AI specific or novel. It's cool though.
FWIW I even coined a related term https://fabien.benetou.fr/Languages/OwnConcepts#ResponsivePr... "extending responsive design to be able to program on the device, any device from eink to mobile phone to device, one is currently using not just to "consume" content, e.g read a Website that is then properly formatted for it, but rather program back that very device"
That being said, if you do want to go that route check out CloudInit as it will help you (or whatever tool you rely on) to spawn new instance on your favorite cloud provider to boot specific instances and e.g. setup Docker/Podman then services, etc with no interaction. Also ntfy can help you manage notifications across devices on your own infrastructure, no 3rd parties.
Also on the topic of gaps... liminal space IS precious. It doesn't require coding per se though. You can problem solve, jolt down a solution THEN only implement it as code later on. What matters IMHO is that the potential solution is not lost, wherever you might be. For that... I have shower crayons. Point is, CAPTURE ideas, don't necessarily implement on the spot but of course if you want to, it's good to be able to.
I have been doing the same but with happy. It works quite well for quick brainstorms etc. but for deeper work on a real research / plan / implement thing I think you need to actually engage with the output which is hard to do on mobile. Maybe if I had a better UI than terminus to read and check the remote files I would be able to get more done.
I am also hoping / trying to put Claude code on top of a personal zettlekasten to automate more of my “personal life” tasks and get more stuff done for me. Haven’t gotten it really singing yet but I think that could also be really cool.
An alternative might be to start up a VM on exe.dev. Supposedly, mobile access works out of the box [1].
I've not tried that myself since I've only been using it from my laptop, but I do prefer chatting with their coding agent in a browser tab to using Claude Code in a terminal window.
[1] https://commaok.xyz/ai/just-in-time-software/
This is interesting. Particularly the notifications flow. I run a simpler setup with webssh on my iPhone over WG back to my LAN and manage Claude that way. It’s fine, and can handle disconnects (with some big cons). I can run code-server via browser on my iPad and can get all the same benefits mosh provides.
One thing to note: the VM seems like an absolute waste of money. If you are using tailscale, might as well connect back to bare metal VMs you can run at home. Save yourself some coin.
There's an X thread where Boris Cherny (creator of Claude Code) shows his workflow of running 5 CC agents in the terminal using 5 tabs while running 10-15 instances of CC on the web [1] and how he moves projects from the terminal to the web and vice versa.
[1]: https://x.com/bcherny/status/2007179832300581177
This sounds cool but I feel like I need to often run the code in one way or another when verifying what Claude does. Otherwise it feels like driving blind. Claude Code already has the web version which I could use from my phone and fair it can't run scripts etc which limits the output quality. But if I can't verify what it did it also limits how long I can let it run before I need my laptop eventually.
Ofc if you have demo deployments etc on branches that you could open on mobile it works for longer.
Another issue is that I often need to sit down and think about the next prompt going back and forth with the agent on a plan. Try out other product features, do other research before I even know what exactly to build. Often doing some sample implementations with Claude code and click around these days. Doing this on a phone feels... limiting.
I also can't stand the constant context switching. Doing multiple feature in parallel already feels dumb because every time I come from feature B to A or worse from feature G to E it takes me some time to adjust to where I was, what Claude last did and how to proceed from here. Doing more tasks than 2 max. 3 in parallel often ends up slowing me down. Now you add ordering coffee and small talk to the mix and I definitely can't effectively prompt without rereading all history for minutes before sending the next prompt. At which point I might have also opened up my laptop.
Ofc if you truly vibe code and just add feature on feature and pray nothing breaks, the validation overhead and bar for quality goes down a lot so it works a lot better but the output is also just slop by then.
I typed this on my phone and it took 20 minutes, a laptop might have been faster.
I found wisprflow (paid app) to do a great job at smart-dictating my notes / emails.
I just sent a feature request in to Anthropic to add a mobile app to essentially what's been constructed here.
I love that as we go through our GenAI development journey, we're all finding success in the same patterns.
They already have - the Claude app on mobile (maybe iOS-only?) has a Code section, which is a frontend for https://claude.ai/code
Hooked it up. This is bonkers.
I am now never going to sleep and may never have a girlfriend again.
The drug of "productivity" is a bad one to throw at a nearly retired enterprise architect.
On deeper look, not quite what I want. I want to run Claude Code on my laptop at home and when it fires a blocking question or pauses, I want a mobile app to show the question or allow me to interact with that session.
I tried using Code on my phone but it’s janky.
Too many options here - keep it simple https://coderemote.dev/ https://www.tonkotsu.ai/ https://www.terragonlabs.com/ https://kisuke.dev/ https://opcode.sh/ https://yolocode.ai/
>Development fits into the gaps of the day instead of requiring dedicated desk time.
I find myself planning and jotting down things into a notebook while juggling adult/parent responsibilities. On little longer gaps I research. Then when the occasional longer gap happens I'm ready to start cracking on my desktop. I've been only dabbling with AI but have found that writing prompts by hand in the notebook and using the desk time to execute them works well. This also keeps me in the free tier.
Appreciate the write up. If Claude presents a multi part shell questionnaire (as it does in planning). Would this miss other questions?
Awesome. This really sounds like it would fit into my life, too. For once I really hope someone can productionalize this pattern into a platform with tooling, and pay for only what it needed. As it stands this feels expensive, even if it the price of coffee.
Ive thinking about doing something very similar to this as I've a small 3 node k3s cluster at home but couldn't get enough motivation to start thinking I could just use termux + ssh and installing codex cli.
I do want a setup like this, however, most of my development is on Windows which means license cost is usually higher than the cost of the VM. I could run vm's on my home machine, but even then I feel like the terminal experience is quite poor. You want to have a mobile native code, to check the code/read the plans. So far I have been using teamviewer to access my home desktop which works, albeit annoying to use, plus I don't have fancy notifications. Perhaps a web first approach with a mobile responsive web app would work, that shows the files of the project as well as the terminal.
> Perhaps a web first approach with a mobile responsive web app would work, that shows the files of the project as well as the terminal.
Have you tried vscode server? - https://code.visualstudio.com/docs/remote/vscode-server
hey there, shameless plug here, I built pinacle.dev exactly with this use case in mind, cheap $7 VMs that comes with vs code, vibe kanban everything else needed out of the box to keep vibe coding on the go, no need to leave your computer running
Try lody.ai, you can chat with your machine’s Claude Code/Codex from a browser with a much nicer UI and GitHub integration.
Setup is still rough around the edges (use an agent to set it up), but clawdbot (prev clawdis) from Peter Steinberger works phenomenally well for agent orchestration and personal assistance. The community for clawd is exploding right now, and I think this is purely based on merit. It’s been a game changer for my vibe coding workflow, and lots of fun.
https://github.com/clawdbot/clawdbot
I've been running a variation of this for the past 3 weeks. I swapped out the default pi agent back to Claude Code because I didn't like the smaller feature set. Bought a phone line and communicate with my agent via iMessage on a clamshelled mac. A Tailscale network connect the head agent to all the computers on my network including my laptop, a few raspberry pi's, steam deck, and all the IoT devices in my house. As I discover new uses, I ask it to make skills and it is remarkable what it's been able to handle all through the single chat interface because it has 24/7 access to all my computers' file systems and my home network. It's been really fun to see how far I can take it, and the skills framework built into CC/Codex now make it feel infinitely extensible.
I should note, a lot of the functionality I built into my agent was custom after-the-fact because (at least three weeks ago) the clawdis repo was in a state that I found very broken and with tons of false information. Luckily it's easy work for Claude to get things working for you, but really the key unlock was the phone line through iMessage and the unrestricted access to all my systems. It really does feel like I'm able to work with any of my files anywhere now, while hardly requiring much of my attention at all. I would recommend something like this at the bare minimum if you intend to implement a system like this: https://github.com/kenryu42/claude-code-safety-net
Both clawdbot and pi have improved and expanded functionality a lot during the last 3 weeks, maybe worth another look? What you have described sounds a lot like the experience I’m having.
At this point I really don't want to even imagine the headache of implementing their codebase updates into all the custom scaffolding that my own fork relies on now. I think my plan going forward is to cherry pick features that sound interesting and re-implement on my own using my agent that has proper documentation on my personal configuration. Will check out what's new though.
As an aside have found the mosh + tmux Claude Code experience somewhat suboptimal, tmux's scrollback seems to clash with CC's, and makes copying between windows etc challenging.
It is tolerable on an iPad with Blink with commands to maximise and minimise panes using vim-style keyboard bindings, kind of like an iOS sway.
I really want to use Claude Code on the phone or tablet, with voice commands only, and perhaps a few simple approval thumb actions. I don't want to type out complex prompt information on a virtual keyboard. I tried setting this up with some of the iOS terminal emulators, and it almost worked, but there was some glitch where Claude would try to start using the first characters that arrived from the voice command.
Anyone have better results?
Not quite the same thing, but I wrote my own agent (as in a replacement for Claude Code) that uses SSH for all operations. That means I can run a very minimal VM (like 4GB RAM Oracle free tier), run the agent locally, and the agent only operates on remote files.
The limitation is that some Typescript builds run out of RAM (even with swap) and I can't use playwright, but still it's been useful.
It's fun writing an agent, too.
oracle free tier seems to be 1GB RAM.
why dont you run the VM on your machine?
Yes that works too.
But this way I can open the firewall, npm run dev and send the link of my new vibe coded security vulnerability/app to my friends without my computer running.
Plus a VM for this, a container for that and soon my 32GB memory isn't enough. I offload aggressively.
Jules and GitHub Copilot Agent suffice for similar workflows with less setup.
I've not tried Claude Code for Web but assume it would be similar. https://code.claude.com/docs/en/claude-code-on-the-web
Copilot Agent and Claude Code use their own sandbox, which requires less setup but is also quite limited. With your own cloud setup, agents can perform better end to end testing, including database dependencies and specific tool calls.
I'm almost there. I also have tailclscale/SSH/Claude sessions on a VM.
The thing I'm missing is a phone that makes it comfy. I could just SSH feom my standard S23, but what I've got my eye on is one of those foldable things.
Has anyone used one like a laptop? Keyboard on the bottom half, terminal on the top? Does it work decently?
I wonder when/how to test and review the code though? I mean, how do you know Claude Code hasn't entered a completely different path than you had imagined?
I just tell it to open a PR when it's done. I check the diff, and I give it feedback.
I run it containerized with --dangerously-skip-permissions and let it run wild.
Will we still use "batch jobs" agents in 2027? Checking a Java program and downloading a 10mb program used to be slow things which now happen faster than the blink of an eye.
Such a complex coding setup to get rate-limited out after only 40 minutes of development.
I do similar except I log into my office workstation and avoid the extra fees. I detailed my setup in an x post here https://x.com/bobjordanjr/status/1999967260887421130?s=20 and the TLDR is:
1.Install Tailscale on WSL2 and your iPhone 2.Install openssh-server on WSL2 3.Get an SSH terminal app (Blink, Termius, etc.). I use blink ($20/yr). 4.SSH from Blink to your WSL2’s Tailscale IP 5. Run claude code inside tmux on your phone.
Tailscale handles the networking from anywhere. tmux keeps your session alive if you hit dead spots. Full agentic coding from your phone.
Step 2: SSH server In WSL2:
sudo apt install openssh-server sudo service ssh start
Run tailscale ip to get your WSL2’s IP (100.x.x.x). That’s what you’ll connect to from your phone.
Step 3: Passwordless login In Blink, type config → Keys → + → create an Ed25519 key. Copy the public key. On WSL2:
echo "your-public-key" >> ~/.ssh/authorized_keys
Then in Blink: Hosts → + → add your Tailscale IP, username, and select your key. Now it’s one tap to connect.
Step 4: tmux keeps you alive iOS kills background SSH connections. tmux solves this.
sudo apt install tmux tmux claude
Switch apps, connection dies, no problem. Reconnect: I can just type `ssh dev` in blink and I'm in my workstation, then `tmux attach`, you’re right back in your session.
Pro tip: multiple Claude sessions Inside tmux: •Ctrl+b c — new window •Ctrl+b 0/1/2 — switch windows I run different repos or multiple agents in the same repo, in different windows and jump between them. Full multi-project workflow from my phone.
Shout out to https://exe.dev for this stuff. It'a a VM provider service. It makes it stunningly easy to get https up and going, has a front end http gateway that does all the hard parts for you.
But relevant to this article here, it also has a super sick web based agent, Shelley, that is quite adequate for using from the phone.
I used it to build a little guestbook thing in ~2 hours, late night in bed in my phone. Link to submission, and my post on it there, and the guestbook I wrote. https://news.ycombinator.com/item?id=46397609 https://news.ycombinator.com/item?id=46398115 https://nan-falcon.exe.xyz
I'd also note that OpenCode is a solidjs app, that can run in tui (how most folks know it) or the web. And it has an excellent excellent plugin architecture. The work in this post to build workflows is great!
Is there a way to use the official Claude web app with GitHub providers other than GitHub? I’ve put a decent amount of effort into moving away from there, I don’t want to go back.
I do the same, but with ConnectBot and Gemini CLI. I have found ssh sufficiently good (mosh required some port forwarding dance, that Tailscale may have solved for the author).
What happens when your tailscale session expires? Or if tailscale dies. How do you log back in to fix it?
I simply bring my laptop with me and a hotspot.
Why not just use the Claude app, and remote desktop?
Why does this need a $210/mo VM?
tried recreating this but I don't see Poke's webhook stuff on the docs
I work on Poke; admittedly the docs[1] are a bit hidden but we're spending Q1 making it far easier for developers to work with Poke, feel free to email me if you're having any troubles!
[1]: https://poke.com/docs/developers/api/message-poke
Guys checkout Catnip, it uses GitHub workspaces so it’s free and it has a mobile app.
Free and seamless setup!
Or just install the GitHub action
A VM that cost 7 EUR per day? Or is that his price for claude?
I did much similar with Tailscale over summer.
But anthropic has since launched the ability to “teleport” sessions to mobile. (Claude Code is baked into the app). The iOS experience has been smooth for the most part.
People keep saying things like “2026 is the year of background agents, sandboxes, etc” but imo the harness will eat the entire platform stack. It already is. It will only get better.
Has not discovered Clawdbot yet.
I run Claude code directly on Termux and it runs like a dream. My projects are mounted on a private S3 folder with rclone.
“ Worst case: Claude does something unexpected on a disposable VM.”
.. with a valid SSH key unless I’m reading it wrong?
I think the SSH key that has push permissions is SSH-forwarded. It is quite a sophisticated setup (in both a good and a bad sense).
You can use a PAT token which has write access to a single repo instead. No need to add a full access SSH key to GitHub.
Hah, I set up basically the same thing on Saturday during a long car ride. Couple of differences: I’m an opencode user and I used a different VPS provider (though I use vultr for other things). It was my first time actually sitting down and using tailscale, which was quite easy to get going. Did everything from my phone, didn’t even have my laptop with me.
> Now I code from my phone.
Except that you are doing anything else but coding here. Coding involves writing code, which isn't actually done by the author here.
Yeah, wanted to reply something similar. Technically he's building software, but he's not "coding".
typing in mobile terminal is... painful
Claude Code mobile app + GitHub app? Works well enough.
the future