>Google Cloud has blocked our account, making some Railway services unavailable. We have escalated this directly with Google. The Railway Platform team has since confirmed access to Google Cloud and is working on restoring access to all workloads. We have access to some of our Google Cloud–hosted infrastructure and are working to restore the rest of the service. We apologize for the disruption.
That's true, unfortunately. Very sorry abut this, and it's my fault that we weren't resilient to this failure. We'll be migrating the webapp to a more resilient architecture as a first priority.
The landing page is up again now but unfortunately will have to default to accepting demo requests for now :(
There's very few startups that I look at these days and don't think to myself, "I could just write a Claude skill for that". This one seems pretty cool. Congrats on launch
>> Superlog scans your codebase and infrastructure to add new alerts, metrics and dashboards, preventing tricky failure modes and observability decay.
This is interesting, and my prior belief here has been that this automates a one time set up, and perhaps a quarterly clean-up or reactive monitoring changes that people do today. Curious what your experience has been - do teams accept these ongoing maintenance PRs at a good rate?
For full disclosure / context: we work in a related space - investigation agents for production issues.
Building on the "investigation > patch" point — running Claude Code, Codex, and Gemini CLI daily, the pattern I keep noticing is that auto-fix is fine on "obvious bug, obvious fix" (off-by-one, null check, missing await, error not propagated). It falls over on "subtle invariant" bugs where the existing code is intentionally weird to preserve something non-obvious — the PR looks right and breaks something three modules away.
The tool I'd actually want isn't "tries harder to fix everything." It's one that credibly says "this touches an invariant I can't see — here's what I think might happen, you handle it." Calibrated humility beats confident patches.
Curious how your high-confidence threshold actually works. Self-reported model certainty (notoriously unreliable), test coverage in the affected area, blast-radius of the change, something else?
Auto-instrumentation is great until you hit cardinality. We turned on auto-traces across our LLM gateway and within a week our trace storage costs were dominated by attribute cardinality from user-id and prompt-text labels. Pulled those out into a separate column store with sampling. Query latency dropped 40x. Curious how Superlog handles attribute cardinality at scale, especially for AI workloads where prompt content is high entropy. The 'installs itself' magic is only useful if it doesn't blow up the bill.
Interesting project - but you need to add some information on where the data goes. As far as I can tell, code goes to some upstream ai provider (for installing, for analyzing).
Telemetry goes to some provider or local hosted solution? And then to your upstream ai provider for analysis?
When you're installing Superlog, you can use any coding agent you'd like, including a local model.
Your telemetry then goes into our data stores, and right now we have one DC on the US west coast.
Whenever there's an error log or trace, Superlog can analyze it and prepare a resolution PR (or a note if something needs to be done manually).
This can be turned off and then the incident can be sent to your own models via a webhook.
We use one of the frontier models for that (it's an upstream AI provider). We're working on our own fine-tuned version of a SoTA model to minimize dependency on other AI providers.
To investigate an incident, we clone the repo in our worker, and pass the repository files to a coding agent in a sandbox. The agent has an MCP that gives it access to the telemetry (logs/metrics/traces) of the project.
The coding agent will then investigate the incident and prepare a patch. It hands over the patch via a tool. The worker then deterministically pushes the patch to a branch and opens the PR.
This way the agent doesn't have full Git access and can't do anything it's not supposed to do in the repository.
investigation is the hard part, not generating patches. we've had prod issues where the fix was obvious once you knew the cause, but finding the cause meant connecting an error trace to a config change from 3 deploys ago. if the MCP only surfaces traces and logs from one service the agent is going to propose workarounds instead of actual fixes. how deep does the investigation context actually go?
Great question! The investigation agent has access to all the telemetry - not only one service. So we can actually trace the root cause in such complex cases!
There are good ways to link operations between different services with OpenTelemetry (for example, passing the parent trace id in an inter-service HTTP/gRPC request). It's a bit tedious to do by hand, that's why we're publishing the skill that does that for you.
And totally agreed on config changes and deploy info. We've seen that having good environment and version control (commit hash, file name, line number) tagging is extremely important for root cause analysis, so we go hard on this in the skills.
We also have many infra integrations in our roadmap to make sure that we can deeply analyze the infra/config side of things.
Interesting product, but had similar question, i think it will take a little time to be mature for production systems: as what i can see right now is very straightforward, most of the observability providers are doing this, in case you already have the observability stack setup. we currently use Openobserve they have an ai agent that provides correlation, cause and fix for any issues . The real differentiator can be on how accurately you can do the investigations, and how brutally you can steelman the ability for it locate the issue, cause and fix. Good luck on the launch
Congrats on the launch, this looks very promising. I hadn't seen any installation that uses a URL to point to a skill, seems like an evolution of wizard scripts
That been said for more complex setups like on kubernetes where you need a collector and an operator I found OTEL to be super painful to setup a couple of years ago. Has it gotten any easier now?
I'm afraid a collector and the operator are still the recommended way to go by OpenTelemetry (https://opentelemetry.io/docs/platforms/kubernetes/getting-s...). We're still working on a custom skill for Kubernetes, but the general skill should give you a sane default already.
A good way to start can be to start sending traces/logs directly by instrumenting the service and putting our backend as the collector.
I also help out personally whenever our clients have any questions on setting up the telemetry :)
Before running the install prompt, I’d want to see a dry run: which files it would touch, what telemetry leaves the box, provider calls, and what “high confidence” means. For debugging tools, generating a PR is the easy part; knowing whether it’s grounded in enough evidence is the part I’d worry about.
I love the launch! Automated observability that feeds back into the product development process is the future of this category vs having to spend a lot of time configuring and managing the infrastructure yourself.
It's something we've thought a lot about at Amplitude. We'd love to talk.
Love the concept! Some feedback: I went to sign up to give it a go, but the set up process left me feeling a bit untrusting - so I backed out for now. I'd prefer more explanation about what to expect, what I will get, how it is safe, etc before asking me to run a prompt.
Right now, the prompt will enumerate all the services and install the OpenTelemetry SDK (https://opentelemetry.io/) in each service.
Then for every service, the skill will make sure that:
- Every time something breaks and an operator needs to take a look, there's an error log
- All important steps in a process emit info/debug logs (so that an issue can be investigated)
- Operations are covered with spans with relevant attributes.
- Cost (LLM tokens), API performance (latency/RED), tenant activity (cost/usage per tenant) are covered by metrics so that you can use Superlog MCP to build cool dashboards.
For most common stacks like NextJS, FastAPI, React Native/Expo etc. we have a custom skill that explains the best practices for this specific technology. For all the other stacks we ask the agent to use general best practices.
We have evals for all custom skills where we start from a starter project, run the agent with the skill and use LLM-as-a-judge to compare it to a human-written 'golden patch'.
In general, we try to:
- minimize diff, so that the instrumentation is easy to review
- make small chunks of additive diffs vs huge indents / moving logic around
- minimize new dependencies
- use well-supported and audited OTel SDKs vs custom libs
It deleted the codebase, which technically.. is a valid way to get rid of all of the bugs.
I kid, nice work. As others have said, investigation, and understanding "the why it was originally done that way", not the patch, is usually the lion share of the work.
I made the Slack onboarding step mandatory for now since we thought that a lot of our value was in sending investigations and PRs, and Slack was what we used ourselves.
What tool do you use for communication around your project? If you don't want to share publicly, could you please shoot a line to:
ash [at] superlog.sh?
Would love to learn about your usecase in more detail too!
Got it! What channel would you prefer instead? Would Telegram/WhatsApp/Signal/iMessage be good?
The platform itself doesn't need Slack to function, we just observed that users got more value if they could get notifications somehow, so I'm more than happy to add more comms platforms :)
Good question! We don't have one as of today, just because we're iterating very quickly and a cloud version is the quickest way for us to keep things lean and up-to-date, but we're not far from having one.
Could you please send me an email at ash [at] superlog.sh? I'd love to hear more about your use case - we might have something for you very soon!
Great question! I like to think about this in two ways:
1. Counter-positioning. Most existing tools have invested heavily in their web platforms and compete on their UI/UX. But actually, what matters to our clients is that bugs are fixed. Our top clients would rather never open our tool at all. If our competitors want to beat us, they essentially have to fight against their established business models that hinge on users looking at their browsers.
2. Evals. In order to have the most accurate RCA analysis you need a very good suite of evals: what was the right root cause in this bug? what is the right fix?. We're investing into this heavily, and as one of the early movers we have a big advantage here.
At the same time, I tend to approach strategy with a lot of caution. A lot of the canonical reasoning behind 'startup positioning' is based on extrapolation from trends, but surprisingly few analogies work in economics.
Our focus right now is:
- talking to our users
- making sure they have the best experience
The typical issues I have seen with LLMs / Agents tend to be reactive in their fixes. So they tend to "patch" the symptom more than "fix" the root cause. Interested to see how you solve this problem.
You're right! It's a big issue and I don't think there's a silver bullet.
We have an eval suite with code+telemetry fixtures and a golden RCA+patches and an LLM-as-a-Judge. So whenever we get feedback from our users and they're OK with it, we use their feedback to create an eval case (it's still quite manual since you have to calibrate the case).
We use Superlog to observe Superlog, so I often extract cases from our own errors. The PRs get better and better, but, of course, it's sort of a continuous improvement process.
Either it's a tool for observing or it's a tool for fixing issues, it cannot be both, by physical principle.
Best case scenario here is that the product succeeds, and then you need to instrument the product itself in order to observe it, like debugging the debugger. But it wouldn't be an observability tool, it would shift the product that needs to be observed from the previous source code that is now a target language into the new source code that is now your product.
Love the analogy! We honestly just wanted to have this product ourselves, and that was our primary motivation behind building it.
I agree with the philosophical principle! If you give a rigid observer an incentive to 'remove bugs', it will happily silence all alerts and report success.
Our goal is to make sure that doesn't happen. The investigation agent is actually a separate agent with a separate goal.
In practice, we rarely see the agent just silencing stuff. When this happens, I get on it and make it an eval case :)
It is a well known phenomenon in haskell or OOP in general that just reading can cause systems to change their execution paths. Which is why it's not sufficient to make write/setters private, but reads/getters too.
However in the case of SuperLog the path to system change is quite direct
"We fix bugs.
Superlog prepares a resolution PR for every incident. If Confidence Gate fails, it posts findings for the investigating team and pulls in the engineers who can add context."
The system literally pushes (or Pull requests, whatever, github is dead anyways) a code change.
Very good point on the confidence gate! We've rolled out feedback collection features on PRs themselves, on Slack notifications and incidents a few days ago so the data is still a bit fresh.
Anecdotally, our top clients accept 80-90% of PRs, with several clients accepting all of them and requesting an auto-merge feature. I myself accept most of Superlog PRs to Superlog. Most PRs that stay unmerged are usually due to a client losing interest in our product / abandoning the instrumented project.
Another interesting point is that not every defect is a PR. Often it's misconfiguration in an external service, so there's a special incident state for that. For example, yesterday I forgot to verify our domain on Resend so some verification emails didn't go through. Superlog pinged me on Slack and explained where to go to fix.
Super glad you like the npx onboarding and the MCP tool :) Please keep the feedback coming!
Not their fault
Railway their hosting provider is entirely down as well
From https://status.railway.com/
>Identified
>Google Cloud has blocked our account, making some Railway services unavailable. We have escalated this directly with Google. The Railway Platform team has since confirmed access to Google Cloud and is working on restoring access to all workloads. We have access to some of our Google Cloud–hosted infrastructure and are working to restore the rest of the service. We apologize for the disruption.
That's true, unfortunately. Very sorry abut this, and it's my fault that we weren't resilient to this failure. We'll be migrating the webapp to a more resilient architecture as a first priority.
The landing page is up again now but unfortunately will have to default to accepting demo requests for now :(
There's very few startups that I look at these days and don't think to myself, "I could just write a Claude skill for that". This one seems pretty cool. Congrats on launch
Thank you! super happy that's how you feel about Superlog. Let us know if you want to try it out and/or have any feedback :)
>> Superlog scans your codebase and infrastructure to add new alerts, metrics and dashboards, preventing tricky failure modes and observability decay.
This is interesting, and my prior belief here has been that this automates a one time set up, and perhaps a quarterly clean-up or reactive monitoring changes that people do today. Curious what your experience has been - do teams accept these ongoing maintenance PRs at a good rate?
For full disclosure / context: we work in a related space - investigation agents for production issues.
Building on the "investigation > patch" point — running Claude Code, Codex, and Gemini CLI daily, the pattern I keep noticing is that auto-fix is fine on "obvious bug, obvious fix" (off-by-one, null check, missing await, error not propagated). It falls over on "subtle invariant" bugs where the existing code is intentionally weird to preserve something non-obvious — the PR looks right and breaks something three modules away.
The tool I'd actually want isn't "tries harder to fix everything." It's one that credibly says "this touches an invariant I can't see — here's what I think might happen, you handle it." Calibrated humility beats confident patches.
Curious how your high-confidence threshold actually works. Self-reported model certainty (notoriously unreliable), test coverage in the affected area, blast-radius of the change, something else?
[dead]
Auto-instrumentation is great until you hit cardinality. We turned on auto-traces across our LLM gateway and within a week our trace storage costs were dominated by attribute cardinality from user-id and prompt-text labels. Pulled those out into a separate column store with sampling. Query latency dropped 40x. Curious how Superlog handles attribute cardinality at scale, especially for AI workloads where prompt content is high entropy. The 'installs itself' magic is only useful if it doesn't blow up the bill.
Interesting project - but you need to add some information on where the data goes. As far as I can tell, code goes to some upstream ai provider (for installing, for analyzing).
Telemetry goes to some provider or local hosted solution? And then to your upstream ai provider for analysis?
Thanks for the feedback!
When you're installing Superlog, you can use any coding agent you'd like, including a local model.
Your telemetry then goes into our data stores, and right now we have one DC on the US west coast.
Whenever there's an error log or trace, Superlog can analyze it and prepare a resolution PR (or a note if something needs to be done manually).
This can be turned off and then the incident can be sent to your own models via a webhook.
We use one of the frontier models for that (it's an upstream AI provider). We're working on our own fine-tuned version of a SoTA model to minimize dependency on other AI providers.
To investigate an incident, we clone the repo in our worker, and pass the repository files to a coding agent in a sandbox. The agent has an MCP that gives it access to the telemetry (logs/metrics/traces) of the project.
The coding agent will then investigate the incident and prepare a patch. It hands over the patch via a tool. The worker then deterministically pushes the patch to a branch and opens the PR.
This way the agent doesn't have full Git access and can't do anything it's not supposed to do in the repository.
I would love to use it but the website is down
"Please check your network settings to confirm that your domain has provisioned.
If you are a visitor, please let the owner know you're stuck at the station."
Would love to learn more and consider being a customer!
Apologies everyone! indeed our webapp unfortunately went down together with Railway - we're working hard to bring it up
investigation is the hard part, not generating patches. we've had prod issues where the fix was obvious once you knew the cause, but finding the cause meant connecting an error trace to a config change from 3 deploys ago. if the MCP only surfaces traces and logs from one service the agent is going to propose workarounds instead of actual fixes. how deep does the investigation context actually go?
Great question! The investigation agent has access to all the telemetry - not only one service. So we can actually trace the root cause in such complex cases!
There are good ways to link operations between different services with OpenTelemetry (for example, passing the parent trace id in an inter-service HTTP/gRPC request). It's a bit tedious to do by hand, that's why we're publishing the skill that does that for you.
And totally agreed on config changes and deploy info. We've seen that having good environment and version control (commit hash, file name, line number) tagging is extremely important for root cause analysis, so we go hard on this in the skills.
We also have many infra integrations in our roadmap to make sure that we can deeply analyze the infra/config side of things.
Interesting product, but had similar question, i think it will take a little time to be mature for production systems: as what i can see right now is very straightforward, most of the observability providers are doing this, in case you already have the observability stack setup. we currently use Openobserve they have an ai agent that provides correlation, cause and fix for any issues . The real differentiator can be on how accurately you can do the investigations, and how brutally you can steelman the ability for it locate the issue, cause and fix. Good luck on the launch
Congrats on the launch, this looks very promising. I hadn't seen any installation that uses a URL to point to a skill, seems like an evolution of wizard scripts
That been said for more complex setups like on kubernetes where you need a collector and an operator I found OTEL to be super painful to setup a couple of years ago. Has it gotten any easier now?
Thank you! Glad you liked the install process :)
I'm afraid a collector and the operator are still the recommended way to go by OpenTelemetry (https://opentelemetry.io/docs/platforms/kubernetes/getting-s...). We're still working on a custom skill for Kubernetes, but the general skill should give you a sane default already.
A good way to start can be to start sending traces/logs directly by instrumenting the service and putting our backend as the collector.
I also help out personally whenever our clients have any questions on setting up the telemetry :)
Before running the install prompt, I’d want to see a dry run: which files it would touch, what telemetry leaves the box, provider calls, and what “high confidence” means. For debugging tools, generating a PR is the easy part; knowing whether it’s grounded in enough evidence is the part I’d worry about.
I love the launch! Automated observability that feeds back into the product development process is the future of this category vs having to spend a lot of time configuring and managing the infrastructure yourself.
It's something we've thought a lot about at Amplitude. We'd love to talk.
Awesome, let's definitely have a chat! I'll shoot an email via BF :)
Love the concept! Some feedback: I went to sign up to give it a go, but the set up process left me feeling a bit untrusting - so I backed out for now. I'd prefer more explanation about what to expect, what I will get, how it is safe, etc before asking me to run a prompt.
Thank you! Very good point.
Right now, the prompt will enumerate all the services and install the OpenTelemetry SDK (https://opentelemetry.io/) in each service.
Then for every service, the skill will make sure that:
- Every time something breaks and an operator needs to take a look, there's an error log - All important steps in a process emit info/debug logs (so that an issue can be investigated) - Operations are covered with spans with relevant attributes. - Cost (LLM tokens), API performance (latency/RED), tenant activity (cost/usage per tenant) are covered by metrics so that you can use Superlog MCP to build cool dashboards.
For most common stacks like NextJS, FastAPI, React Native/Expo etc. we have a custom skill that explains the best practices for this specific technology. For all the other stacks we ask the agent to use general best practices.
We have evals for all custom skills where we start from a starter project, run the agent with the skill and use LLM-as-a-judge to compare it to a human-written 'golden patch'.
In general, we try to:
- minimize diff, so that the instrumentation is easy to review - make small chunks of additive diffs vs huge indents / moving logic around - minimize new dependencies - use well-supported and audited OTel SDKs vs custom libs
You can read the skills here: https://github.com/superloglabs/skills.
I'll make sure to add this to our landing and print this out as the agent writes the code!
Thank you for the feedback!
It deleted the codebase, which technically.. is a valid way to get rid of all of the bugs.
I kid, nice work. As others have said, investigation, and understanding "the why it was originally done that way", not the patch, is usually the lion share of the work.
Genuine question! What makes this any different than any other thing like it? Why should i use this over something else
I would love to try it but I got stuck when it asked for Slack since I dont use that.
Hm, sorry about that!
I made the Slack onboarding step mandatory for now since we thought that a lot of our value was in sending investigations and PRs, and Slack was what we used ourselves.
What tool do you use for communication around your project? If you don't want to share publicly, could you please shoot a line to:
ash [at] superlog.sh?
Would love to learn about your usecase in more detail too!
Lots of places use various slack alternatives or teams/google.
For my current project I would use webhooks/email just like I do currently for my monitoring and alerting.
I don’t use Slack either. What about solo indie founders who don’t use “team communication”?
Got it! What channel would you prefer instead? Would Telegram/WhatsApp/Signal/iMessage be good?
The platform itself doesn't need Slack to function, we just observed that users got more value if they could get notifications somehow, so I'm more than happy to add more comms platforms :)
Congratulations with the launch and it looks good, do you provide any API for other tools to integrate superlog.
on your pricing page:
> Start with one repo. Price the rest when the signal is real.
which makes it sound like possibly the $150/mo price is per-repo?
I think that could use some clarification - if I have 10 services in a monorepo vs 10 individual service repos, does that 10x my cost?
Very good point, thank you! Let me remove this phrase, you're right, it's misleading.
The pricing is only by usage (traces/logs/metrics) and investigation credits. We don't charge extra for repos :)
Congrats on the launch
Congrats on the launch
Seems very useful, congratulations on the launch!
Any plans for an on-prem version?
Good question! We don't have one as of today, just because we're iterating very quickly and a cloud version is the quickest way for us to keep things lean and up-to-date, but we're not far from having one.
Could you please send me an email at ash [at] superlog.sh? I'd love to hear more about your use case - we might have something for you very soon!
What's your moat?
Great question! I like to think about this in two ways:
1. Counter-positioning. Most existing tools have invested heavily in their web platforms and compete on their UI/UX. But actually, what matters to our clients is that bugs are fixed. Our top clients would rather never open our tool at all. If our competitors want to beat us, they essentially have to fight against their established business models that hinge on users looking at their browsers.
2. Evals. In order to have the most accurate RCA analysis you need a very good suite of evals: what was the right root cause in this bug? what is the right fix?. We're investing into this heavily, and as one of the early movers we have a big advantage here.
At the same time, I tend to approach strategy with a lot of caution. A lot of the canonical reasoning behind 'startup positioning' is based on extrapolation from trends, but surprisingly few analogies work in economics.
Our focus right now is: - talking to our users - making sure they have the best experience
This is a very interesting idea and im excited to see where this goes. Congrats!
Thank you! :)
The typical issues I have seen with LLMs / Agents tend to be reactive in their fixes. So they tend to "patch" the symptom more than "fix" the root cause. Interested to see how you solve this problem.
You're right! It's a big issue and I don't think there's a silver bullet.
We have an eval suite with code+telemetry fixtures and a golden RCA+patches and an LLM-as-a-Judge. So whenever we get feedback from our users and they're OK with it, we use their feedback to create an eval case (it's still quite manual since you have to calibrate the case).
We use Superlog to observe Superlog, so I often extract cases from our own errors. The PRs get better and better, but, of course, it's sort of a continuous improvement process.
Sorry to be crude, but this sounds either dead on arrival, or at least needing a pivot, or a rephrasing of the pitch:
The moment something changes the system, it no longer observes it, in fact observing something might cause it to change ( https://en.wikipedia.org/wiki/Observer_effect_(physics) )
Either it's a tool for observing or it's a tool for fixing issues, it cannot be both, by physical principle.
Best case scenario here is that the product succeeds, and then you need to instrument the product itself in order to observe it, like debugging the debugger. But it wouldn't be an observability tool, it would shift the product that needs to be observed from the previous source code that is now a target language into the new source code that is now your product.
Love the analogy! We honestly just wanted to have this product ourselves, and that was our primary motivation behind building it.
I agree with the philosophical principle! If you give a rigid observer an incentive to 'remove bugs', it will happily silence all alerts and report success.
Our goal is to make sure that doesn't happen. The investigation agent is actually a separate agent with a separate goal.
In practice, we rarely see the agent just silencing stuff. When this happens, I get on it and make it an eval case :)
How does a grep or read affect the observing system?
I guess the change in voltages, arrangement of registers, filling of buffers in the network stack are changing but... what?
It is a well known phenomenon in haskell or OOP in general that just reading can cause systems to change their execution paths. Which is why it's not sufficient to make write/setters private, but reads/getters too.
However in the case of SuperLog the path to system change is quite direct
"We fix bugs. Superlog prepares a resolution PR for every incident. If Confidence Gate fails, it posts findings for the investigating team and pulls in the engineers who can add context."
The system literally pushes (or Pull requests, whatever, github is dead anyways) a code change.
The observer effect is that observing itself modifies the system. This is nothing to do with that.
[flagged]
[dead]
[flagged]
[flagged]
[dead]
[dead]
[flagged]
Very good point on the confidence gate! We've rolled out feedback collection features on PRs themselves, on Slack notifications and incidents a few days ago so the data is still a bit fresh.
Anecdotally, our top clients accept 80-90% of PRs, with several clients accepting all of them and requesting an auto-merge feature. I myself accept most of Superlog PRs to Superlog. Most PRs that stay unmerged are usually due to a client losing interest in our product / abandoning the instrumented project.
Another interesting point is that not every defect is a PR. Often it's misconfiguration in an external service, so there's a special incident state for that. For example, yesterday I forgot to verify our domain on Resend so some verification emails didn't go through. Superlog pinged me on Slack and explained where to go to fix.
Super glad you like the npx onboarding and the MCP tool :) Please keep the feedback coming!
It's a slop account.