Autonomous software maintenance
Bugstack finds what breaks, proves the fix, and opens the pull request.
We are building toward the end of manual software maintenance. Spend the time on what you want to make.
What Bugstack does on its own today
- Fixes crashes
- Fixes failing CI
- Fixes failed deploys
- Fixes bugs your customers report
- Upgrades packages
- Patches vulnerable packages
- Finds committed secrets
- Runs security audits
- Answers review comments
A real fix on Bugstack's own repository: detected, fixed, merged, and still holding in production a day later.
It finds the work before anyone files a ticket.
Bugstack watches your running app and your repository. When something breaks or goes out of date, it starts on its own.
Coming soon
- Silent error finding
- Errors in your users' browsers
- Performance regressions
- More hosting providers
- Repositories outside GitHub
Every bug fix is proven in a sandbox first.
Bugstack copies your repository into an isolated sandbox and writes a test that reproduces the bug. The test has to fail on your code, then pass with the fix.
The test must fail with the bug's own error
A test that fails for anything else, a typo or a bad import, never counts.
Nothing gets in or out
Dependencies install, then the network is cut before the first test runs. Your secrets never go in.
The pull request includes the test
The proof travels with the fix, and your own CI runs on it too.
You set how much it does alone.
Every rule that can hold a change is a setting you own, for each repository.
Autonomy
- Writes the fix
- Opens the pull request
- Merges it
BugstackYou
Proof before merging
Bugstack allowed to merge
Bugstack may
First place on SWT-Bench Verified,
scored by its maintainers.
SWT-Bench measures whether an agent can write a test that fails on a real bug and passes once the bug is fixed. Writing that test unaided is a cornerstone of autonomous software engineering, and it is the step Bugstack runs before every bug fix.
433 real GitHub issues. Scored by the benchmark's maintainers, 1 Nov 2026. See the leaderboard at swtbench.com · How the run was scored
What it asks of an agent
Each task is a real GitHub issue from an open source Python project. The agent gets the issue and the code, and has to write a test that fails before the fix and passes after it.
Why that matters
An agent that can reproduce a bug on its own can show that its fix worked, with nobody checking by hand. Every proven fix starts with that test.
How Bugstack compares
Bugstack writes its tests with Sonnet 5.5 and uses Opus 5.5 to pick the best one. Claude Code scored 87.8% on Opus 5.5 and 75.8% on Sonnet 5.5.
43 of 44
fixes held on an exam of bugs we planted in our own code, sealed before we tuned the fixer.
245 of 252
held on the full set of planted bugs, in TypeScript, Python, Go and Ruby.
Hand off the upkeep.
Go build.
$100 a month. Start with a 14-day trial: $50 of credit and a first security audit on us.
Or book 15 minutes with the founder