The pain
“Done” stopped meaning done
Code nothing calls. Tests that only agree with themselves. Fixes that break under a retry. All green, all reported finished.
Why AI-generated code fails →● Guardrails for AI coding agentsAgent says done.
It doesn't work.
Your agent ran the suite, saw green, and told you it was finished. It wasn't. Cotterpin is guardrails for AI coding agents: it checks every “done” on your machine, before the commit, so AI slop never reaches your code base — or your users.
Start free
$ pipx install cotterpinfrom 0.3.0
No credit card required Runs on your machine
Works with the agents you already use
- Antigravity
- Claude
- Codex
- Grok
Demo: Cotterpin fixing a Java payment bug
A developer asks Cotterpin to fix customers being charged twice when the payment provider retries a webhook. Cotterpin analyses the code, writes a test that reproduces the bug, lets the agent fix it, validates the build and tests, then sends an adversary to break the fix. The adversary finds a race condition, the agent fixes it again, the fix survives, a receipt is signed, and the branch is pushed with a pull request opened for review.
Shift left
Catch AI slop where it starts: in the agent session, before the commit
Agentic coding is a new way of working. CI and code review sit after the commit, once a problem has left the developer's machine. Cotterpin moves quality and security as close to the developer as they can go — into the agent session itself — so only checked work enters your code base. Why it matters.
What a developer or a business needs to know
Trust
Verdicts that can't be talked round
Real builds, the real tests your change reaches, an adversary that attacks the fix, and receipts bound to the exact inputs they checked.
How it works →Time and money
Stop re-checking by hand
Review a verdict with evidence, not every diff. Delta-only checks, a light hook, and a plain price — measured on your own code.
Time and money saved →Security
Your code stays on your machine
No Cotterpin cloud sees your source. Destructive agent actions wait for your approval, signed by a key only root can read.
Security →Fast, frugal, and it keeps itself in shape
Speed
Optimised for speed
Only the tests your change reaches run, a test already proven on the same inputs is reused, and every stage has a time budget it is held to. A commit's checks take seconds.
Cost
Fewer tokens, the same results
A model and effort matrix gives each task the model it needs, not the strongest one: a search runs on a fast model, a hard fix on the strongest. On a monthly plan the same limit holds more work.
Tuning
Always measuring, always tuning
Every request's planned model is recorded against what actually ran, and every route run gets a speed review, so the balance between cost and quality is tuned on measurements, never guesses.
Resilience
A harness that heals itself
When a run stops, Cotterpin matches the failure against a shared pool of healing skills and applies the fix within safe bounds. Every new failure becomes a skill, so the same stop never happens twice.
Memory
Memory that stays with your repository
What your agents learn about the codebase is kept as a knowledge bundle in the repository itself: plain files your team reviews and commits. Every session starts from it, so nobody pastes the same context twice.
Parallel
Safe parallel runs, sized to your machine
Cotterpin measures your machine's cores and memory and runs as much work side by side as it can hold, each task in its own isolated worktree, and backs off when memory runs short, so your laptop stays usable.
Built for the people who answer for the code
For developers
- Keep your agent and your habits. Antigravity, Claude, Codex or Grok, prompted the way you do now.
- Try it without risk. Watch is free with no time limit: it reports what every gate would catch and never blocks.
- Easy to leave. Every change is a commit;
cotterpin --uninstallremoves it.
For companies
- One policy for every team, with single sign-on on Team and Enterprise.
- Evidence for every decision, and a decision log agents can't rewrite.
- Answers for your security review, in one place.
Questions people ask
What is Cotterpin?
Cotterpin is guardrails for AI coding agents. It checks every “done” your agent claims, on your machine and before the commit, with real builds, tests and gates, so unverified AI-generated code never lands in your code base.
Does Cotterpin replace my coding agent?
No. It works inside the agent you already use (Antigravity, Claude, Codex and Grok) through the agent's own hooks, and you keep prompting the way you do now.
Does my code leave my machine?
No. Cotterpin runs on your machine and never receives your source, diffs or prompts. Only your own model provider sees your code, under your own account.
Is Cotterpin free?
Yes, to start. Watch is free with no time limit: it reports what every gate would catch and never blocks. Pro and Team enforce the gates; prices are on the pricing page.
Your agent will say “done” again today. Find out if it's true.
Free, with no time limit. No credit card required.
Start free
$ pipx install cotterpinfrom 0.3.0