Kontrua

Your agent instructions are wrong and nothing checks

Agents obey your AGENTS.md, even when it lies. Kontrua reads every claim it makes, proves each one against the real code, and hands you the verdict on the PR that broke it.

A verdict Kontrua produced on a real public repository, refreshed every six hours.

Kontrua
72 claims read in 1.2s
homebridge/homebridgeat edf549370/70 verifiednothing broken
  • CLAUDE.md:7 npm run build
  • CLAUDE.md:7 npm run clean && tsc
  • CLAUDE.md:8 npm run watch
  • CLAUDE.md:9 npm run dev
  • CLAUDE.md:10 npm run lint
  • CLAUDE.md:10 npm run lint:fix
  • CLAUDE.md:10 eslint.config.js
  • CLAUDE.md:11 npm run test
Deterministic verdict. No model decides what is true.See the whole verdict
Works with
GitHub
GitLab
Context rot

Agents obey it, right or wrong

Claude Code, Cursor and Copilot plan from the context files in your repo. Every merged PR widens the gap between what they say and what the code does.

  • Provably true. Every command and path resolved against the real code. Not a model's opinion.
  • Lower token bills. Without it, agents re-discover your repo on every task: exploratory reads, wrong plans, retried diffs. A dense, true AGENTS.md is the cheapest context an agent can load.
  • Your file, kept honest. Kontrua edits only the lines the code made false. What you wrote stays yours.
What the correction PR contains

Proposed on a branch you review

The artifacts it owns
AGENTS.md and the other context files Kontrua generates arrive as an ordinary diff, at whatever paths your config gives them.
The files you wrote
One line at a time. Every correction is line-local and shows the line before and the line after, so you review a sentence, not a rewrite.
What it declines to touch
A claim it cannot repair safely is left as is and reported with its file, its line and the reason. A line that only describes something you deleted is reported too, never deleted for you.

Every other line in the file stays byte-identical.

How a claim is resolved
A command
against the scripts and targets your manifests define
A path
against the file tree at HEAD

Every claim lands in one of three states

  • Verifiedthe anchor resolves
  • Brokenit resolves to nothing
  • Unverifiablereported, never failed
Verification

Every claim resolved, or reported

Commands are matched to your manifests, paths to your file tree. No model decides what is true: an anchor resolves or it does not.

  • Verified, broken, or unverifiable. Three states. Never two.
  • Never a false alarm. Undecidable is not broken.
  • A status you can require. Enforceable in branch protection.
  • Section-scoped edits. Untouched lines stay byte-identical.
How it works

Connect, read the verdict, order the fix

The audit runs on its own and costs nothing. The correction is the part you ask for.

  1. Connect a repo

    GitHub · GitLab

    Install the GitHub App or add your GitLab project. The first audit runs in about a minute and tells you what your context files get wrong today: every command and path they claim, resolved against the code at HEAD.

  2. Every push and every PR gets a verdict

    Free and unlimited

    Nothing to add to your pipeline. Kontrua posts a kontrua/context commit status and one comment naming each broken claim with its file, line and reason. A claim it cannot decide is reported, never failed, so the check stays worth requiring.

  3. Ask for the correction PR when you want the fix

    Metered

    The correction lane opens a pull request that repairs the drifted lines: the context artifacts Kontrua owns, plus single-line corrections to the files you wrote. You review it like any other PR. This is the part that is metered.

Catch it at the merge that breaks it

The audit needs no trigger: every push and every PR gets its verdict, free. These are the ways to launch a correction, the run that rewrites what broke.

Trigger on PR approval

Install the GitHub App or connect your GitLab project. Kontrua audits every pull request for free, and when the repository is set to trigger on approval it runs the correction for the claims that pull request broke.

  • No workflow file and no pipeline step to add
  • The verdict arrives as a kontrua/context status plus one comment
  • The correction is committed to the PR branch, so code and context merge together
  • Works with GitHub and GitLab

Trigger mode is per repository: on PR approval, on merge, or manual only.

Nothing to paste

Kontrua receives the pull request event from the GitHub App or your GitLab project, so both the audit and the correction start without a snippet in your repository.

Comparison

The gap compounds with every merge

Hand-maintained context vs. Kontrua.

What is checked
Nothing. Agents obey whatever the file says
Every command and path, against HEAD
When you find out
Mid-task, from an agent following a dead path
On the pull request that broke it, as a failed check
Proof
A reviewer’s memory
A verdict pinned to a commit, no model in the loop
False alarms
Guesswork either way
Undecidable claims are reported, never failed
The fix
Someone rewrites the file by hand, eventually
A correction PR touching only the false lines
Cost
Agents re-discover the repo on every task
Audits free and unlimited; you pay only for fixes
Pricing

The verdict is free. You pay for the fix.

Audits and PR verdicts are unlimited on every plan: connect a repo, push, open a pull request, get a verdict on every claim. The allowance counts corrections only, and one run is one correction pass. The API trigger and the Notion sync work on every plan; scheduling needs a paid one. Open-source repos always get Max features, free.

Free
$0forever

For solo devs and small projects

  • Unlimited audits and PR verdicts
  • 20 corrections / month
  • 2 repository setups / month
  • 2 repositories
  • 1 developer, invites on Pro
Get started
Most popular
Pro
$30/developer/month

or $24 per developer per month, billed annually

For growing engineering teams

  • Unlimited audits and PR verdicts
  • 30 corrections / developer / month
  • 3 repository setups / developer / month
  • Unlimited repositories
  • Unlimited team members
  • Priority email support
Start 14-day free trial
Max
$60/developer/month

or $48 per developer per month, billed annually

For teams shipping at scale

  • Unlimited audits and PR verdicts
  • 60 corrections / developer / month
  • 8 repository setups / developer / month
  • Unlimited repositories
  • Unlimited team members
  • Priority email support
Get Max
Enterprise
Custom

Custom solutions for your team

  • Everything in Max
  • Custom correction allowances
  • Dedicated support
  • Custom integrations
Talk to us

Frequently Asked Questions

The context files agents read before they plan: AGENTS.md, your conventions, your docs. Every command and path they claim is resolved against the real code, on every push and every pull request.

No. Commands are matched against your manifests and paths against the file tree. A claim is verified, broken, or reported as unverifiable. No model sits in the verdict path, so the verdict cannot hallucinate.

It corrects them. When the audit finds a broken claim, the correction lane can open a pull request that rewrites only the false lines, and you review it like any other PR. Kontrua does not invent documents from scratch, and never ships a sentence it cannot anchor to the code.

No. Your file is the asset. A correction edits only the lines the code made false; everything else stays byte-identical.

Any tool that reads context files from your repo: Claude Code, Cursor, Copilot, and anything else following the AGENTS.md convention. Standard files, no SDK, no lock-in.

Audits and PR verdicts are free and unlimited on every plan, including Free. You pay per developer for corrections, the runs that rewrite what broke.

Kontrua never stores your source. Diffs and file contents are analyzed in memory during a run and discarded. Writes only ever land on a Kontrua branch, never on your default branch.

Find out what your docs get wrong

Connect a repo. The first audit takes about a minute, costs nothing, and every claim gets a verdict.