nocap 🧢→✅

Is your AI agent lying about being done? Catch it in 1 line.

A lie detector for AI coding agents. It runs your checks itself and lets the real exit code overrule the agent's self-report — PROVEN ✅ or OVERCLAIM ❌.

receipts over self-report

nocap (slang, "no cap"): no lie, for real. An agent is "capping" when it claims a task is done that it didn't actually finish. nocap catches that.

The problem

Your AI agent says "bug fixed, all tests pass ✅" — but the tests never ran, the build is broken, or the file wasn't actually changed. You trust it, merge, and break prod. Every Cursor / Claude Code / Cline / Aider user has been here.

nocap is the receipt. It doesn't trust what the agent says — it runs the check itself and tells you whether the claim is PROVEN or an OVERCLAIM.

Quick start

Zero install, zero config — try it right now:

npx @zoeyx/nocap demo

Adjudicate a single command (exit 0 → PROVEN, anything else → OVERCLAIM):

npx @zoeyx/nocap verify "npm test" --claim "bug fixed, tests pass"
# exits 0 if the real exit code is 0, exits 1 on OVERCLAIM

Declare a claim, then check it

Create a .nocap.yml in your repo (or run npx @zoeyx/nocap init):

claim: "login form validates email and shows errors"
verify:
  - id: unit-tests
    run: "npm test"
  - id: build
    run: "npm run build"

Then:

npx @zoeyx/nocap check

nocap runs every verify step itself and adjudicates: all green → PROVEN, any red → OVERCLAIM. The agent's narrative never counts — only the receipt.

Use it everywhere

CI — GitHub Action

# .github/workflows/nocap.yml
on: [pull_request]
jobs:
  nocap:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: 20 }
      - run: npm ci
      - uses: zhouyuanxinand/nocap@v1
        with:
          command: check

The check fails (nonzero exit) on OVERCLAIM, so a lying agent can't get its PR merged.

Git hook — block the push

npx @zoeyx/nocap hook install

Installs a pre-push hook that runs nocap check and blocks the push while any verify step is red. Remove it with nocap hook uninstall.

Machine-readable output

nocap --json check

{
  "verdict": "PROVEN",
  "claim": "login form validates email and shows errors",
  "steps": [
    { "id": "unit-tests", "command": "npm test", "exitCode": 0, "provenance": "harness" }
  ]
}

How it works (the one idea)

nocap distills a single principle — receipt > observation — into a tiny CLI:

• An observation is what the agent says ("done", "tests pass").
• A receipt is a command nocap ran itself, capturing the real exit code and hashing the output.
• Only a receipt can prove a claim. The agent cannot self-approve.

That's it. No prompts, no models, no theory — just a deterministic check whose exit code is the verdict. The agent is free to claim whatever it wants; nocap runs the proof.

For the curious: receipts are persisted under .nocap/runs/<id>/receipts/ with SHA-256 hashes of stdout/stderr, so a captured result can be re-verified later (tamper detection).

Commands

Command	What it does
nocap demo	self-contained "catch an agent lying" scenario — no setup
nocap verify "<cmd>"	run one command, adjudicate it
nocap check	read .nocap.yml, run every verify step, adjudicate
nocap init	scaffold a .nocap.yml (auto-detects your test command)
nocap hook install / uninstall	manage the pre-push git hook
--json	machine-readable output (any command)

Exit code: 0 on PROVEN, 1 on OVERCLAIM, 2 on usage error.

Works with every AI coding agent

nocap is model- and host-agnostic. It doesn't care which agent wrote the code — Cursor, Claude Code, Cline, Aider, Windsurf, Continue, GitHub Copilot Workspace, or your own. If it can lie about being done, nocap can catch it.

Install

npm install -g @zoeyx/nocap        # global
# or just use it ad hoc:
npx @zoeyx/nocap check

Requires Node ≥ 18.

How is this different from…

	what it checks	how it decides
nocap	an agent's completion claim	re-runs the receipt (real exit code)
Guardrails AI / NeMo	LLM output content	validates generated text
OpenAI Evals / promptfoo	model quality offline	scores a dataset
Langfuse / Opik	agent traces	observes, doesn't adjudicate

nocap is the only one that exists to answer one question: did the agent actually do what it said it did?

Acknowledgements

The "receipt > observation" idea is inspired by agent-governance (MIT) — a deeper governance protocol. nocap is the distilled, model-agnostic, 30-seconds-to-set-up version of its core check.

License

MIT