Install the context
Put CheckYourself in or next to your project, or point your AI assistant at the folder as its operating context.
completion evidence for AI-built apps
Your agent says it’s done. CheckYourself makes it prove it.
A read-only audit that runs the checks itself, shows what the project proves, and keeps what it could not check in plain sight.
audit ready
example output
checkyourself audit ./your-app
01read project signalsobserved
02execute verifier checksexecuted
03map coverage + scorescoped
04name what stays unprovenvisible
Source and docs: github.com/KyaniteLabs/checkyourself — Apache-2.0, local-first, no account.
the gap between “done” and ready
The happy path is not the whole app.
“It works on my screen” is a starting point, not a receipt.CheckYourself, for the moment after your agent stops typing
AI coding tools are excellent at momentum. They are less helpful when the missing piece is an auth edge case, a release check, or the thing nobody thought to test.
CheckYourself turns completion into something you can inspect. It reads the project, runs verifier-owned challenges, maps the production surface, and gives you a bounded 0–100 score with the gaps attached.
from claim to receipt
Four moves. One calmer launch decision.
the audit loop
Make the unknowns legible.
Every step keeps the evidence close to the action, so “not checked” cannot quietly turn into “safe.”
Run a read-only audit
The audit detects the project shape and sweeps the relevant launch surfaces before it suggests a fix.
Let the verifier execute
Committed challenges run the checks themselves—tests, validators, and gates—with receipts tied to the claim.
See the score and the gaps
Review the score, coverage, complete findings register, safe approval batch, and a learning plan built from what your app missed.
what’s inside
A second pass with a memory.
CheckYourself is free, open-source, model-agnostic, and designed to run where your code already lives.
Broad enough to catch the quiet failures.
The sweep covers 20 canonical surfaces, grouped into 10 scored categories, backed by 150+ tests. It records what was observed, what was inferred, and what is still untested.
CLI + MCP · no SaaS account required · local-first by design
CursorClaudeCodexCopilotWindsurfReplitLovableBolt
surfaces in the sweep20 mapped
- product purpose + harm
- frontend UX + accessibility
- API + backend behavior
- auth + permissions
- data + recovery
- secrets + runtime config
- tests + quality gates
- CI/CD + dependencies
- deploy + rollback
- observability + response
Evidence, not a guarantee.
The score reflects what this audit could prove from the project at that moment. It does not certify production safety, and it cannot see what the project never exposes.
That boundary is the point: anything not checked is reported as unproven, not hidden behind a confident number.
start with reality
Give “done” somewhere to land.
Add the folder to your project, point your assistant at it, and run the local CLI. Keep the first pass read-only; approve fixes only after you can see the evidence.
run the audit
python3 tools/checkyourself.py /path/to/your/project
The same workflow is available through MCP for native agent clients.
See the audit loop