Open-source reliability and governance checks for AI agents.
Ten opinionated checks plus a deterministic production gate. The coverage score shows where controls are strong or weak. The gate returns PASS, REVIEW, or BLOCK from declared answers and evidence references. No tooling lock-in and no vendor dependency.
→ Full writeups: www.goclawproof.com/checks → Interactive assessment: www.goclawproof.com/assessment
| # | Check | Category |
|---|---|---|
| 01 | Tool Permissions & Least Privilege | Security |
| 02 | Logging & Audit Trails | Quality |
| 03 | Prompt Injection & Data Exfiltration | Security |
| 04 | Human-in-the-Loop & Escalation | Governance |
| 05 | Rollback & Kill Switches | Operations |
| 06 | Secrets Management | Security |
| 07 | Evaluation & Regression Testing | Quality |
| 08 | Data Boundaries & RAG Governance | Governance |
| 09 | Cost Controls & Rate Limiting | Operations |
| 10 | Multi-Agent Coordination | Quality |
Each check is a single YAML file with the failure mode, two verification questions, a production checklist, common pitfalls, and a link to the full writeup.
A packaged skill is included at skills/clawproof-audit/. Drop it into any Claude Code or Claude Agent SDK project and an agent can audit a codebase, runbook, or configuration against all 10 checks on demand.
# Clone into your project's skills directory
git clone https://github.com/lexbeam-software/clawproof-checks.git /tmp/clawproof
cp -r /tmp/clawproof/skills/clawproof-audit ./.claude/skills/
# In Claude Code:
> /skills
# Select: clawproof-audit
> audit this agent for production readinessThe skill produces a scored audit report (0-100) with prioritized remediation, linked back to the full check writeups.
Every check is also bundled into a single JSON file for programmatic consumption:
curl -O https://raw.githubusercontent.com/lexbeam-software/clawproof-checks/main/skills/clawproof-audit/checks.jsonSchema: see checks/SCHEMA.md.
The CLI uses Python's standard library and sends nothing over the network. Start with the example declaration, replace the evidence references with your own, and run it:
git clone https://github.com/lexbeam-software/clawproof-checks.git
cd clawproof-checks
./bin/clawproof gate examples/gate-pass.json --format markdown --target my-agentExit codes are stable for CI:
| Exit | Decision | Meaning |
|---|---|---|
| 0 | PASS |
Every answer is yes and has evidence |
| 2 | REVIEW |
A control or evidence reference is incomplete |
| 3 | BLOCK |
A critical control explicitly failed |
| 4 | INVALID |
The declaration or selected policy fails structural validation |
The score is diagnostic, not the release decision. A critical failure blocks even at
95/100, and an affirmative claim without an evidence reference cannot produce PASS.
The gate does not open or authenticate those references. See the versioned policy at
policy/clawproof-gate.v1.json and copy the example
workflow from examples/github-actions/clawproof-gate.yml.
The gate evaluates what you declare. It does not inspect or certify a production environment. Evidence references are included in the report for a human or CI process to verify.
Because agent governance that only lives inside one consultancy's deck is not governance, it is slideware. These checks are the inputs I wish I had when I started running agents in production. If they save someone a Monday morning incident, the repo has paid for itself.
New failure modes, additional checklist items, language-specific playbooks — all welcome. See CONTRIBUTING.md.
MIT. Use these checks in commercial products, internal tooling, client deliverables, or anywhere else. Attribution is appreciated but not required.
Maintained by Werner Plutat / Lexbeam Software.
For enterprise rollout support in DACH: www.agentklar.de.