clawsig

Verify what your AI agents actually did.

Cryptographic proof bundles for every agent action — tool calls, file changes, network connections, subprocess spawns.

Quick Start

# Wrap any agent command
npx clawsig wrap -- python3 my_agent.py
npx clawsig wrap -- pi "Fix the bug in auth.ts"
npx clawsig wrap -- node scripts/deploy.mjs

# Initialize a project with a verification policy
npx clawsig init

# Verify a proof bundle offline
npx clawsig verify proof-bundle --input .clawsig/proof_bundle.json

What It Does

clawsig wrap transparently observes your agent through 6 layers:

1. Ephemeral DID — unique Ed25519 identity per run
2. Local Proxy — intercepts LLM API calls, generates receipts
3. Causal Sieve — parses tool calls from HTTP streams
4. Sentinel Shell — captures shell commands via BASH_ENV + trap DEBUG
5. FS Sentinel — watches file changes with content hashes
6. Interpose Sentinel — hooks connect(), open(), execve() via LD_PRELOAD

On exit, everything is compiled into a signed proof bundle at .clawsig/proof_bundle.json.

Documentation

• Protocol Spec
• Troubleshooting
• GitHub

License

MIT