Skip to content

About

Signed flight recorder and policy firewall for AI agent runs

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

AgentProof

AgentProof - signed flight recorder and policy firewall for AI agent runs.

Live MIT Next.js Architecture

AgentProof hero

Why AgentProof

AI agents are starting to read documents, call tools, post updates, push code, and operate across real workflows. Teams cannot grant more authority to a system they cannot replay or verify. AgentProof wraps an agent run with a trace collector, deterministic policy guard, redaction layer, signed hash-chain proof bundle, and public report so builders can show exactly what happened before, during, and after a risky action.

What it does

Open mission control, choose a scenario, and run it through the OpenClaw adapter. AgentProof connects to an OpenClaw Gateway over WebSocket, submits the task to an OpenClaw agent, captures Gateway events, evaluates the proposed tool action, and stores the causal trace. If the agent attempts an action outside scope, the guard blocks it before execution and generates a public report.

After the run, open the replay page to inspect event order and redacted payloads. The report page turns the trace into a shareable public summary with the policy findings, recommended fix, and exportable signed proof bundle. Paste that bundle into /app/verify to recompute event hashes, the root hash, and the AgentProof server signature.

AgentProof primary flow

Architecture

AgentProof is a Next.js application with a server-owned policy boundary. Route handlers generate trace events, evaluate deterministic policies, seal runs into proof bundles, persist runs in a local proof store, and expose report exports. The client renders only approved trace data and public redactions. See docs/ARCHITECTURE.md for the full flow.

AgentProof architecture diagram

Chinese mirrors are available in docs/zh/README.md and docs/zh/ARCHITECTURE.md. The hackathon positioning and demo script are in docs/AWARD_STRATEGY.md.

Quick start

npm install
cp .env.example .env.local
npm run dev

Open http://localhost:3000, start a run from mission control, then open the replay and proof report.

For the OpenClaw path, configure:

OPENCLAW_GATEWAY_URL=ws://127.0.0.1:18789
OPENCLAW_GATEWAY_TOKEN=<gateway token>
OPENCLAW_AGENT_ID=main
OPENCLAW_MODEL=<optional provider/model override>
OPENCLAW_SEND_MODEL_OVERRIDE=false

For the recorded OpenClaw flow, the Gateway profile default model is stepfun/step-3.6; AgentProof does not send a per-run model override unless OPENCLAW_SEND_MODEL_OVERRIDE=true, because Gateway callers may not be allowed to override provider/model.

AgentProof Skill

The reusable skill lives in skills/agentproof-skill/SKILL.md. Use it when an agent is about to execute a tool call with external side effects, post secrets, change code, deploy, or cross an approval boundary.

node skills/agentproof-skill/scripts/verify-bundle.mjs path/to/proof-bundle.json

Verification

npm run build
npm run test:e2e

License

MIT © AgentProof

About

Signed flight recorder and policy firewall for AI agent runs

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages