The AI trust layer between agents.
Agents can talk. Now they can build trust.
OpenClasp verifies agent identity, formalizes agreements, surfaces task-specific risk, recommends safeguards, and learns from signed outcomes. It runs on top of the Agent2Agent (A2A) Protocol: A2A handles communication; OpenClasp adds assurance before, during, and after agents work together.
Try OpenClasp · Run the demo · Read the protocol
- Verifies publishers, cloud runtimes, and agent identities.
- Records explicit agreements before agents act.
- Predicts success for the relevant task, agent, and version.
- Asks one high-value question at a time to expose material risk.
- Recommends safeguards and human approval when needed.
- Learns which questions, answers, and safeguards correlate with better outcomes.
- Produces signed outcomes, private bilateral feedback, and shareable agent cards.
Conversation bodies stay directly between agent runtimes. OpenClasp does not receive, relay, or store them. Predictions are advisory, not guarantees.
The deterministic demo requires no API key:
git clone https://github.com/aryaamin/OpenClasp.git
cd OpenClasp
corepack pnpm install --frozen-lockfile
corepack pnpm demoIt exercises identity, scoped delegation, signed agreements, private warnings, policy enforcement, conflict resolution, outcome receipts, contextual learning, and tamper/replay rejection.
Requires Node.js 24.x and Corepack.
corepack pnpm install
corepack pnpm schemas
corepack pnpm lint
corepack pnpm typecheck
corepack pnpm test
corepack pnpm dev- Dashboard:
http://localhost:5173 - API:
http://localhost:3100 - OpenAPI:
http://localhost:3100/openapi.json
Claude powers the optional AI layer when ANTHROPIC_API_KEY is configured. A conservative
deterministic fallback keeps assurance available during cold start or provider failure.
See Architecture, Runtime Connector, Privacy Model, and Security Policy.
OpenClasp is Apache-2.0 licensed and free during the launch beta. Contributions are welcome.