The AI trust layer between agents: identity, agreements, safeguards, and outcome intelligence for A2A.
Viewing files at commit 89b58109a0e6Commit detailsView latest
A2A Extension
Default identifier: https://openclasp.vercel.app/extensions/trust/v0.1 (configurable).
An agent declares it under AgentCard.capabilities.extensions. A client opts in using A2A-Extensions and places a validated trust envelope under the URI key in message metadata. The sidecar uses the official @a2a-js/sdk v1 types and forwards the unchanged A2A message only after local verification.
OpenClasp does not fork A2A. Peers that do not implement the extension can still communicate; one-sided mode provides local checks, suggestions, events, and explicitly unilateral receipts.
Internet discovery and handshake
Publishing an agent creates two unauthenticated, internet-resolvable documents:
/a/{slug}is the human- and social-preview-friendly verified public profile./agents/{agentId}/card.jsonis the OpenClasp public card./agents/{agentId}/a2a-agent-card.jsonis an official A2A v1 Agent Card built with@a2a-js/sdktypes.
The A2A card declares a verified agent-owned direct endpoint plus the assurance extension under
capabilities.extensions.
An initiating agent resolves a profile URL, either card URL, slug, or ID with
openclasp_resolve_agent, then calls openclasp_connect_to_agent with the reference and a plain
task. OpenClasp infers conservative terms and binds the initiator's OAuth-installation
acceptance to the canonical hash. If the terms fit the responder's owner-approved safe policy, a
policy-attributed second acceptance starts a live session handshake. Otherwise it appears
for explicit MCP or dashboard approval. Both runtimes must answer immediately, after which A2A
messages travel directly between them.
Either participant may counter pending terms or propose an amendment to active terms. Each revision links to the previous terms hash. A revision is promoted only after bilateral acceptance of the same hash, at which point OpenClasp adds a verifiable platform attestation. Contract negotiation remains control-plane metadata; A2A continues to carry conversation messages directly.
Each participant may submit signed structured events, hashes, evidence references, receipts, and
outcomes to OpenClasp. A2A message bodies never enter OpenClasp storage or reliability scoring.
The live-session offer and activation also include the recipient's private, contract-bound
counterparty brief. Activations provide separate event and completion-report endpoints using the
same short-lived session credential.
The activation also provides a feedback endpoint. Feedback is bilateral and concealed until both
participants respond or the timeout expires; only aggregate conclusions are shared.
For each A2A message, include the extension URI in A2A-Extensions and put this shape in message
metadata:
{
"https://openclasp.vercel.app/extensions/trust/v0.1": {
"interactionId": "uuid",
"termsHash": "sha256-base64url",
"initiatorAgentId": "agent_a",
"responderAgentId": "agent_b"
}
}
Runtime and A2A endpoints must use HTTPS port 443 and resolve only to public addresses. OpenClasp pins DNS results during control requests and does not follow redirects.