/.well-known/ai-agent-manifest.json.
It is the agent-facing version of a sitemap. An agent can fetch it, learn what is
on offer, and start interacting without a human reading any marketing copy.
What It Describes
Fetch It
Designed For Both Sides
The manifest is written for two readers at once. Agents readcapabilities and
endpoints to act. Humans read human_oversight to know what stays under their
control. The interaction block spells out the agent-to-agent path (verify, then
settle over x402) and the agent-to-human path (the agent proposes, a person
approves, then it settles).
The Card Is Also An ERC-8004 Registration File
Our agents #0 on Arc Mainnet, #0 on Robinhood Chain, #9759 on Celo and #0 on Robinhood Chain testnet point their tokenURI at/.well-known/agent-card.json, so the same file is
their ERC-8004 registration file (registration-v1) as well as the A2A card. Besides the
A2A fields it carries:
A registration says that a token is this agent, and explorers check it both ways, so the
list holds the agents whose tokenURI is this file. X Layer #6271 is the one exception: it is
the same oracle, listed on OKX.AI, and OKX.AI hosts its file. Arc testnet #849980 (Meridian)
is ours too, but its tokenURI is its own inline metadata, so it is a different agent and is
not listed.
The
A2A service names version 0.3.0, the A2A AgentCard schema the file validates
against, although the card’s own protocolVersion says 1.0: an A2A 1.0 interface entry
needs protocolBinding and protocolVersion, and the card’s interfaces are its MCP and
x402 endpoints. The service is for discovery only: A-Identity serves no A2A message
endpoint. The skills the card lists are HTTP tools at its x402 interface, and four of them
are also sold on Celo through the paid Celo tools.
What We Deliberately Do Not Declare
The A2A agent card at/.well-known/agent-card.json tracks A2A 1.0 and lists only
capabilities that have an implementation behind them. Two absences are choices rather
than gaps, and they are written down here so nobody has to guess:
AP2, the Agent Payments Protocol. The card declares no AP2 extension. AP2 answers
“did the human authorise this, and within what limits”, which is a real question and one
adjacent to ours. But we do not accept an AP2 Payment Mandate as an input today, and
declaring an extension we do not implement is the same species of claim this project
refuses everywhere else. When a mandate can actually be handed to risk_check and come
back with a verdict bounded by it, the extension goes in the card on the same day.
A payment handler for someone else’s checkout. We publish paid endpoints and we sell
over x402, but the card does not advertise A-Identity as a drop-in payment handler for a
third-party commerce flow. That would imply an integration surface that does not exist.
The rule behind both: a capability appears in the card when a caller can exercise it, not
when we intend to build it. The same rule governs
/.well-known/http-message-signatures-directory, where the published signing key has a
test asserting the signer it advertises is actually reached from our outbound requests.
Keep the page source LLM-parsable. Plain structure, clear labels, and the manifest
link in the page head let an agent navigate without guessing.
