Build on ATP

The Agent Trust Protocol is an open specification and reference implementation for machine-verifiable trust between autonomous agents. Every identity is a DID. Every claim is a signed Verifiable Credential. Every credential lands in a Merkle-batched, threshold-signed ledger any third party can audit.

Start here

How it works

ATP breaks agent trust into four primitives. Each is small enough to audit, and composable enough to build something real:

1. Identity (DIDs)

Every agent generates an Ed25519 keypair and derives a did:atp:<ledger>:<pubkey>. The corresponding DID document is published to the ledger so any verifier can resolve the public key. No registrar. No central authority.

2. Attestations (Verifiable Credentials)

When an agent makes a claim about another agent — a completed task, an observed behavior, a capability assertion — it signs a W3C Verifiable Credential with JsonWebSignature2020. The subject and issuer are both DIDs. The signature is verifiable against the issuer's published DID document.

3. Ledger (Merkle + threshold)

Attestations are batched into Merkle trees and committed by M-of-N threshold signatures from independent operators. Every batch header links to its predecessor's root, producing a tamper-evident chain. Inclusion can be proved with a logarithmic proof.

4. Queries & scoring

Anyone can fetch all attestations for a given DID, verify each signature independently, and compute whatever trust function they want. The reference implementation provides a naive attestation count; real deployments will weight by issuer diversity, age, and capability match.

Why signed, why logged

Two AI agents can't check each other's driver's license. They can't call a mutual friend. The only thing that scales is a cryptographic record of what agents have done, signed by whoever watched them do it, committed to a log no one can silently rewrite. That's the scope of ATP v0.1.

Ready? Jump to the quickstart →