Ninety seconds, no product demo.
You want the primitive, not the pitch. VAID is an open standard for verifiable agent identity: one canonical signing path, three reference implementations, and a set of conformance vectors you can run against any of them. Everything on this page is Apache-2.0 and installable now, without an account or an API key.
Install it
cargo add vaid-pop # 0.2.1 pip install vaid-pop # 0.2.0 npm install vaid-pop # 0.3.0
Those numbers are read from the registries by CI, not typed here. Ten packages across three registries are checked the same way.
One canonical signing path
RFC 8785 JCS → SHA-256 → Ed25519 over the digest. That is the whole of it. No negotiation, no suites, no versioned algorithm registry to implement.
The bytes agree, and the claim is bounded
Rust, Python and TypeScript reproduce the published conformance vectors byte-for-byte, with no shared runtime between them. The claim is asserted at the vector, not at the version, which is narrower than "the three implementations agree", and is the one you can check.
Apache-2.0
The spec, the vectors and the reference verifiers are open. On crates.io the vectors and tests ship inside the crate rather than as an installable binary, an asymmetry worth knowing before you wire your CI.
Status, from the manifest
Ten shipped, one planned.
These badges are rendered from a capability manifest that is a byte-identical copy of the one
in the vaid repository, drift-checked in CI. Prose on this site does not assert
capability status; that file does, and it is the same file the SDKs are released against.
The full list of eleven, with each blocker named, is on the capability page →
Further reading: the repositories, verify a VAID in your browser, state of the standard.
Before you build on it
Three things worth knowing now rather than later.
The reference mint's revocation is in memory only
Restart it and a VAID you revoked beforehand verifies clean. A VAID is a bearer credential, so that is a real exposure and not a cosmetic one. Plan for a revocation store you actually run, or scope lifetimes short enough that you do not need one.
Verifying across deployments is Planned
Verifying a VAID issued by another deployment additionally needs that deployment's key, and binding a verifier to an issuer identity is an irreversible naming decision that has not been made. Within one deployment, and against a published anchor, verification works today, see the trust anchor.
Two framework templates exist, and LangChain is an adapter
Governed agents are provisioned from ADK and OpenAI templates; those are the only two.
vaid-langchain is published and installable, and it is a
request-signing adapter, a different and smaller claim than governing a
LangChain agent. Anything you read elsewhere about CrewAI is wrong; it exists in no
template, no adapter and no test.
The ending
Go and check the bytes.
This is where this route stops. No form, no call, nothing gated to run it yourself.
Hitting something the docs don't cover, or integrating this into something bigger than a twenty-line script? Talk to us about integration →
Wrong door?
I am responsible for securing agents my organisation runs.
Security leaders about 6 minutes 02My business is adopting AI and I need to know what could go wrong.
Business owners about 4 minutes 03I want this operated for us, on our terms.
Enterprise about 3 minutes 04None of those, or not sure yet.
Everyone else about 2 minutes