VaultysClaw
VaultysClaw is a trust plane for AI agents. It gives every agent, device, sensor, and human in your organisation a cryptographic identity, records every permission any of them holds as a signed, revocable certificate in one append-only ledger, and provides a protocol that lets anyone — the control plane, another agent, a third party — check whether a given permission is still valid right now.
It is deliberately not an agent framework, a workflow engine, or a chat product. Those exist and are good. What is missing at most organisations is the answer to four questions, and that is all VaultysClaw is built to answer:
| Question | VaultysClaw's answer |
|---|---|
| Who is this agent? | A VaultysId DID, proven per connection by an SRP-style challenge/response handshake. Not an API key. |
| What is it allowed to do? | A CapabilityCertificate — signed by the control plane and by the agent itself, independently verifiable offline by anyone. |
| Is that still true? | The cert_status protocol: a signed, timestamped status response any party can request, cache, or forward. Revocation is a ledger write, not a hopeful push. |
| What actually happened? | One append-only audit log, every entry attributed to a DID, keyed to the exact certificate that authorised the action. |
The Zero Trust framing
VaultysClaw is designed against Anthropic's "Zero Trust for AI Agents" guidance, and we publish our self-assessment against it rather than claiming compliance in the abstract.
That assessment is a living document, not marketing copy: it names what is built, what is partial, and what is not there at all — including the domains where we currently score zero.
- Zero Trust overview — the framework, and how VaultysClaw maps onto it
- The compliance matrix — all twelve domains, tier by tier, with current status
- Gaps and roadmap — what is missing and in what order it is being closed
What is in the box
Everything on that diagram is an Actor — one entity, one registration flow, one ledger, one audit trail. Humans are Actors too. A sensor is not a special table; a proxy is not a special protocol. The only thing that differs per kind is the configuration it carries and the admin panel that edits it.
Where to start
Evaluating VaultysClaw? Read the Zero Trust matrix first — it is the most honest single page on this site — then Concepts → Certificates.
Deploying it? Quickstart, then Bootstrapping the first admin and Onboarding actors.
Building an agent against it? Agent kinds and the WebSocket protocol.
A note on maturity
VaultysClaw is in public alpha, and the control plane described by these docs
is a ground-up rebuild (packages/controlplane) that lives alongside the
older proof-of-concept (packages/control-plane). The rebuild deliberately
removed a large amount of product surface — workflow orchestration, human chat
channels, Teams bridges, the in-app notification stack, and the ts-rest REST API —
to do a much smaller thing properly.
If you are looking for docs on those features, see What was removed, and why. They are not coming back; VaultysClaw's job is agent identity and trust, and orchestration is better served by tools built for it.