Cerberus blocks the lethal trifecta at the tool boundary — see the 525-run evidence set.

Guardian · Developers

Developers

Register an agent, give it a Guardian credential instead of a master key, and let Guardian decide and execute each action.

Available in the pilot

How an agent works with Guardian today

Your agent never holds a mailbox token. It holds a Guardian credential that can request actions and cannot approve them.

  1. 01Register the agent

    In the Guardian app, create an agent and copy its one-time credential. The agent never sees your Gmail token.

  2. 02Grant authority

    Choose Allow, Ask or Off for each connected mailbox. Each change is a new policy version.

  3. 03Request a capability

    The agent sends the exact action it wants to Guardian’s agent API.

  4. 04Claim, if you were asked

    When the verdict is ASK, the agent waits until you decide in the app, then claims the approved capability.

  5. 05Execute and verify

    Guardian performs the send with the sealed credential, verifies it, and returns a receipt.

Agent API

Three calls: request, claim, execute

Guardian agent API routes
RoutePurpose
POST /v1/agent/requestsRequest a capability for one canonical action. Returns ALLOW, ASK or DENY.
POST /v1/agent/claimsClaim a capability a person approved.
POST /v1/agent/executeExecute the capability once. A repeat is suppressed.

MCP · in development The same three steps are available to MCP clients as Guardian tools: guardian_request_capability, guardian_claim_approved_capability and guardian_execute_capability.

Planned SDK A thin client library around the same contract:

guardian.authorize(action)
guardian.execute(capability)
guardian.verify(receipt)

Developer access is by invitation during the pilot. API reference and credentials are provided when you are onboarded.

Get started

Let AI act for you without letting it become you.

Odingard Security · Trust infrastructure for autonomous systems