One sentence
Federated Agent Use (FAU) lets a remote AI assistant answer narrow questions about a person’s email without ever receiving their inbox credentials — mail stays on their laptop; permission is granted per request on their phone.The problem
Today, “connect Gmail to my AI” usually means: give the product an OAuth token (or an app password), sync mail to their servers, and hope policy and logging are good enough. That fails three ways:- Credential blast radius. A stolen inbox token is the whole mailbox.
- Data gravity. Once mail is copied off-device, revocation is theater.
- Opaque access. Users cannot see which messages an agent touched for which question.
Mental model
Think of three parties and a blind post office:Goals
- Assistants can get useful, narrow answers from mail the user already has.
- The user keeps custody of mail and credentials on their own machine.
- Every access is scoped, revocable, and auditable in language a person can read.
- The public rendezvous (relay) learns as little as possible and never decrypts node traffic.
- Anyone can attempt contact (open addressing); permission stays local (closed authorization).
Non-goals
FAU deliberately does not:- Replace Gmail, IMAP, or the user’s mail client.
- Host mail, embeddings, or full-text search in the cloud for agents.
- Provide a global registry of “approved” AI vendors (except optional claim enrollment on the user’s node).
- Guarantee the Mac is online 24/7 without the offline queue (and even then, envelopes are size-limited).
- Stop a compromised Mac from reading its own mail (the node’s host is inside the trust boundary).
- Make the relay unable to see metadata (timing, byte volume, HMAC of email, hostname). Blind ≠ invisible.
Terminology
Use these words exactly. Do not invent synonyms mid-document.How the system is decomposed
Two planes
Control plane — TLS terminated by the relay atrelay.<domain>:
- Agents: resolve, offline queue
- Nodes: register, ACME, tunnel WebSocket, self-service
<random>.relay.<domain>:
- Agents:
GET /.well-known/fau-node,POST /mcp - Phones:
/papi/*
*.relay.).
Permission model (the part people get wrong)
- Open addressing. Anyone can
POST /v1/resolvewith an email. Unknown addresses get a decoy hostname with the same JSON shape (anti-enumeration). - Closed authorization. Tool calls are signed as an agent identity. The node’s gateway checks grants and may require a phone approval before retrieving mail for the model.
- Retrieve before the model. The gateway filters which messages exist in the candidate set first. A prompt injection in an email cannot widen retrieval, because retrieval already happened under the grant.
- Phone is the human factor. Raw read and send are always prompted. Structured tools may auto-pass once a grant exists.
What each party can observe
Invariants (always true)
- Assistants never receive an inbox OAuth token or app password from FAU.
- The relay never decrypts data-plane TLS for a node hostname.
- A phone decision is accepted only with a signature from a paired device key (or an explicit local dashboard decide, recorded as such).
- Push notification payloads must not carry mail content; they are wake signals.
/v1/resolveresponses for unknown emails are shape-compatible with known ones.- Device signatures cover the full request target including query string.
- Agent signatures cover method, target URI, authority, and body digest; nonces are single-use inside the expiry window.
Design posture (why it looks like this)
How to read the rest of these docs
Protocol version throughout:
fau/1.