| title | Security | ||
|---|---|---|---|
| description | The security model for IT admins - OAuth 2.1 with PKCE and dynamic client registration, per-tool scopes with audience binding, audited reads, proposal-only writes, server-enforced revocation, structural blindness, EU residency, and self-serve account deletion. | ||
| sidebar |
|
For IT admins evaluating a connection:
- Authentication is OAuth 2.1 with PKCE and dynamic client registration. Approval happens in the user's browser; no API key or token is pasted or written to a file.
- Tokens carry per-tool scopes (
profile:read,files:read,docs:read,updates:write,shared:read) and support audience binding (RFC 8707 resource indicators). - Every access an AI makes leaves a record in the user's audit trail.
- The single write tool is proposal-only: it files a pending suggestion the user reviews in the app. Nothing an AI does edits the user's profile or files directly.
- Access is revocable from the Connect page: one app at a time for an app you connected by signing in, or everything at once. Revocation is enforced on the server: the token stops working immediately, everywhere. A token you pasted into a tool by hand ends when you sign out of the browser you made it in, or when you disconnect everything.
- The public-facing surface is structurally blind: it declares no data bindings, so it holds no readable user content, which lives on an isolated per-user data plane behind it.
- For EU users, the profile store (profile, facts, learnings, audit, compiled context), the handle registry and their backups are pinned to EU jurisdiction. Uploaded file bytes and the search index are not region-pinned today.
- A user can delete their whole account themselves, from Settings, with no support ticket and no waiting period. Profile, facts, files, projects, share grants, connections and audit trail are hard-deleted from primary systems, deleted data leaves every backup within 14 days (we keep only the three most recent backup generations), and the account's @handle is released. Deletion starts immediately; the run's duration scales with how much the account holds, and the response is a metadata-only receipt of what was removed, which cannot be re-fetched because the session dies with the account. An active paid plan must be cancelled first, and a user who is the sole admin of a team must disband it first - the server refuses in both cases rather than half-deleting. Deletion is a browser-only action: a token minted for a connected AI client is refused, so nothing an assistant does can erase an account. The named list of what survives is in the privacy policy.
- Every request to the MCP endpoint needs a valid token, starting with the first handshake. Without authentication the server answers with a sign-in challenge (HTTP 401 pointing at its OAuth metadata), so a client asks the user to sign in before anything else. Nothing is served anonymously over MCP: not the tool metadata (names, descriptions, schemas), and not the public context behind an @handle. A profile whose owner has set it to public is readable without an account on the web instead: at its public page, or as plain markdown at
https://usemycontext.ai/@<handle>.md.
The tool-by-tool scope mapping is in The thirteen tools. More detail is in the privacy policy.