The problem
Sharing one mailbox across sessions mixes codes between runs, and polling it for new mail is slow and fragile. Using a person's own address puts their account at risk and their inbox in the agent's context.
A browser agent also reads whatever lands in that inbox. A verification email is the one message it will open without hesitation, which makes it a good carrier for instructions aimed at the model.
How it runs on nofeb
- Create an inbox per session with ttl_seconds, so it cleans itself up. Inboxes are unlimited on paid plans, so there is no reason to reuse one.
- Type the address into the form, then call wait_for_email with has_otp=true. The call returns as soon as the code arrives, up to a minute later.
- Read the code from the message's extracted codes, or the magic link from its classified links. No HTML parsing.
- Every message arrives with a prompt-injection score and its reasons; above your threshold it is quarantined before the agent sees it. What the agent does read comes through safe_body, with hidden text and look-alike characters removed.
- For an agent that watches several sessions at once, a WebSocket stream delivers message.received for the inboxes its key can reach, resumable from a cursor after a reconnect.
- Give the session a key bound to its one inbox, so a compromised session can read nothing else.
Where nofeb stops
- nofeb does not solve CAPTCHAs, phone verification, or sites that refuse sign-ups from agents. Respect each site's terms.
Status of everything this relies on
| Capability | What it does | Status |
|---|---|---|
| Unlimited inboxes | Create as many inboxes as the work needs, with a generated or chosen address and an idempotent client reference. | Shipped |
| Disposable inboxes with a time to live | Give an inbox a lifetime in seconds; it and its mail are removed when it expires, and an event says so. | Shipped |
| Wait for an email | Long-poll an inbox for up to a minute until a message matching sender, subject or has-a-code arrives. | Shipped |
| One-time code and link extraction | Verification codes and classified links are pulled out of every message and returned as fields. | Shipped |
| Scoped API keys | Keys bound to the organization, one pod or one inbox, holding only the scopes they are given. A key cannot mint a key wider than itself. | Shipped |
| Agent self-signup with a human sponsor | An agent creates its organization, first inbox and API key in one unauthenticated call. Until the human sponsor it names verifies by email, it may only reply to people who wrote to it first. | Shipped |
| Injection scoring and quarantine | Every inbound message gets a prompt-injection score from 0 to 1 with its reasons: hidden text, invisible and look-alike characters, instructions aimed at an AI, role and chat-template tokens, tracking beacons, data-carrying links and encoded payloads. Above the threshold you set per organization, pod or inbox, the message is quarantined (or only flagged) for a person to release or delete, and its content stays out of the delivery event. | Shipped |
| safe_body for untrusted mail | The message's new text with zero-width and direction-altering characters, hidden HTML and forgeable quoted history removed, wrapped in a provenance header that tells the model whose words these are. | Shipped |
| Realtime stream | A WebSocket stream of the same events as webhooks and the feed, in order, resumable from a cursor with no gap and no duplicates, and filtered by inbox, pod or event type within the credential's own reach. | Shipped |
| MCP sign-in with OAuth | MCP clients connect by signing in instead of pasting a key: OAuth 2.1 with discovery, dynamic client registration, client ID metadata documents, PKCE, tokens bound to the MCP server, refresh-token rotation and immediate revocation. A person approves the scopes, and optionally one inbox, in the console, and can disconnect the app at any time. API keys keep working beside it. | Shipped |