Human-approved sending
The AI requests approval for a provider-side draft. The owner sees the exact draft, creates a short-lived approval envelope, and authorization is checked again immediately before dispatch.
InboxPort
Permissioned Gmail and Outlook access for AI clients, with revocable access and human-approved sending.
overview / problem
I built InboxPort after running into a connector limitation myself: connecting more inboxes solved the account problem, but created a harder permissions problem.
The question became: what should a particular AI client be allowed to do inside a particular email account?
what I built
Multiple Gmail accounts, multiple Outlook accounts, one MCP endpoint, per-account and per-installation permissions, account pause and revocation, and permission gating around search, read, draft, and send-request actions.
Message bodies are fetched from the provider on demand. The system is not designed to retain a mailbox corpus.
architecture
AI client → MCP endpoint → installation permissions → account permissions → Gmail API / Microsoft Graph → provider-side messages and drafts
interesting decisions
Draft hash validation
Immediately before sending, the provider-side draft is fetched again and compared with the approved draft hash. If it changed, sending is blocked.
Permission boundaries
Search, read, draft, and send-request permissions are checked per account and per AI client installation instead of treating OAuth as blanket access.
verification
The August 2, 2026 verification run documented 227 passing unit and security tests, along with clean lint, TypeScript, and production build checks.
limitations
- Invite-only prototype.
- Public GitHub repository is a technical case study.
- Application implementation remains private.
- Gmail OAuth remains subject to Google's restricted-scope verification requirements.
- No independent security audit has been claimed.