Mailbox MCP Server

Mailbox

One REST API for all your mailboxes, whichever provider they are at, a Python client for it and an MCP server on top. Scripts, tools and AI agents work with your mail through one door you control.

Status: beta, version 0.4.1. Usable with real accounts. A breaking change of the API or the configuration is announced in the changelog. Stored data is carried forward by migrations.

Contents

• What it is for
  • Working with an AI agent in your mail
  • More uses
• What it can do today
• Providers
• The four packages
• Getting started
• Documentation
• Development
• Licence

What it is for

Most people and small companies have more than one mail account: a personal address at GMX or web.de, a company mailbox at Microsoft 365, an info@ address at some hoster, maybe a Gmail account. Each speaks its own dialect: IMAP here, Microsoft Graph there. Every tool that wants to work with mail has to learn all of them and has to be given the passwords.

Mailbox turns that around. It has four packages. mailbox-service runs on your own machine or server and holds the connections to all accounts. Everything else talks to that service only, through one REST API that looks the same for every provider. mailbox-client is the Python client of that API, for scripts and apps. mailbox-mcp, the MCP server, builds on the client and opens the service to AI agents. mailbox-common holds what the service and the MCP server share. The service decides who may do what. A script, an app or an AI agent gets a token of its own. That token opens exactly the accounts and operations it was given, nothing more.

That makes a few things simple that are hard otherwise.

Working with an AI agent in your mail

Through the MCP server, Claude or another AI agent works with every account you gave it. A few requests it handles today:

• "What came in since last night?"
• "Move every invoice from Company XY to the folder Accounting."
• "Forward today's invoices to our tax advisor."
• "Show me all mail from Company XY between the 1st and the 15th of September."
• "Which mails with attachments came this week? Read the invoice and tell me the amount and the due date."
• "Sum up the mails about the offer and write a draft reply."
• "Mark last week's newsletters as read and move them to the archive."

Its rights decide whether it may send on its own or only leave drafts for a person to send. A grant can also name the recipients it may write to, such as the tax advisor's address alone, and how many mails it may send a day. A mail to anyone else is refused. It can keep the agent to some folders, such as the invoices, and end after a week.

More uses

• Controlled access. A bookkeeping tool reads the invoice folder of one account and nothing else. It may neither send nor delete, and it never sees a mail password, only its own token.
• Automation across accounts. A script files invoices, archives newsletters and forwards order confirmations. It works the same way for GMX, Microsoft 365 and your own mail server.
• Sending with guard rails. A newsletter job sends from info@, to a fixed list of recipients, at most a few times a day.
• React instead of polling. Signed webhooks tell your own system about new mail, a ticket tool or an internal dashboard for example.
• Traceable. Every send is recorded in the audit: by whom, when and under which grant. So is every sign-in and every change to users, tokens, roles, accounts and webhooks.
• One inbox for your own tools. A dashboard or a small internal app lists, searches and answers mail from every account with one client.
• Self-hosted. The service stores the credentials encrypted. Mail goes straight from the provider to you, never through a third party's cloud.

What it can do today

• Accounts: IMAP with SMTP for sending, set up from the address alone (autodiscovery), JMAP servers such as Fastmail and Stalwart, POP3 for mailboxes without IMAP (the inbox only), and Microsoft accounts (Outlook.com, Microsoft 365) over Microsoft Graph with OAuth sign-in. Credentials are encrypted (AES-256-GCM) under a master key kept outside the database.
• Reading: folders, lists and search, one account or all at once, messages as text and HTML, attachments, the raw source. Message ids stay the same when a message moves, kept by a background sync.
• Writing: flags, moving, deleting, folders, drafts, and sending with reply, reply to all and forward. A retried send is not sent twice (Idempotency-Key).
• Changes: a change feed names each message created, changed or deleted since a point you keep, in IMAP, JMAP, POP3 and Microsoft accounts, flags set in other mail clients included where the IMAP server offers CONDSTORE. Webhooks post the same events, signed, to a URL of your choice, a host in your local network included.
• Users and rights: users, roles and grants per account and per operation, down to folders, grants that expire, API tokens, limits on sending (allowed recipients, sends per day), an audit of every send and an audit of administration.
• MCP server: reads, sorts, writes drafts, sends and tells what is new, over stdio or streamable HTTP, offering only the tools its token may use. Mail content reaches the model marked as foreign text.
• Configuration UI in the browser under /ui: accounts, users, rights, tokens, reading and writing mail, the send audit, webhooks, the status of accounts and sync, the audit of administration, the service log.
• Operation: encrypted backup and restore, container images, compose files for operation with HTTPS through Caddy and MCP server instances.

Planned next: threads across folders. The order is in docs/ROADMAP.md.

Providers

Not supportable: Tuta, which offers no IMAP and no API. Details per provider in docs/CONCEPT.md, section 5.3.

The four packages

A release publishes all four to PyPI under the same version, as benethos-mailbox-service, benethos-mailbox-client, benethos-mailbox-mcp and benethos-mailbox-common, and the service and the MCP server as container images on ghcr.io. The MCP server is listed in the official MCP Registry as io.github.benethos-hub/benethos-mailbox-mcp, and every release carries it as a Claude Desktop extension, benethos-mailbox-mcp.mcpb. Each package has its own README. It explains how to install, start and configure it, for the service and the MCP server with the container image and the compose file.

Getting started

1. Start the service and create the first user: packages/mailbox-service.
2. Open http://127.0.0.1:8080/ui, sign in as admin with the one-time password, choose a password of your own and connect your accounts.
3. Give your scripts or your AI agent a user with the rights they need and a token (step by step), and for an agent, add the MCP server to it. Claude Desktop installs it as an extension from the release, other clients start it with uvx: packages/mailbox-mcp.

Documentation

• docs/CONCEPT.md: the design, the API, the security model
• docs/ROADMAP.md: the phases and what is done
• docs/ARCHITECTURE.md: the layers, the modules and the rules for new code
• docs/LIMITS.md: every rate limit and how they work together
• docs/PERMISSIONS.md: users, roles and rights
• docs/AUTHENTICATION.md: signing in to the configuration UI, the password and the second factor
• docs/LOGGING.md and docs/AUDIT.md: the service log and the audit of administration
• docs/UI.md: how the pages of the configuration UI look and behave
• docs/REFACTORING.md: how the layout of the code came to be
• docs/IDEAS.md: collected, not decided
• docs/MICROSOFT.md: connecting Microsoft accounts
• docs/GOOGLE.md: connecting Gmail accounts, the Google client of your own
• docs/openapi.json: the API contract. A running service shows it at /docs
• CHANGELOG.md
• SECURITY.md: how to report a vulnerability
• containers/: the Dockerfiles, and compose files for operation, for development and for a test mail server

Development

One uv workspace, one lockfile, four distributions.

uv sync
uv run pytest -q
uv run ruff check . && uv run ruff format --check .
uv run mypy

The tests run offline. CI runs these checks too. It also installs the packages on the lowest versions their pyproject.toml allows, runs the tests there and checks those versions against OSV for known vulnerabilities, and checks the brand assets in assets/brand/ with the organisation's workflow. Every one of these checks has to pass before a change reaches main. Manual checks against real test accounts live in live/ and are described in CLAUDE.md, which also holds the working rules for this repository.

Licence

MIT, see LICENSE.