IMAP Mailbox

imap-mcp

A Model Context Protocol (MCP) server for any IMAP mailbox. It speaks IMAP rather than one vendor's API, so it works with whatever provider you already have.

Lets MCP clients like Claude Code, Claude Desktop or Codex read and search your mail, organise it into folders, save attachments and draft replies — with every message fenced as untrusted content, and the write tools off unless you turn them on.

Eleven tools, not fifty: a mail account is a workflow, not an API surface, so related operations are folded into one tool with a mode rather than split across many. And eleven is the ceiling, not the floor — IMAP_ALLOW_TOOLS=essential registers a curated six instead, and under the read-only default that narrows to four. See choosing which tools load.

What makes it different

It cannot send mail. That is the feature. An agent with access to private data, exposure to untrusted content, and a channel to the outside world is exploitable by anyone who can put a message in the inbox — the pattern that produced EchoLeak, where one crafted email exfiltrated internal data from Microsoft 365 Copilot with no user interaction. This server has the first two and deliberately not the third. save_draft writes the reply into your Drafts folder; you send it from your own mail client. No amount of clever text in a message can make this server post anything anywhere.

Writes are off until you turn them on. With only IMAP_HOST, IMAP_USER and IMAP_PASSWORD set, the server registers six read tools and nothing else. The mailbox tools appear with IMAP_READ_ONLY=false — note the default is true, the opposite of the other servers in this family, because this one reaches a mailbox. Tools that are off are not registered at all — a capability the model cannot see is one it cannot be talked into using.

Mail is treated as hostile input, because it is. Anyone in the world can put text in your inbox. Message bodies are fenced between markers carrying a per-call random nonce, and every line inside them is prefixed with that nonce, so the "this is data" signal does not stop at the edges of a long forwarded thread. A reminder follows the block, because otherwise the last instruction-shaped sentence in the model's context is the attacker's. Zero-width characters and directional overrides are stripped before the model sees anything, hidden HTML elements are dropped on a best-effort basis (the fencing, not the stripping, is what carries the weight), and markdown image syntax — inline and reference style — is defused so a rendering client cannot be made to fetch a tracking URL.

That covers folder names too: a folder name is chosen by whoever created the folder, which on a shared mailbox is not necessarily you. list_mailboxes returns the name twice — path exactly as the server spelled it, because that is the handle every other tool takes, and display_name cleaned up for reading, with a warning on the entry when the two differ.

Alongside the message you get a server-side assessment: the SPF/DKIM/DMARC verdicts with the authserv-id they came from, which prompt-injection shapes matched, and which words mix Latin with Cyrillic or Greek letters. When something matches, the warning is the first thing in the result rather than a field buried in JSON.

Those verdicts carry a forgeable flag, and by default it is always true. A sender can write an Authentication-Results header of their own, and if your provider does not add one, theirs is the only one there — nothing inside the message distinguishes the two. Set IMAP_TRUSTED_AUTHSERV_ID to the id your provider stamps (it is the first token of the header on any message you already have) and only that id counts as authentic. Until you do, spf=pass is reported as what it is: a claim, from a header anyone could have written.

"New mail" that actually works. The server tags messages it has handed over with a custom IMAP keyword (AiSeen by default), so list_new_messages returns each message once. The human \Seen state is never touched — everything is read with BODY.PEEK.

Deleting and moving ask a person. Where the client supports MCP elicitation, delete_messages, move_messages and deleting a folder raise a real dialog that the model cannot answer on its behalf. Where it does not, they fall back to a two-call token — and say so, rather than implying somebody approved. ELICITATION=false takes that fallback deliberately; it never removes the guard. See Asking a person.

Requirements

• Node.js 22 or newer
• An IMAP account. Providers with two-factor authentication generally need an app-specific password.

Configuration

Booleans are compared against the literal string true; 1, yes and True are not true. IMAP_READ_ONLY is the mirror image: only the literal false turns it off, so a typo leaves the write tools unregistered.

Choosing which tools load

IMAP_ALLOW_TOOLS and IMAP_DENY_TOOLS take comma-separated tool names; a trailing * matches a whole family. essential is a curated preset of six — list_mailboxes, list_new_messages, list_messages, get_message, set_message_flags and move_messages. Four of those are read tools, so it stays useful under the read-only default.

IMAP_ALLOW_TOOLS=essential
IMAP_ALLOW_TOOLS=list_new_messages,get_message,move_messages
IMAP_DENY_TOOLS=delete_messages

One boundary the list cannot draw: move_messages copies as well as moves (mode: "copy"), and the two are one tool. Denying move_messages removes both; there is no way to keep moving and forbid copying, or the other way round. Both modes ask for confirmation.

An entry that matches no tool aborts startup and names it, so a typo cannot silently hide a tool — an absent tool is not something anyone traces back to an environment variable. A filtered tool is never registered, so it is absent from tools/list and unknown to tools/call alike, exactly like a write tool under IMAP_READ_ONLY.

It covers tools. The attachment resources this server also exposes are not filtered.

If you run several of these servers at once, mcp-hub is the other answer — its /hub endpoint replaces every server's tools with six meta-tools. The password is deleted from the process environment as soon as it is read, so it is not visible to child processes or in /proc/<pid>/environ.

Without IMAP_DOWNLOAD_DIR this server never writes to the filesystem. The three size limits are separate on purpose, because they answer three different questions: IMAP_MAX_ATTACHMENT_BYTES protects the model's context window, IMAP_MAX_DOWNLOAD_BYTES protects your disk, and IMAP_MAX_EXTRACT_BYTES bounds how much hostile input one parser is handed. Raising any one of them is not a request to raise the others.

The server starts without credentials on purpose — it completes the handshake and lists its tools, and every call then fails with setup instructions instead of reaching a server.

Installation

Claude Code

claude mcp add imap-mcp \
  -e IMAP_HOST=imap.example.net -e [email protected] -e IMAP_PASSWORD=… \
  -- npx -y @ni-c/imap-mcp

Claude Desktop

{
  "mcpServers": {
    "imap-mcp": {
      "command": "npx",
      "args": ["-y", "@ni-c/imap-mcp"],
      "env": {
        "IMAP_HOST": "imap.example.net",
        "IMAP_USER": "[email protected]",
        "IMAP_PASSWORD": "…"
      }
    }
  }
}

Codex

[mcp_servers.imap-mcp]
command = "npx"
args = ["-y", "@ni-c/imap-mcp"]
env = { IMAP_HOST = "imap.example.net", IMAP_USER = "[email protected]", IMAP_PASSWORD = "…" }

Docker

docker run --rm -i \
  -e IMAP_HOST=imap.example.net \
  -e [email protected] \
  -e IMAP_PASSWORD=… \
  ghcr.io/ni-c/imap-mcp

Saving attachments needs a writable directory, and the image runs as uid 1000 — so a bind mount has to be owned by it on the host: -e IMAP_DOWNLOAD_DIR=/data -v "$PWD/attachments:/data" with chown 1000:1000 attachments. Without IMAP_DOWNLOAD_DIR the container never writes anything.

Through mcp-hub

A client that cannot spawn a local process — ChatGPT connectors, Claude on the web, Cursor, LibreChat — reaches imap-mcp through mcp-hub: one container serves many stdio MCP servers over Streamable HTTP, with an OAuth 2.1 login behind a single password and long-lived tokens for the clients that cannot do OAuth. Its /hub endpoint puts every server behind six meta-tools, so one connector reaches all of them without N×tool schemas in the model's context, and it speaks both protocol revisions — a question this server asks travels through it to the person at the far end.

Its /config/mcp.json uses Claude Code's format, so the entry is the one you already have:

{
  "mcpServers": {
    "imap-mcp": {
      "command": "npx",
      "args": ["-y", "@ni-c/imap-mcp"],
      "env": {
        "IMAP_HOST": "imap.example.net",
        "IMAP_USER": "[email protected]",
        "IMAP_PASSWORD": "…",
        "IMAP_ALLOW_TOOLS": "essential"
      },
      "denyTools": ["delete_messages"]
    }
  }
}

allowTools and denyTools there are the hub's own per-server filter, which is not the same thing as *_ALLOW_TOOLS in env — the difference, and the mistake it invites, are in the client guide.

Tools

Read — always registered

Mailbox — needs IMAP_READ_ONLY=false

👤 raises a dialog the model cannot answer · 🔒 needs a confirmation token: call once to receive one, then again with it.

copy is confirmed as well as move, because the thing that cannot be taken back is not the deletion — it is the disclosure. A destination is a free-form folder name, and on a shared account or a public namespace one call hands every message to everyone who can read it, leaving the source folder untouched. For the same reason set_message_flags refuses to add \Deleted: it is half a deletion, and the next client to close the mailbox may finish it. Use delete_messages, which asks.

Neither a confirmation nor a dialog quotes a mailbox name inside its own sentence — folder names come from the account, which on a shared mailbox means a colleague chose them.

Structured output

Every tool declares an outputSchema and answers with structuredContent alongside the text block, so a client can use the result without parsing prose:

{
  "untrusted": true,
  "source": "imap",
  "mailbox": "INBOX",
  "total_matching": 214,
  "offset": 0,
  "returned": 25,
  "next_offset": 25,
  "messages": [{ "uid": 4711, "subject": "…", "from": "…", "seen": false }],
}

Every tool that reports anything out of the mailbox carries untrusted: true and source: "imap" as fields — a sender display name, a folder name a colleague chose and an attachment filename are all attacker-controllable, and they reach the model through the listing tools long before anyone opens a message. Only get_server_info and the five write tools are without it: those report this server's own configuration, or what it just did with the uids it was given.

get_message and a text attachment keep the per-call nonce fence in the text block — the structured half states the same fields, so a client is not made to parse the fence to find them. An image attachment keeps its bytes in the content block, where a client renders them, rather than repeating the base64.

A refusal is now an error result: an attachment the policy rejects, one whose bytes are an executable whatever it claimed, one too large to inline. Each was a plain result that read like an answer.

Attachments are also available as MCP resources at imap://message/{uid}/part/{partId}, which matters where the server has no useful filesystem. The resource path runs the same allowlist, size limit and magic-byte check as the tool — it is not a second, unguarded door.

Not exposed, on purpose

No sending, no SMTP, no raw IMAP passthrough, no APPEND of arbitrary MIME, no HTML composition, no OAuth2, and no OCR — a scanned PDF has no text to extract and says so rather than guessing. The first is the whole security argument (see SECURITY.md); the second would make every guard here optional; the last is planned but needs a test account before it ships.

And one thing the tool filter does not cover: attachment resources. IMAP_ALLOW_TOOLS narrows tools/list, not resources/list, so a server with a narrow allow list still serves those. IMAP_DOWNLOAD_DIR and the content-type allowlist are what constrain them — worth knowing before concluding that a filtered install reaches less of the mailbox than it does.

Safety

• Every result carrying mailbox content is marked untrusted, message bodies additionally fenced with a per-call nonce and marked line by line.
• Attachments pass two independent gates. The declaration is checked against a content-type allowlist, an executable-extension refusal list and a size ceiling; the bytes are then checked against magic numbers. An executable renamed to .pdf and declared application/pdf clears every declaration check and fails on its bytes — including when saving to disk, where it would be more dangerous, not less.
• A part_id must come from a listing call, so the body cannot be pulled out through the attachment tool and escape its framing.
• Documents are parsed in a process that can be killed. mode: "text" reads a PDF or Office file with a bundled PDF.js and a ZIP reader — the only place this server parses a binary a stranger sent. It runs in a child process with a heap limit and a timeout, its stdout discarded rather than shared with the JSON-RPC transport, PDF.js's eval support off, compressed streams measured against a ceiling before PDF.js inflates them, and an entry allowlist that decides what is decompressed before the buffer is sized. Nothing in that path touches the network or the filesystem.
• Extracted text says what it is. Extraction returns every text-drawing instruction in a file, including text set below one point or drawn in the colour of the paper, and returns nothing that was drawn as a picture. The result says so above the fence, because "the document says X" is otherwise a claim the user has no way to check.
• Downloads cannot escape their directory. The target comes only from the environment, the filename is sanitised, the resolved path is re-checked, and the file is opened with wx and mode 0600 — so nothing is overwritten and no planted symlink is followed.
• Mailbox names, flags and addresses are refused if they contain line breaks. IMAP is a line protocol and a draft is a mail header; a CR is an injection primitive, not a typo.
• TLS is never disabled globally. IMAP_INSECURE_TLS is scoped to the connection it names; NODE_TLS_REJECT_UNAUTHORIZED appears nowhere.
• Every change to the mailbox is logged to stderr with the UIDs and folder — never the subject. stderr is the one channel the model does not read.
• Responses are bounded. Whole items are dropped rather than the JSON being sliced, and the truncation notice comes first so the recovery hint survives.

SECURITY.md has the trust model, what these measures do not cover, and how to report a vulnerability.

Documentation

The full guide, tool reference and security notes live at imap-mcp.ni-c.de (source in docs/).

Development

npm install
npm test
npm run build

The test suite runs against an in-memory IMAP fake, so it needs no server and no network. For a live server to point the real thing at, see CONTRIBUTING.md — it starts a throwaway mailbox in a container.

Releasing

1. Add the CHANGELOG entry and bump package.json.
2. npm run lint && npm run build && npm run test:coverage
3. Commit, then push a signed tag: git tag -s vX.Y.Z -m "vX.Y.Z" && git push origin main vX.Y.Z

The release workflow publishes to npm (Trusted Publishing, with provenance), creates the GitHub release from the CHANGELOG section and updates the MCP Registry entry.

Contributing

Issues, discussions and pull requests are welcome — see CONTRIBUTING.md. For vulnerabilities please use private reporting rather than a public issue; the policy is in SECURITY.md.

License

MIT © Willi Thiel