pylos-mcp

The gate between your mailbox and your LLM.

Anyone in the world can put text in your inbox, and the moment an AI assistant reads that inbox, anyone in the world can put text in front of your assistant. pylos-mcp is an email MCP server built around that fact. It lets Claude, or any MCP client, search, read and draft your mail while treating every message as what it really is, input from a stranger.

It runs on your machine and speaks plain IMAP, so it works with Gmail, iCloud, Yahoo, GMX, Fastmail, mailbox.org, Posteo, Proton via Bridge, or anything self-hosted, and your credentials never leave home. Out of the box it can read and draft. Anything riskier, moving, sending, deleting, is a separate switch that stays off until you flip it.

Why this one

Quick start

Add the server to your MCP client's config. For Claude Desktop that file is claude_desktop_config.json.

{
  "mcpServers": {
    "pylos-mcp": {
      "command": "npx",
      "args": ["-y", "pylos-mcp"],
      "env": {
        "PROVIDER": "mailbox.org",
        "EMAIL_USER": "you@example.com",
        "EMAIL_PASSWORD": "your-app-password"
      }
    }
  }
}

Use an app password rather than your account's regular login password, the table below says which providers insist on one. Restart the client and the read and draft tools appear. Later config changes need the same treatment, a newly enabled capability only registers its tools after a full client restart, and in Claude Desktop toggling the server off and on is not always enough.

Provider setup

Gmail, iCloud, Yahoo and Fastmail refuse regular account passwords over IMAP, so an app password is the only way in. mailbox.org, GMX and Posteo accept the account password, though a dedicated app password is still the wiser choice. Each provider's account settings cover creating one.

Capabilities

Capabilities are independent switches, not a ladder. Reading is always on, drafting starts on, everything else stays off until you list it in CAPABILITIES. A switched-off tier has its tools left out of the tool list entirely rather than merely refused, so a model never even learns a disabled tool exists.

Enable more with a comma-separated list, for example CAPABILITIES=drafts,manage,delete.

Suspicion warnings

The server also tells you what is suspicious about a message. Three detectors annotate get_email results with a line above the content, written entirely in the server's own words and never quoting the content that tripped them.

Warnings: hidden_text (412 hidden characters via display:none), encoded_blob (base64 run of 600 characters)

• Hidden text. Text concealed with the common CSS tricks, display:none, invisible or one-pixel fonts, matching text and background colors, off-screen positioning, aria-hidden. It covers inline styles and attributes, a tripwire rather than a rendering engine. Newsletters legitimately hide short preview text, so the warning fires only past a threshold, unless the hidden text itself contains an instruction-like phrase or an encoded run, which warns at any length. The text stays in the body by default. STRIP_HIDDEN_TEXT=true drops it instead, with a note of how much was dropped.
• Instruction patterns. A deliberately small set of phrases that address an AI as an instruction target, like "ignore previous instructions". Small so that an inbox merely talking about AI stays quiet. Extend it with FLAG_EXTRA_PATTERNS, pipe-separated phrases matched as case-insensitive literals.
• Encoded blobs. Long contiguous base64 or hex runs in the body, reported with their length and never decoded.

Warnings annotate, they never withhold. The message always comes back, and each detector has its own toggle in the reference below.

What this can never do

Mail is attacker-controlled text, so the hard limits live in the architecture rather than in a prompt. No message can talk the server out of any of these.

• No raw HTML ever reaches the model. Bodies come from the plain-text part when one exists or are converted to text otherwise, and invisible characters that could hide instructions from a human reader while staying readable to a model are stripped along the way.
• Untrusted content is fenced. Everything that came from a mailbox, bodies, subjects, sender names, folder listings, Sieve script text, is wrapped in a labeled delimiter before the model sees it, and the delimiter sequence is neutralized wherever it appears inside the content, so a malicious email cannot forge a closing marker and write instructions outside the fence. The few lines that live outside the fence (the metadata line above a message, the one-sentence confirmations a tool returns after it acts, the Warnings line) are server-authored, stripped of hidden characters, collapsed onto a single line and length-capped, and the Warnings line never carries message content at all.
• No bcc field exists anywhere, not on drafts and not on sent mail. A bcc recipient receives a full copy of a message while appearing nowhere in it, and that invisibility is exactly what an injected email would want, a silent extra recipient that no review of the draft or of the sent copy could ever catch. The field is absent rather than guarded, so there is nothing to talk the model into.
• Deleting a message moves it to Trash. There is no expunge and no permanent-delete option, and the tool result never claims a permanence this server does not offer.
• Sieve access is read-only, permanently. Server-side filter rules can forward, auto-reply and notify, each an exfiltration channel that survives revoking the app password or uninstalling this server. Scanning uploaded scripts for dangerous commands would only be safe if this project's parser agreed with the mail server's parser exactly, and any disagreement between the two is a bypass, so write access is left out entirely rather than defended.

Sending is the other risky door, so it starts closed even once the send capability is on. Until SEND_ALLOWLIST says who may be addressed, every send is refused, and the refusal names the two ways to open the gate. Choosing SEND_ALLOWLIST=* allows anyone, visibly and on purpose.

Fencing reduces prompt-injection risk, nothing eliminates it. The model still reads text written by strangers, so treat every response that includes message content as untrusted input rather than ground truth.

Configuration reference

All configuration is environment variables, validated at startup. Invalid configuration fails immediately with an actionable message, never partway through a conversation.

Maintenance expectations

pylos-mcp is built for the author's own daily use and maintained on that basis. Issues and pull requests are welcome, and CONTRIBUTING.md carries a wishlist of directions that would genuinely help. The scope stays narrow on purpose, so if you need something wider than the security posture allows, fork away. The codebase is deliberately small enough to make that pleasant.

Development

npm install
npm test                  # unit and MCP-layer tests, entirely offline
npm run test:integration  # starts a disposable local Dovecot container, tests against it, tears it down
npm run build

No test in this project connects to a real mailbox, in development or in CI. npm test runs in-process fakes, and npm run test:integration brings up its own local Dovecot container over Docker, seeded with synthetic fixture messages, and removes it when the run finishes.