mcp-agent-mail

Encrypted email + PGP for AI agents, as a local MCP server. Read, send, and reply to GPG-encrypted mail; manage a contact book with key provenance; archive and search everything you read.

Built to be called by any MCP-aware agent harness (Claude, opencode, etc.) over stdio. Works with Gmail and any IMAP/SMTP provider.

Standalone rewrite. This project is a clean, hardened rebuild of the earlier prototype, built around one hard-won invariant: consequence-critical bytes must never enter the model's context window. The prototype earned that invariant the hard way — a low-parameter local model reading and repeating high-bit key material corrupted it. This rewrite carries the invariant over verbatim; several deliberately-added behavioral changes are flagged with ⚠ below.

Why this design

The model is the least reliable component in the pipeline. Small local models (the ones that run on consumer hardware) occasionally mangle byte-exact data — a quoting battle, a garbled regex, a corrupted key block. For ordinary text that is noise; for private keys, passphrases, and ciphertext it is unacceptable. So the design rule is:

Anything whose bytes must not change never enters the model's context.

That rule, not a threat model, is the primary reason for every invariant below. They are reliability controls: they keep high-consequence data on the side of the boundary the model cannot corrupt. The same controls also happen to harden the tool against an untrusted model and a partially-trusted host — useful, but derived. The primary enemy here is entropy, not malice.

There's a third, pragmatic face of the same rule: context is the scarcest resource on a local model, and high-entropy bytes are its worst possible consumer. A key block, an armored ciphertext, a base64 payload is nearly incompressible noise — it burns tokens at maximal density and returns zero usable signal. Keeping it out of the context window doesn't just protect its bytes; it wins back expensive context the model would otherwise waste repeating and re-mangling noise it was never going to use.

• Inbound PGP key interception — public keys that arrive by email are imported and linked to the sender's contact before the body reaches the model. The block is replaced by a notice, never shown.

• No silent downgrade (⚠) — email_send encrypts by default. If no key is on file for the recipient it refuses to send; the model must explicitly choose encrypt=False to send in the clear. A model that guesses wrong silently is worse than one that stops and asks.

• Full fingerprints only — 16-char key IDs are rejected everywhere (Evil32 collision attack).

• Secrets are opaque — SecretString wraps passphrase/password; reprs, logs, tracebacks show ***. Private keys are never exported.

• Fail-fast config (⚠) — missing secrets abort startup with a list, never a warning and never a fallback default. A half-configured server is a guesser; fail-fast is fail-safe.

• Immutable identity entries (⚠) — the agent (EMAIL_ADDRESS) and the owner (OWNER_EMAIL) each have exactly one contact entry, fixed at setup. No tool can add, remove, re-key, or clear them, and the server refuses to start until both exist. Identity is a provisioning decision, not a model action.

Related MCP server: Mail MCP Server

Tool surface

Requirements

• Python 3.13+

• uv

• GnuPG — Gpg4win on Windows, gnupg2 on Linux/macOS

• An email account with IMAP/SMTP app-password access (Gmail: enable 2FA, create an app password)

Install

git clone <repo-url> mcp_agent_mail
cd mcp_agent_mail
uv sync

Setup

1. Generate a dedicated agent PGP key pair:

   gpg --full-generate-key

2. Create .env from the template and fill it (see .env.example for passphrase quoting pitfalls):

   Copy-Item .env.example .env

   Required: EMAIL_ADDRESS, OWNER_EMAIL, EMAIL_PASSWORD (app password), GPG_KEY_ID (the agent key's full 40-char fingerprint), GPG_PASSPHRASE.

   The env backend stores secrets in plaintext on disk. For production use, follow docs/SECURITY.md and plan to move to a credential-store backend (M2 roadmap: Windows Credential Manager / KeePassXC / gpg-agent pinentry).

3. Sanity check:

   uv run mcp-agent-mail setup      # deps + key presence + secret status
   uv run mcp-agent-mail doctor     # offline config diagnostics
   uv run mcp-agent-mail doctor --live   # opt-in IMAP + keyring checks

4. Provision the identity entries (see below). doctor shows a [!!] line and the server refuses to start until they exist.

Identity & setup

Two contact entries pin the identity boundary of the whole system and are immutable from the tool surface. They are provisioned once, by hand (or by a future owner-facing CLI) — never created or edited by the tools:

• Agent entry — the record whose email equals EMAIL_ADDRESS. It is the agent's own identity inside the contact book and carries the agent key fingerprint (GPG_KEY_ID). Conventional surname: agent of <owner given name>.

• Owner entry — the record whose email equals OWNER_EMAIL. That is the human running the server.

contact_add, contact_remove, contact_link_key, contact_set_fingerprint and contact_clear_key refuse to touch either entry (status protected). The server fails fast at startup if an identity email is absent or held by more than one record.

Minimal data/contacts.json with both identities provisioned:

{
  "Hermes": {
    "added": "2026-09-13T00:00:00",
    "given_name": "Hermes",
    "surname": "agent of Chris",
    "email": "agent@example.com",
    "gpg_key_fingerprint": "AAAABBBBCCCCDDDDEEEEFFFF0000111122223333",
    "key_source": "keyring_uid_match",
    "key_linked_at": "2026-09-13T00:00:00",
    "key_cleared_at": "",
    "notes": "",
    "updated": "2026-09-13T00:00:00"
  },
  "Chris": {
    "added": "2026-09-13T00:00:00",
    "given_name": "Chris",
    "surname": "Example",
    "email": "owner@example.com",
    "gpg_key_fingerprint": "4444555566667777888899990000AAAABBBBCCCC",
    "key_source": "keyring_uid_match",
    "key_linked_at": "2026-09-13T00:00:00",
    "key_cleared_at": "",
    "notes": "",
    "updated": "2026-09-13T00:00:00"
  }
}

The added/updated timestamps are ISO 8601. key_source values: "keyring_uid_match", "manual", "cleared".

Usage

Run as an MCP server (stdio — what harnesses expect)

uv run mcp-agent-mail

To serve over HTTP instead:

uv run mcp-agent-mail serve --http --port 8000

Register in an MCP client

Point the client at the project — the server reads .env directly and does not need the environment pre-seeded:

{
  "mcpServers": {
    "mcp-agent-mail": {
      "command": "uv",
      "args": ["--project", "C:/path/to/mcp_agent_mail", "run", "mcp-agent-mail"]
    }
  }
}

HTTP transport with curl (manual probing)

The HTTP mode speaks the standard MCP streamable-http transport. A raw HTTP call is not a single request-response: MCP is session-based. You initialize, then send JSON-RPC messages that carry an Mcp-Session-Id header, and results come back as Server-Sent Events (text/event-stream), not plain JSON.

Start the server:

uv run mcp-agent-mail serve --http --port 8000

1. Initialize the session — the mcp-session-id header in the response is the session token for every following call:

curl -s -D - -X POST http://127.0.0.1:8000/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"initialize","id":1,"params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"curl-test","version":"1.0"}}}'

2. Signal the client is initialized (protocol requirement; produces no response body):

curl -s -X POST http://127.0.0.1:8000/mcp \
  -H "Content-Type: application/json" \
  -H "Mcp-Session-Id: <session-id>" \
  -d '{"jsonrpc":"2.0","method":"notifications/initialized"}'

3. Call a tool — list the three most recent inbox messages:

curl -s -X POST http://127.0.0.1:8000/mcp \
  -H "Content-Type: application/json" \
  -H "Mcp-Session-Id: <session-id>" \
  -d '{"jsonrpc":"2.0","method":"tools/call","id":2,"params":{"name":"email_check_inbox","arguments":{"limit":3}}}'

The result arrives as an SSE frame; the useful payload is the "text" field inside the JSON (example, sanitized):

event: message
data: {"jsonrpc":"2.0","id":2,"result":{"content":[{"text":"{\n  \"folder\": \"INBOX\",\n  \"count\": 1,\n  \"messages\": [\n    {\n      \"uid\": \"100\",\n      \"folder\": \"INBOX\",\n      \"from\": \"John Smith <john.smith@example.com>\",\n      \"to\": \"agent@example.com\",\n      \"subject\": \"Hello\",\n      \"date\": \"Tue, 01 Jan 2026 10:00:00 +0100\"\n    }\n  ]\n}","type":"text"}],"isError":false}}

Reading runs the same PGP pipeline as stdio: inbound ciphertext is decrypted, sender key blocks are intercepted and linked in the contact book, and key material never reaches you as raw bytes:

curl -s -X POST http://127.0.0.1:8000/mcp \
  -H "Content-Type: application/json" \
  -H "Mcp-Session-Id: <session-id>" \
  -d '{"jsonrpc":"2.0","method":"tools/call","id":3,"params":{"name":"email_read","arguments":{"uid":"100"}}}'

stdio is what MCP harnesses expect and what the server runs by default. HTTP serves manual probing, remote access, and web-based clients — it changes the transport only, never which tools are exposed.

Tighten your harness's tool-use prompt

Small local models occasionally misread an otherwise-unambiguous tool contract on first use — e.g. batching two local MCP calls into one invocation array and getting the batch rejected. Don't fight this in the server; fix it in the harness. Add a one-line tool-use rule to the harness's always-injected context (SOUL.md, AGENTS.md, CLAUDE.md, .cursorrules — whichever the harness loads from user-owned data, so updates don't clobber it):

One local tool invocation is exactly one entry per tool_call; only connectors__-type names may be batched together. Mixed or multi-local batches are rejected.

The rule costs tokens once per session prefix, not per call, and removes the whole class of first-tool-call failures.

CLI commands

Development

uv run ruff check .
uv run pytest -q

127 tests cover: secret handling, fail-fast config, identity immutability, fingerprint/key-block validation, contact provenance, archive dedup/search, the outbound encryption gate, and the full tool surface. No .env or real account is needed — tests seed fake secrets and mock the transports.

Configuration reference

Documentation

• docs/SECURITY.md — threat model, guarantees, trade-offs

• docs/ARCHITECTURE.md — module map and data flow

License

MIT © 2026 Christof Milius