What it feels like

You talk to Claude. Claude talks to your mailbox. You stay in the loop on anything that matters.

"What did I miss from the climbing group this week?" → Claude searches the local mirror, reads the thread, summarizes it. No prompt — reading is safe.

"File all the newsletters under Reading and mark them read." → Claude moves and marks them. Organizing is gated, but quiet.

"Reply to Alice that I'm in for Saturday, and send it." → A Touch ID prompt appears: To: alice@example.com · Subject: Re: gear list. You tap. It sends. You didn't.

Every read is served from a local SQLite mirror, so it's fast and works offline. Every write is governed by a per-tool policy. Every send re-prompts, every time, showing the literal recipients — that fingerprint tap is the line between "Claude drafted it" and "Claude sent it."

What Claude can do

34 tools, grouped by what they touch. Reads run free; everything that changes state is deny-by-default and Touch-ID gated.

Full list with descriptions: docs/cli-reference.md.

Why it's safe

proto-mcp is built so that an LLM driving your mailbox is a convenience, never a liability. The guarantees that make that true:

• 🔐 Your fingerprint on every send. Each write fires a native prompt showing the literal recipients and subject. mail_send has a TTL of zero, so it re-prompts every single time. No blanket approvals for sends.
• 🛡️ Default-deny by construction. Unknown tools don't run. A tool with no policy entry fails to register — you can't accidentally ship an unguarded write.
• 🍎 Signed, notarized, and self-checking. Hardened-runtime, Developer-ID-signed, Apple-notarized binaries, plus a SHA-256 integrity check at startup that refuses to run a swapped daemon.
• 🔒 Locks when you walk away. Screen lock, sleep, or an idle timer zero the in-memory session; resuming takes Touch ID.
• 🧾 Honest, redacted audit log. Every call is logged — secrets scrubbed, bodies reduced to {sha256, bytes}, recipients kept literal so the verification chain stays truthful.
• 🏠 Local-only. The daemon listens on a 0600 Unix socket, never a network port. Mail content goes to Proton over TLS; nothing else leaves.

What a prompt actually looks like:

┌──────────────────────────────────────────────┐
│ Send mail_send?                              │
│                                              │
│ To: alice@example.com                        │
│ CC: charlie@example.com                      │
│ Subject: Re: gear list                       │
│                                              │
│ [ Cancel ]              [ Send & Touch ID ]  │
└──────────────────────────────────────────────┘

The full threat model — including the risks proto-mcp doesn't defend against — is in docs/security.md. Read it before you point this at a live mailbox.

How it works

One background daemon holds your Touch-ID-unlocked session and serves every tool over a local socket. Claude Desktop and Claude Code each attach through a tiny forwarder, so they share one session: unlock once, use everywhere; lock once, everything locks.

Claude Desktop ─┐                          ┌─ go-proton-api + GPG
Claude Code ────┼─ shim ─ socket ─ protonmcpd ┼─ SQLite mirror + FTS5
                ┘     (0600)               └─ Touch ID + policy + audit

The full design — every binary, package, and the local mirror — is in docs/architecture.md.

Build from source

Requires macOS 13+, Go 1.26.4+, and Xcode Command Line Tools (for swiftc).

git clone https://github.com/just-an-oldsalt/proto-mcp.git
cd proto-mcp
make all                          # builds bin/* + the Swift helpers
./bin/protonmcp login
./bin/protonmcp backfill
./bin/protonmcp daemon install
./bin/protonmcp install

Source builds are ad-hoc signed by default and work fully (the Touch ID gate, policy, audit, and lock/unlock all run the same). For a locally-signed build, see scripts/signing-setup.md.

Good to know

• macOS only. The keystore and biometric helpers use Security.framework, LAContext, and AppKit. Linux builds compile for testing, but the auth flow won't work.
• Be a good Proton citizen. proto-mcp currently sends Proton Bridge's AppVersion header while a dedicated identifier is requested from Proton (see docs/proton-appversion-request.md). Don't rate-abuse, scrape, or run multi-account automation through it — anything that violates Proton's Terms is no less a violation for borrowing Bridge's header.
• Cached bodies are plaintext-in-SQLite. Decrypted message bodies are cached locally (TTL-bounded, secure_delete on). On a stolen, imaged disk that's recoverable cleartext until purged. Envelope encryption (SQLCipher) is a post-1.0 item. protonmcp purge --older-than 7d --vacuum shrinks the window now.
• Personal use. Built for one person and their mailbox on their Mac.

Documentation

Issues, defects, and roadmap are tracked in Jira (project PROTO), the source of truth. TODO.html and DEFECTS.html are retained as historical design records from the build-out.

Quickstart

brew tap just-an-oldsalt/proto-mcp
brew install --cask proto-mcp

protonmcp login            # Proton SRP password + 2FA + key unlock
protonmcp backfill         # one-time: pull your message envelopes into the local mirror
protonmcp daemon install   # register + start the background daemon
protonmcp install          # connect it to Claude Desktop + Claude Code

Restart Claude, and the tools show up under protonmcp in /mcp. That's it — signed, notarized binaries, no Gatekeeper warning, no network listener.

Prefer to build it yourself? See Build from source.

Configuration

Tune per-tool policy, rate limits, allowed recipients, the idle-lock timer, and the cached-body TTL with a single YAML file. For example, to cap LLM-driven sends and restrict them to one domain:

tools:
  mail_send:
    decision: prompt
    rate_limit: 5/hour
    allowed_recipients: ["@mydomain.com"]
idle_lock_minutes: 30

Full reference, plus locking and the audit/observability commands: docs/configuration.md.

License & acknowledgements

GPLv3 — see LICENSE. proto-mcp depends transitively on proton-bridge (also GPLv3) via go-proton-api.

• Proton AG for proton-bridge and go-proton-api, on which the entire crypto + transport layer rests.
• Anthropic for the Model Context Protocol and the Claude clients this server targets.
• Every defect that took the shape it did because cmd-r, claude-review, claude-security-review, or a live testing session looked at the code more carefully than I would have alone.