mail-mcp
An Apple Mail MCP server that enables reading emails, listing mailboxes, searching messages, handling attachments, and creating drafts, with no ability to send mail.
README
mail-mcp
An Apple Mail MCP server with a supply chain small enough to read in an
afternoon: zero runtime dependencies, no network code, no shell, and no way
to send mail. It drives Mail.app through its scripting interface (JXA over
osascript) and exposes thirteen tools for reading mail, listing structure,
handling attachments, and creating drafts.
Everything load-bearing here traces to measurements on real hardware,
recorded in docs/apple-mail/ and re-checkable with node research/verify.mjs.
The security posture, stated plainly
This server cannot send mail, and that is the design, not a gap.
- There is no send tool, no send code path, and no code that could grow one
quietly: a test walks the dependency closure and fails if anything imports
node:net,node:tls,node:http,node:https,node:dgram, callsfetch, or contains asend(call in a Mail script. Do not add a send tool; a draft-then-send flow with an in-band confirmation token does not actually establish human review, because a model can chain both calls in one turn. The user reviews drafts in Mail.app and presses Send there. - Drafts are the only mutation. No delete, no move, no mark-as-read, no rule editing. Email content is attacker-controlled input to any model reading it; the small write surface bounds what injected instructions can do.
- No credentials, no sockets, no Keychain. Mail.app does its own syncing; this server never sees a password or opens a connection.
save-attachmentis the only filesystem writer. It writes only under an allowlisted root (MAIL_MCP_SAVE_ROOT, default~/Downloads), resolves symlinks before checking containment, sanitizes filenames from mail, and never overwrites. There is deliberately no tool that returns attachment bytes into the conversation.- No user value is ever interpolated into script text. Scripts are static
files on disk; values travel as
argv, spawned without a shell. package.jsonhas nodependencieskey. Ever. A test enforces this too.
Requirements
- macOS with Mail.app configured (built against Mail 16.0 on macOS 26.x)
- Node.js 20 or newer
- Automation permission for whatever runs the server (see below)
Install and register
No build step and no dependencies to install:
git clone https://github.com/007hacky007/mail-mcp.git
cd mail-mcp
Register with Claude Code (available in every session):
claude mcp add --scope user mail-mcp -- node "$(pwd)/src/index.mjs"
Or for the current project only:
claude mcp add mail-mcp -- node "$(pwd)/src/index.mjs"
For any other MCP host, configure a stdio server running
node /absolute/path/to/mail-mcp/src/index.mjs.
Verify with the host's tooling (claude mcp list should show
mail-mcp ... Connected), then call the doctor tool.
Automation permission
On the first call into Mail, macOS prompts to allow the host application to
control Mail.app. Approve it. To check or repair the grant: System Settings >
Privacy & Security > Automation, find the host app (Terminal, Claude, etc.),
enable Mail. The doctor tool reports whether the grant is working.
The ad-hoc-signed Node problem
TCC identifies programs by code signature. A Homebrew-installed Node is
ad-hoc signed and its signature changes on every upgrade, so macOS forgets
the Automation grant and prompts again after each brew upgrade. If the
re-prompting bothers you, use the official Node.js installer (Developer ID
signed, stable identity) at a stable path. doctor reports which kind of
Node is running.
Tools
| Tool | What it does |
|---|---|
doctor |
Diagnostics: Mail running, Automation granted, accounts visible, Node signing, warm round-trip timing. Run first when anything misbehaves. |
list-accounts |
Accounts with enabled state, type, addresses. Disabled accounts are reported as disabled, never hidden. |
list-mailboxes |
Mailbox tree with full paths and (optionally) counts. Full paths are the identifiers everything else expects. |
get-unread-count |
Cheap unread counts; defaults to every enabled account's INBOX. |
list-messages |
Newest N of one mailbox: id, subject, sender, date, read state, mailbox path. |
search-messages |
Substring search over subject/sender with date range, across accounts or scoped. Reports every mailbox scanned and skipped. |
get-message |
One message in full: headers, recipients, plain-text body; includeHtml opt-in fetches and parses the source for the HTML part. |
get-thread |
Related messages via References/In-Reply-To, with visible normalized-subject fallback. |
list-attachments |
Attachment names, types, sizes, parsed from the MIME source. Metadata only. |
save-attachment |
Writes one attachment under the allowlisted root. Never overwrites. |
create-draft |
New draft with recipients, subject, body, optional attachments. Saved to Drafts, never sent. |
reply-draft |
Reply (or reply-all) draft with your body above the quoted original. Never sent. |
forward-draft |
Forward draft to given recipients with an optional note. Never sent. |
Things worth knowing before filing a bug
- Message ids are per-mailbox integers and do not survive moves. Every tool returns the full mailbox path with the id, and lookups need both. If a message was moved since you listed it, re-list or search again.
- Size guards are deliberate. Bulk reads degrade worse than linearly
with mailbox size (measured: a 52k-message mailbox is far more than 3x
slower than a 17k one). Mailboxes over the guard are refused or skipped
and reported, never silently dropped. Raise
maxMessagesper call to scan them anyway and expect tens of seconds. - Search covers subject and sender, not bodies. Fetching every body through the scripting bridge would take minutes and gigabytes.
- A failed operation never looks like an empty result. Transport failures (Mail unresponsive, timeout) surface as errors; domain refusals name the real candidates; partial scans label their totals as floors and list what was skipped.
- Ambiguity is refused, not guessed. Real mailboxes can have two
same-named siblings (this machine has twin
Junkmailboxes in one account); a path matching more than one mailbox is an error naming them. - Gmail accounts may hide standard mailboxes from the scripting bridge.
On this machine the Gmail-backed account exposes no
All Mail, no[Gmail]container, and noDraftsmailbox; its INBOX is real and holds the mail. Created drafts for such an account exist in Mail's unified Drafts even when no per-account Drafts path can list them.
Configuration
| Variable | Meaning | Default |
|---|---|---|
MAIL_MCP_SAVE_ROOT |
Root directory save-attachment may write under |
~/Downloads |
Testing
npm test # tier 1: 219 unit tests, no Mail.app, runs anywhere
node research/verify.mjs # tier 2: replays every recorded probe against real Mail
Tier 1 gates every commit: input guards, MIME parsing, path containment,
error mapping, and the no-network/no-send/no-deps enforcement tests. Tier 2
is the drift detector: it fails when a macOS update changes Mail's object
model or performance characteristics that this design rests on (bulk fetch,
byId cost, mailbox structure).
Tier 3 is drafts: creating them is the only mutation, so automated runs are
manual-by-choice. The live checklist used during development is described in
docs/apple-mail/README.md (draft recipes section); it creates clearly
labeled test drafts that you delete afterwards.
Architecture
src/mcp/ stdio JSON-RPC framing: initialize, tools/list, tools/call
src/tools/ one file per tool: validate input, call, shape output
src/mail/ domain logic that runs in Node: MIME parsing, path containment
src/jxa/ the runner plus one .js script per operation, run via osascript
Scripts are ordinary files, independently runnable from a terminal for
debugging (osascript -l JavaScript src/jxa/list-accounts.js). The runner
serializes all Mail calls (Mail's scripting dispatch is single-threaded),
charges queue wait against each call's own deadline, kills overruns with
SIGKILL, and treats non-JSON output as a transport failure, never as data.
The knowledge behind every design decision lives in docs/apple-mail/:
measured costs, the whose-clause trap, message identity semantics, the
draft-content trap, and the corrections to folklore that turned out to be
wrong on real hardware.
License
No license granted yet. All rights reserved.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。