emailsRboring
A safe, redacting MCP server for Apple Mail on macOS that enables an LLM to read, search, triage, draft, and confirm sending email while stripping secrets and blocking dangerous actions.
README
emailsRboring-mcp
A safe, redacting MCP server for Apple Mail on macOS. It lets an LLM read, search, triage, draft, and (with confirmation) send email — while a single policy layer strips secrets, blocks dangerous actions, and refuses to send without your approval.
It's a thin proxy in front of two excellent existing servers, adding the safety layer neither has on its own:
- imdinu/apple-mail-mcp (Python) — fast FTS5 full-body search over your whole mailbox. Read/search.
- sweetrb/apple-mail-mcp (TypeScript) — a 40-tool AppleScript write surface. Draft/organize/send.
⚠️ Read SECURITY.md before using. This tool can read all your mail and send messages. Its protections are real but heuristic — the confirm-only send gate is not proof against prompt-injection, and it keeps a plaintext local cache of your email bodies.
Status: v0.1.0 — source on GitHub (not yet on npm), macOS only, requires the two upstream servers below. · Contributing or using a coding agent? See AGENTS.md. · Changes: CHANGELOG.md.
What it does
MCP client (Claude Desktop / Claude Code)
│ one server: "emailsRboring"
▼
┌──────────────── emailsRboring-mcp (policy proxy) ─────────────────┐
│ • fail-closed allowlist (send-serial / delete / rules are hidden) │
│ • redaction chokepoint: strips structuredContent, masks OTPs/codes │
│ • confirm-only send; reply/forward default to drafts │
│ • untrusted-content fence on reads; attachment-exfil block │
│ • 25k char-limit truncation; per-tool safety annotations │
└──────────┬───────────────────────────────────────┬────────────────┘
read/search │ │ write/organize/send
▼ ▼
imdinu (serve -r) sweetrb (build/index.js)
Tools (26, all mail_*):
- Read/search (8):
mail_list_accounts,mail_list_mailboxes,mail_get_emails,mail_get_email,mail_search,mail_get_email_links,mail_get_email_attachment,mail_get_attachment - Organize (10):
mail_move_message,mail_batch_move_messages,mail_flag_message,mail_unflag_message,mail_batch_flag_messages,mail_batch_unflag_messages,mail_mark_as_read,mail_mark_as_unread,mail_batch_mark_as_read,mail_batch_mark_as_unread - Mailboxes/info (5):
mail_create_mailbox,mail_rename_mailbox,mail_get_unread_count,mail_list_attachments, (+ reads above) - Compose (3):
mail_create_draft,mail_reply_to_message,mail_forward_message(draft by default) - Send (1, gated):
mail_send_email(requiresconfirm: true)
Not exposed (by design): mass/mail-merge send, delete/trash, and mail-rule editing.
Requirements
- macOS with Apple Mail configured and running
- Node.js ≥ 18
- Python ≥ 3.11 (for the imdinu upstream)
- Full Disk Access for the process that runs imdinu's indexer (it reads
~/Library/Mail) - Automation permission (Apple Events → Mail) — granted on first run via a macOS prompt
Install
1. The read/search upstream (imdinu)
pipx install apple-mail-mcp # or: uv tool install apple-mail-mcp
# Grant Full Disk Access to your terminal/host, then build the index:
apple-mail-mcp index # one-time; ~40s for ~40k messages
2. The write/organize upstream (sweetrb)
git clone https://github.com/sweetrb/apple-mail-mcp sweetrb-apple-mail
cd sweetrb-apple-mail && npm install && npm run build
# note the absolute path to build/index.js
3. This proxy
git clone https://github.com/ahmadzafar-code/emailsRboring.git
cd emailsRboring && npm install && npm run build
4. Point the proxy at the two upstreams
Set these env vars (the proxy reads them at startup):
| Env var | Value |
|---|---|
EMAILSRBORING_IMDINU_CMD |
path to imdinu's apple-mail-mcp (e.g. ~/.local/bin/apple-mail-mcp, or a venv path) |
EMAILSRBORING_SWEETRB_ENTRY |
absolute path to sweetrb's build/index.js |
5. Wire into a client
Claude Code:
claude mcp add emailsRboring -s user \
--env EMAILSRBORING_IMDINU_CMD=/path/to/apple-mail-mcp \
--env EMAILSRBORING_SWEETRB_ENTRY=/path/to/sweetrb/build/index.js \
-- node /path/to/emailsRboring-mcp/build/index.js
Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json (merge into any existing mcpServers):
{
"mcpServers": {
"emailsRboring": {
"command": "/ABSOLUTE/PATH/TO/node",
"args": ["/path/to/emailsRboring/build/index.js"],
"env": {
"EMAILSRBORING_IMDINU_CMD": "/path/to/apple-mail-mcp",
"EMAILSRBORING_SWEETRB_ENTRY": "/path/to/sweetrb/build/index.js"
}
}
}
}
⚠️ Use the absolute path to
node, not bare"node". Claude Desktop (a GUI app) launches with a minimalPATHthat usually doesn't include your Node install (e.g. nvm), so"command": "node"fails to start. Get the path withcommand -v node(e.g./Users/you/.nvm/versions/node/vX.Y.Z/bin/node). The same applies toEMAILSRBORING_IMDINU_CMD— use a full path, notapple-mail-mcp.
Then:
- Fully quit Claude Desktop (⌘Q) and reopen — closing the window isn't enough.
- Verify: Settings → Developer should list
emailsRboringas running (or show the error if it failed); the tools control in a chat shows themail_*tools. - On the first Mail action, approve the macOS "Claude wants to control Mail" (Automation) prompt.
- Troubleshooting: logs are at
~/Library/Logs/Claude/mcp-server-emailsRboring.log.
Claude Code uses the same idea via claude mcp add (above), which also accepts absolute paths.
Configuration
emailsRboring.config.json (next to build/) carries optional policy knobs (the tool allowlist and redaction rules are baked into code and cannot be loosened by config):
{
"sendAllowlist": [],
"fullBodyDefault": false
}
sendAllowlist— empty = off (send is confirm-only). If non-empty,mail_send_emailrecipients must match an entry (e.g."you@work.com"or"@work.com"). Recipients can never be derived from email content. Strongly recommended if you enable real sending — see SECURITY.md.
Usage examples
Three worked workflows (what you say → which tools run → what happens):
1. Morning triage
You: "What needs my attention in my inbox today?"
The agent calls mail_get_emails (filter: today), groups senders, and surfaces the few human/actionable items above the newsletter noise. Any verification codes in the listing come back [REDACTED]. Read-only — nothing changes.
2. Find a thread and draft a reply
You: "Find the contract-renewal thread with Dana and draft a reply saying I'll join the call Thursday."
The agent calls mail_search (query: "contract renewal", full-body FTS5), mail_get_email to read the latest message, then mail_reply_to_message with no confirm → a properly threaded draft lands in your Drafts. Nothing is sent; you review and hit send in Mail.
3. Bulk organize
You: "Archive everything from the K1 Speed newsletter and flag anything from my advisor."
The agent calls mail_search to collect ids, then mail_batch_move_messages (→ Archive) and mail_flag_message. All reversible; delete is not available by design.
Sending (gated)
You: "Email finance@acme.com the summary." → the agent shows you recipient/subject/body and calls
mail_send_emailonly after you say "send it" (it setsconfirm: true). Without your approval, the send is refused.
Verify your install
The proxy requires the two upstream paths in its environment (same ones your client config uses), so export them first:
export EMAILSRBORING_IMDINU_CMD=/path/to/apple-mail-mcp
export EMAILSRBORING_SWEETRB_ENTRY=/path/to/sweetrb/build/index.js
node build/index.js # boots: "[emailsRboring] ready — 26 tools (8 read / 18 write)" then waits on stdio (Ctrl-C)
python3 verify.py # 22 checks (surface, redaction w/ a live code, send-gate, exfil) → "22/22 ALL GREEN"
Without those env vars the server exits with a clear message telling you to set EMAILSRBORING_SWEETRB_ENTRY.
Credits & License
- Read/search upstream: imdinu/apple-mail-mcp — GPL-3.0. Installed separately; this proxy spawns it as a subprocess and does not include or modify its code.
- Write upstream: sweetrb/apple-mail-mcp — MIT.
- This proxy: MIT (see LICENSE).
Huge thanks to both upstream authors — emailsRboring is just the safety layer on top of their work.
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。