Mailoo - Email MCP Server

Mailoo

Mailoo is an open-source MCP server you run yourself to connect AI assistants and agents to multiple email accounts over IMAP and SMTP.

• Find what you need — search and read email, threads, and attachments across accounts.
• Act on email — prepare drafts, reply, and schedule messages through your assistant or agent.
• Choose your setup — guided setup for Claude Desktop or Cursor, or a standard MCP connection for your own agent.

Maintained by Bitfloo, forked from email-mcp. Licensed under LGPL-3.0-or-later; see NOTICE for attribution.

Quickstart for Claude Desktop or Cursor

Install Node.js 24 or newer and prepare your email account first, including an app password if your provider requires one. Then run these commands in a terminal:

npx -y @bitfloo/mailoo@latest account add

npx -y @bitfloo/mailoo@latest config edit npx -y @bitfloo/mailoo@latest install

In config edit, press Enter to keep the rate limit (10 for a new config), then choose Yes for Read-only mode? (disables all write operations). This saves read-only mode in TOML. Choose your client in the installer, restart it, and ask: “Summarize up to five unread emails. Include each sender, subject and any requested next step. Ask which account to use if needed.”

The wizard prints a read-only/core manual client entry; the installer does not apply those environment overrides. This guided path uses the read-only setting you saved in TOML. With untouched default settings, the installer's entry runs writable/full.

Follow the guided setup and check your first result →

Quickstart for agent builders

With Node.js 24 or newer and account credentials ready, create a local account using npx -y @bitfloo/mailoo@latest account add. Connect your MCP client over stdio:

{

"mcpServers": { "mailoo": { "command": "npx", "args": ["-y", "@bitfloo/mailoo@latest", "stdio"], "env": { "MCP_EMAIL_READ_ONLY": "true", "MCP_EMAIL_TOOLSET": "core" } } } }

Initialize the connection, list tools, then call list_accounts, list_emails, and get_email with the returned account name and message ID.

Follow the MCP request sequence and inspect its results →

Both quickstarts configure read-only access for the first read. The walkthrough also covers feature requirements and fallbacks, including AI triage, local calendars, ManageSieve, and experimental OAuth.

To enable writes later, run config edit, keep the rate limit, and choose No for read-only mode; for the builder entry, set MCP_EMAIL_READ_ONLY to "false". Restart the server after either change. Request sending or changes explicitly when you intend them. You can request a dryRun: true preview separately; Mailoo does not require a preview before sending.

Mailoo connects to the mail servers you configure. Your MCP client and its model provider receive the mail content you request. Optional TypeSafe classification and webhooks send data to their configured services when enabled. Usage reporting is off by default. See Capabilities & data flows for destinations and fields.

Highlights

Up to 57 tools across full, core, and minimal profiles, up to 7 prompts, and 6 resources. Includes calendar and reminders, scheduling, ManageSieve filters, and optional AI triage. See API for tools and docs/tools.md for behaviour and limits.

Table of Contents

• Highlights
• Quickstart for Claude Desktop or Cursor
• Quickstart for agent builders
• Security & Sandboxing Architecture
• LLM Ergonomics & Toolset Profiles
• Docs
• Background
• Install
• Usage
• Capabilities & data flows
• API
• Maintainers
• Contributing
• Attribution & License

Security & Sandboxing Architecture

Policy and how to report a vulnerability: [SECURITY.md](SECURITY.md).

Mailoo applies controls to specific network, filesystem, and tool operations. It also labels untrusted mail for the model. Those labels do not guarantee how an LLM will respond to instructions inside an email.

Enforced operation controls:

• Zero Shell Execution (`execFile` only) — Native OS integrations (macOS osascript for Calendar/Reminders/notifications, Linux notify-send, Windows powershell, and crontab installation) invoke child_process.execFile with argument arrays (shell: false), avoiding shell parsing of arguments.
• SSRF & DNS-Rebinding Protection (`net.BlockList`) — Outbound alert webhooks validate URLs against an IPv4/IPv6 net.BlockList covering loopback, RFC 1918 private ranges, link-local/cloud metadata (169.254.169.254, 100.100.100.200), CGNAT, NAT64, 6to4, and reserved prefixes both at configuration time and at socket connection time (createWebhookConnectLookup), preventing DNS-rebinding attacks (src/safety/validation.ts).
• Symlink-Free Write Containment (`savePath`) — Attachment and .eml exports via savePath are strictly confined to the working directory, walk every path segment with lstat, refuse hidden dotfiles and OS persistence directories (Library/LaunchAgents, .ssh, .gnupg), and open targets with O_CREAT | O_EXCL | O_NOFOLLOW under a 50 MiB cap (5 MiB inline base64 cap). An .eml export is capped by its own size; an attachment by its encoded size in the message, which for base64 runs above the file size (details).
• Strict Credential File Permissions (`0600`) — config.toml is written and verified with mode 0600 inside a 0700 config directory ($XDG_CONFIG_HOME/mailoo/config.toml). Passwords and message bodies are redacted from the JSONL write audit log (src/safety/audit.ts).
• TLS & STARTTLS Verification — tls enforces implicit TLS with certificate verification (verify_ssl = true by default); starttls fails closed if the server does not advertise STARTTLS. Turning both off is accepted only for a loopback host. ManageSieve refuses AUTHENTICATE PLAIN over unencrypted sockets without STARTTLS.
• Operational Guardrails (`MCP_EMAIL_READ_ONLY`, `MCP_EMAIL_TOOLSET`, `dryRun`) —
• MCP_EMAIL_READ_ONLY=true disables all write tools, background watchers, hooks, and schedulers at startup, exposing only the 31 read-only tools.
• MCP_EMAIL_TOOLSET=full|core|minimal scopes the registered tool surface to match the agent's capability tier and context budget.
• dryRun: true on high-risk tools (send_email, reply_email, forward_email, send_draft, delete_email, bulk_action, delete_mailbox, sieve_put_script, sieve_delete_script) validates all parameters and account resolution and returns a [DRY RUN] preview without modifying server state.
• Per-Account Rate Limiting — rate_limit (default 10 sends/minute) enforces an independent token bucket per account (src/safety/rate-limiter.ts). Buckets live in memory, so each server process and each mailoo scheduler check run counts on its own.

Untrusted-content labels: Email subjects, sender headers, bodies, previews, and authentication results are wrapped in << >> … << >> delimiters. Mailoo strips forged copies of these markers from the payload before wrapping it. This fencing identifies external content as data for the receiving model; it is not a sandbox for model behaviour or a guarantee against prompt injection.

LLM Ergonomics & Toolset Profiles

Mailoo is designed to work across both frontier models (Claude, GPT, Gemini) and local/small 3B–8B parameter models (llama.cpp GBNF grammars, Ollama, vLLM):

• Toolset Profiles (`MCP_EMAIL_TOOLSET` / `settings.toolset`):
• full *(default)* — All 57 tools (31 in read_only mode), including Calendar, Reminders, ManageSieve, Watcher, Scheduler, Templates, Contacts, and Analytics.
• core — 19 tools (10 in read_only mode) covering core email workflows (accounts, mailboxes, emails, threads, send, drafts, manage, bulk, attachments, health).
• minimal — 14 tools (8 in read_only mode) tailored for 3B–4B models with tight context budgets (accounts, mailboxes, emails, send, manage, health).
• Optional `account` & Smart Account Resolution — When only a single account is configured or default_account (MCP_EMAIL_DEFAULT_ACCOUNT) is set, the account parameter is completely optional across tools. Account lookup also matches case-insensitively by profile name or email address (e.g., passing "you@example.com" resolves the matching account), and returns actionable error messages listing valid accounts (name (email) pairs) when ambiguous.
• Lossless Parameter Normalization & GBNF Compatibility — Single-message tools accept emailId, id, or email_id as either string or integer (42 or "42"). Batch tools (get_emails, bulk_action) accept string or numeric ID arrays (["42", "43"] or [42, 43]). list_emails and search_emails accept both hasAttachment and has_attachment. All Zod schemas avoid regex lookaheads so llama.cpp GBNF grammar compilation succeeds without errors.
• Context Window Protection — get_email and get_thread default to format="text" and enforce a default maxLength of 20000 characters (appending a clear truncation notice when exceeded) so large HTML newsletters never overflow small context windows.
• MCP `outputSchema` & `structuredContent` — Tools declare structured outputSchema contracts and return machine-readable structuredContent alongside human-readable text blocks.
• Outgoing attachment paths must be regular files under the working directory or home — Security considerations

Docs

| IMAP credentials, Gmail / iCloud app passwords, and experimental Google