MCP email server for AI agents

@loftbox/mcp is the official Model Context Protocol server for LoftBox. It gives Claude Desktop, Cursor, or any MCP client a real email mailbox: 28 tools for sending, inbox reading, threads, domains, and — unusually for this category — human-in-the-loop approval built into the tool surface.

Install and run

LOFTBOX_API_KEY=lb_live_xxx npx -y @loftbox/mcp

For Claude Desktop, add to claude_desktop_config.json:

{
  "mcpServers": {
    "loftbox": {
      "command": "npx",
      "args": ["-y", "@loftbox/mcp"],
      "env": { "LOFTBOX_API_KEY": "lb_live_your_key_here" }
    }
  }
}

Cursor and other MCP clients register the same command (npx -y @loftbox/mcp) with the LOFTBOX_API_KEY environment variable. Transport is local stdio.

The 28 tools

Group	Tools
agents	`agent_create` `agent_get` `agent_list`
mailboxes	`mailbox_create` `mailbox_list` `inbox_list` `inbox_ack`
messages	`message_send` `message_get` `message_list` (full-text `q` search, label & status filters)
labels	`label_add` `label_remove`
human-in-the-loop	`approval_queue_list` `message_approve` `message_reject` `approval_policy_list` `approval_policy_create`
threads	`thread_list` `thread_messages`
domains	`domain_create` `domain_list` `domain_status` (DNS guidance)
suppressions	`suppression_list` `suppression_add` `suppression_remove`
attachments	`attachment_list` `attachment_url`
events	`event_list`

Every tool carries MCP annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) so clients can reason about side effects before calling.

Why approval tools matter

An agent that can message_send autonomously is a liability. With LoftBox, sends that match an approval policy queue as pending_approval instead of going out — and a human (or a supervising agent) resolves them with approval_queue_list → message_approve / message_reject. Governance lives in the same MCP tool surface the agent already uses.

Scope your key

Use a scoped API key for the agent and keep approve/reject rights on a separate key: the tools message_approve, message_reject, approval_policy_create, and suppression_add/suppression_remove deserve stricter scopes than day-to-day sending.

Get an API key — free Read the docs

Package: npm @loftbox/mcp · Source: GitHub · Also see the LangChain and CrewAI integrations, or how LoftBox compares to other agent email providers.