@zenderock/useinbox-mcp

Give your AI assistant its own email address. An MCP server for useinbox.email agent inboxes: the assistant can create mailboxes, read what arrives as conversations, and reply inside them.

Setup

Create an API key at https://www.useinbox.email/developers.html, then:

Claude Code

claude mcp add useinbox --env USEINBOX_API_KEY=neus_... -- npx -y @zenderock/useinbox-mcp

Any MCP client (claude_desktop_config.json, .cursor/mcp.json, …)

{
  "mcpServers": {
    "useinbox": {
      "command": "npx",
      "args": ["-y", "@zenderock/useinbox-mcp"],
      "env": { "USEINBOX_API_KEY": "neus_..." }
    }
  }
}

Set USEINBOX_BASE_URL as well if you run the API elsewhere.

Permissions

Create the key from Agent Mode → API keys and tick only what the assistant needs — typically inboxes:read, messages:read and messages:send. Leave inboxes:write off unless the assistant should be able to delete mailboxes. A key with nothing ticked has full account access.

Tools

| Tool | | |---|---| | list_inboxes | find an inbox id | | create_inbox | new address, live immediately | | delete_inbox | stops accepting mail at once | | list_threads | conversations with previews; filter by unread or label | | read_thread | full bodies and attachments, oldest first | | send_message | start a new conversation | | reply_to_thread | answer inside an existing one | | mark_thread_read | clear unread state | | set_thread_labels | replace a thread's labels | | get_inbox_activity | how much mail an inbox has handled, with a daily breakdown | | get_usage | plan consumption this month — check before a long run of sends | | check_delivery | delivery, bounce and complaint rates, and what failed |

What it lets an assistant do

"Check my support inbox and draft replies to anything unread."

The assistant lists unread threads, reads each one for context, and replies within the conversation — the recipient sees the answer threaded under their own message, not as a new email. Labels are how it records what it already handled, so the next session does not answer the same thread twice.

Errors come back as readable tool results rather than protocol failures, so the assistant can tell you "the account has reached its inbox limit" and act on it.