Connect Claude Code

MailMoose ships an MCP server, so you can work with your mail by talking to Claude Code or Claude Desktop instead of opening the app.

You stay in charge. Claude reads and drafts, and your client shows you every action before it runs.

Connect Claude Code

Run this in your terminal:

claude mcp add --transport http mailmoose https://mailmoose.dev/api/mcp

Then type /mcp in Claude Code and choose to sign in. MailMoose opens in your browser, you approve what the connection asks for, and Claude Code receives a token of its own. No key is copied anywhere, and you can withdraw the token later without touching anything else.

The connection belongs to the organization you had selected when you approved it. The same command sits behind Connect agent, next to the organization name in MailMoose, so you never have to type the address by hand.

Connect a different agent

Cursor, Codex, VS Code and the rest each want the server written into their own configuration file, in their own shape. Rather than looking yours up, paste this into the agent itself:

set up https://mailmoose.dev/SKILL.md

This is not a terminal command. It goes in the chat, and the agent fetches the page and writes the configuration for whichever client it is running in. You still sign in yourself afterwards, in a browser, exactly as above.

Two things it cannot do. A hosted assistant like ChatGPT has no access to its own settings, so it will hand you the steps instead of taking them, see Connect ChatGPT. And it needs to be able to fetch a URL and edit a file, so an agent you have kept away from both will say so.

Read what it proposes before you approve it, the same as with any other change an agent offers to make.

Connect with an API key instead

Use a key when the client cannot open a browser, or when you want the permissions fixed in advance and revocable on their own. This runs a small program on your machine, so it needs Node 20 or later and a key from Settings, see API keys.

Start with a read only key. Add sending once you are comfortable with what it does.

{
  "mcpServers": {
    "mailmoose": {
      "command": "npx",
      "args": ["-y", "@mailmoose/mcp"],
      "env": {
        "MAILMOOSE_API_KEY": "mm_live_your_key_here"
      }
    }
  }
}

Restart Claude Code. Ask it "what is unread in my inbox" to check the connection.

Claude Desktop uses the same block, in its own configuration file.

What you can ask for

1. "What came in overnight across all my domains?"
2. "Summarise the thread with Jane about the January quote."
3. "Reply to that and say Thursday at 14:00 works."
4. "Has anyone from that supplier emailed us before?"
5. "Create hello@learnfarsi.app."
6. "Connect learnfarsi.app and tell me what to add at my registrar."
7. "I've added those records, check them."
8. "Turn on sending for learnfarsi.app."

Reading and searching happen without fuss. Anything that sends shows you the message first.

What it will not do

The server is built so that a message arriving in your inbox cannot turn into an action against you.

1. Replies cannot be redirected. Replying takes a conversation and a body, never a recipient. The address comes from the conversation's own records, so nothing written in an email can change where your reply goes.
2. New recipients need your word. Sending to an address you have never corresponded with is refused until you confirm it yourself.
3. Message content is marked as untrusted. Text from senders is labelled so Claude treats it as information rather than instruction.
4. There is no tool for creating keys, changing webhooks, or deleting a domain. Those stay in the app, where a person does them deliberately.

A word on trust

Anyone can send you an email, so anything inside a message is text written by a stranger. A message might contain something that reads like an instruction, for example asking for a reply to go to a different address.

The measures above remove the ways that could turn into a real action. They do not make the problem disappear, and no product has solved it. Read what Claude shows you before approving anything that sends, and be especially careful with a message carrying a failed authentication badge.

Common issues

1. Claude cannot see the tools. Restart the client after editing the configuration. On a signed in connection, run /mcp and check it says connected.
2. Every call says the key is invalid. The key may be revoked or expired. Create a new one under Settings, API keys.
3. Reading works but sending is refused. Your key lacks the send permission. Create a new key with it.
4. Sending stops partway through the day. You have reached the key's daily send limit. The default is 100 sends per key and 1,000 per organization over the previous 24 hours. Wait for older sends to leave that window.