Email in Claude Code, Codex and Cursor: connect an inbox over MCP

01What you need before you start

You need three things: a current version of the coding agent, an email service with an MCP server, and an account there that you can sign in to in the browser. Command Mail's server runs at https://commandmail.app/mcp and uses OAuth 2.1 with PKCE and dynamic client registration, so you do not create API keys or OAuth clients yourself.^[5]

Decide beforehand what the agent may do. On Command Mail's sign-in screen you switch each permission on or off: mail:read for reading, searching and organizing within each mailbox's autonomy level, which includes archive, spam, trash and rules (switching a retention period on or shortening it also needs mail:send and level auto); mail:draft for drafts; mail:send-approved for sending a draft a person approved; mail:send for sending and replying without individual approval and, together with mail:read, forwarding; and mailboxes for connecting addresses. Sending without approval starts switched off.^[5]

The free plan covers one mailbox, with further mailboxes paused but still receiving, and the last 90 days. It reads, searches, sorts and drafts, and it does not send, not even approved drafts. Sending needs Pro, which costs €19 per month for the first 100 paying Pro accounts, for as long as that subscription runs, and €29 per month after that; there are also yearly prices of €190 or €290, a 7-day Pro trial for €1 and a Team plan at €39 per person per month (prices on 2026-10-20).^[5]

02Connect email to Claude Code

Claude Code adds remote servers with one command and asks for sign-in inside the session.^[1]

1. Add the server: claude mcp add --transport http commandmail https://commandmail.app/mcp. Without further flags, the entry lands in local scope, meaning your user and the current project. Add --scope user to make it available in every project.
2. Start claude and type /mcp. Select commandmail and sign in; the browser opens Command Mail's consent screen.
3. Choose the permissions and confirm. Back in the terminal, /mcp shows the server with its tool count.
4. Check from the shell: claude mcp list shows ✔ Connected, ! Needs authentication or ✘ Failed to connect for each server.

If your team should get the same server, --scope project writes it to .mcp.json in the repository. Each person then approves the server once when running claude in that folder, and each person signs in with their own account. A cloned repository cannot approve its own servers.^[1]

Claude Code loads MCP tools through tool search by default, with exceptions such as older models or an ANTHROPIC_BASE_URL that points to a host other than Anthropic. Claude finds the mail tools when a task needs them, so a large tool list does not crowd every request.^[1]

03Connect email to Codex

Codex keeps MCP servers in config.toml, and the Codex CLI, the IDE extension and the ChatGPT desktop app share that file.^[2]

1. Add the server: codex mcp add commandmail --url https://commandmail.app/mcp.
2. Sign in: codex mcp login commandmail. On a machine without a browser, --no-browser prints the authorization URL and waits for you to paste back the callback URL from the browser where you signed in.
3. Choose the permissions on the consent screen and confirm.
4. Check: codex mcp list shows the configured servers; inside the codex TUI, /mcp shows the active ones.

The resulting entry in ~/.codex/config.toml is short:

[mcp_servers.commandmail]
url = "https://commandmail.app/mcp"

For a single repository, put the same table in .codex/config.toml in the project; Codex reads it only in trusted projects. Two optional keys are useful for mail: enabled_tools limits which tools Codex may call, and tool_timeout_sec (default 60 seconds) helps with slow searches across a large inbox.^[2]

Codex prefers the scopes a server advertises when it starts the OAuth login. Command Mail advertises all of its permissions, so the decision happens on Command Mail's consent screen, where you switch off what the agent should not have.^[2][5]

04Connect email to Cursor

Cursor reads MCP servers from mcp.json: ~/.cursor/mcp.json applies to every project, .cursor/mcp.json to one project. If the same server name appears in both files, the project file wins.^[3][4]

1. Open or create ~/.cursor/mcp.json and add:

{
  "mcpServers": {
    "commandmail": {
      "url": "https://commandmail.app/mcp"
    }
  }
}

2. Save the file and restart Cursor.
3. Cursor shows the server under Customize → MCPs. Follow the sign-in prompt and choose the permissions on the consent screen.
4. In an Agent chat, the tools list at the top of the chat panel shows the mail tools. Click a tool name to switch it on or off. Switching off send_message alone does not stop sending, because reply and forward_message send too; leave mail:send off on the consent screen instead.

Cursor uses dynamic client registration when the server supports it. The static CLIENT_ID fields in Cursor's documentation are only for providers without it, so you do not need them for Command Mail.^[3]

Cursor's documentation describes approval for MCP tools by default. In version 3.6 and later, the default Auto-review mode under Cursor Settings → Agents → Approvals & Execution lets allowlisted MCP tools run immediately and sends the rest through a safety classifier.^[4] Check that setting before you give a Cursor agent a sending permission.

05When the agent cannot use OAuth: the command line

Some setups cannot open a browser, for example a remote container or a CI job. Command Mail also has a command-line tool: curl -fsSL https://commandmail.app/install.sh | sh, then commandmail login.^[5] The coding agent can call the CLI directly, and commandmail mcp starts a local MCP server for agents that only speak stdio.

06Test the connection with a first task

The first call in every session should be capabilities. It returns the account owner, the granted permissions, each mailbox with its autonomy level, the allowed operations and the limits.^[5] Ask the agent: "Call capabilities on commandmail and tell me what you are allowed to do." If the answer matches what you chose on the consent screen, the connection works.

A good second task is read-only: "Search my inbox for invoices from the last 30 days and list sender, date and amount." That tests reading and search without changing anything. Only then try a draft, and approve it yourself before anything is sent.

Command Mail marks message content as untrusted data and flags suspected prompt injection. That matters more in a coding agent than in a chat app, because the same session may also run shell commands. Treat instructions inside an email as text to read, never as a task.^[5]

07Troubleshooting: missing tools and sending

Most problems fall into one of five patterns.

Claude Code shows "Needs authentication". The server is configured but nobody has signed in yet, or the token has expired. Open /mcp, select the server and sign in again. /mcp reconnect all retries every server that failed or needs authentication; in the interactive terminal that needs Claude Code v2.1.284 or later.^[1]

Codex lists the server but calls fail. Codex can connect to a server without credentials if no sign-in was completed.^[2] Run codex mcp login commandmail and start a new session.

Cursor shows no tools after editing the file. Cursor reads mcp.json at startup, so restart it. Check that the JSON is valid and that the key is url, not command.^[4]

A project server in Claude Code stays at "Pending approval". Servers from a repository's .mcp.json need your approval. Run claude in the folder and accept the workspace trust dialog and the server.^[1]

The agent says it cannot send. With Command Mail, check the plan first and the permission second. All sending needs Pro. With mail:send-approved, the agent sends only a draft revision that you approved in Command Mail, and not from a mailbox on level read. The direct tools send_message and reply exist only with mail:send; they also need a mailbox that can send, autonomy level auto and remaining room in the daily limit, and Command Mail refuses agent mail to automated senders, mailing lists and no-reply addresses.^[5]

To grant different permissions, sign in again: in Claude Code, choose Clear authentication for the server in /mcp and sign in; in Codex, run codex mcp login commandmail again; in Cursor, remove the server under Customize and add it again. A token refresh can only narrow the granted permissions, so widening them needs this new sign-in. The earlier grant stays valid on the server until it is revoked or expires; the consent screen links to your connections in Command Mail's settings.^[1][2][4][5]

08How the three agents compare for email work

The three tools differ less in what they can do with mail than in where configuration lives and who approves tool calls.

Sources: vendor documentation retrieved on 2026-10-20.^[1][2][3][4]

Whichever agent you choose, the mail-side permissions stay the same, because the email service enforces them on the server. Claude Code, Codex and Cursor each add their own approval layer on top.

09Next steps

If you work with several addresses, read using multiple email accounts with one AI agent. For the permission model in detail, see permissions and autonomy. If you are still choosing a server, the email MCP server comparison lists the options.

Start Command Mail free and connect your coding agent.

Frequently asked questions

Sources

1. Anthropic: Connect Claude Code to tools via MCP code.claude.com, accessed October 20, 2026
2. OpenAI: Codex MCP documentation developers.openai.com, accessed October 20, 2026
3. Cursor: Model Context Protocol (MCP) cursor.com, accessed October 20, 2026
4. Cursor Help: MCP cursor.com, accessed October 20, 2026
5. Command Mail: llms.txt with plans, scopes and connection options commandmail.app, accessed October 20, 2026