Mailloop for AI Agents

Mailloop now has an MCP server and an Agent Skill. Your AI coding agent can create a sandbox, trigger your app to send, wait for the email to land, and check the subject and links, all from the chat. It works today with Claude Code, Cursor, VS Code, Windsurf, Zed, and over OAuth with claude.ai, ChatGPT, and Claude Desktop.

The old loop

You finish the password reset code. Now you want to know it works.

You run the app and submit the form. You alt-tab to a mail client or a testing inbox. You wait for the message, refresh a couple of times, and find it. You open it, read the subject, and check whether the reset link points where you expect. The link looks fine, so you click it and follow the flow by hand. Something is off in the HTML, so you switch back to the editor, change a line, and run the whole sequence again.

Every pass costs a context switch out of the editor, a manual wait at the inbox, and an eyeball check that you have to repeat for text, HTML, and each link. The agent that wrote the code could not see any of it, so the proof step was always yours.

The new loop

You stay in the chat. You say:

The agent runs the sequence itself. It creates a sandbox with create_sandbox, points the signup at the capture address, and triggers the send. Then it calls wait_for_email, which blocks until a matching message arrives instead of polling a UI. When the email lands, the agent calls get_email and reads the sanitized content with every link already extracted, so it can check the subject, follow the verification link by its returned url, and tell you what passed.

For a one-off run, the agent can use create_temporary_sandbox instead. That makes a throwaway inbox that expires on its own, which is handy in CI where you do not want to clean up after yourself.

Get started

For Claude Code, install the plugin. It adds both the MCP server and the skill. These are two slash commands, and you run them one at a time. Run this first:

After that, run /mcp, pick mailloop, choose Authenticate, and approve in the browser. Claude Code uses OAuth by default, so there is no API key to manage.

If you want the server without the skill:

To use an API key instead of OAuth, add a Bearer header. Create the key under Settings, then Organization, then API keys:

OAuth works with claude.ai, ChatGPT, and Claude Desktop too. Cursor, VS Code, Windsurf, and Zed connect to the same endpoint with an API key.

Full setup is in the MCP docs, the playbook is in the Agent Skill docs.