Connect your AI agent
One prompt to paste into your agent. It sets up the connection and you approve the access.
Open Connect your agent in the sidebar, or Developer tools, and copy the setup prompt. Paste it into your AI agent. With a terminal, it installs the CLI, the most capable route. A hosted assistant such as claude.ai or ChatGPT connects the MCP server instead. It walks you through signing in.
The prompt is a pointer, not the instructions. It tells the agent to read
/agent.md, which holds the
steps for each client:
- Claude Code
- Codex
- Cursor
- VS Code with GitHub Copilot
- a hosted assistant with custom connectors
- any other client that can run commands, or that supports remote MCP
What you do
You sign in, choose the workspace and approve the permissions, in your browser, on our page. The agent never sees a password, and the prompt contains no key or token.
Then the agent reads your account to prove the connection works. That is a read-only call: setup creates, changes and spends nothing.
If the agent cannot configure itself
A client with no terminal or configuration access cannot install a connector from a prompt alone. It will give you the endpoint to add by hand:
https://mcp.atelier201.com/mcpAdd it as a custom remote MCP connector in that client, then choose Connect and complete the sign-in. The MCP overview has the generic JSON entry.
Start before signup
The website’s Start with your AI agent button copies a short prompt. An HTTP-capable agent can prepare a site without an account, build it, and give you a demo link and a claim link. You sign up or sign in, review the site, and choose whether to connect the agent. The claim link works for 7 days.
Declining access still keeps the site. Approving access lets the agent receive its token through OAuth. Nothing is published and no domain is connected during the claim: publishing stays a separate instruction after you have seen the preview. See Guest sites.