Until now, pointing Claude Code or Cursor at a Framer project meant setting up Framer’s agent CLI and a local bridge, or reaching for an unofficial plugin that had to stay open in the editor the whole time. Framer’s new official MCP server, announced October 22, drops all of that. You approve the connection with your Framer account once, and your AI tool can work on your projects from wherever it runs, including your phone.

What changed

Framer MCP is a remote server that any AI tool supporting remote MCP connections and OAuth can talk to. Framer lists one-click setup for ChatGPT, Codex, Claude, Claude Code, Cursor, and Grok Bot, plus install guides for a longer list including VS Code, Windsurf, Notion, and Warp. Anything else can connect manually with the server address on Framer’s MCP page.

Once connected, the agent can work with the canvas, components, supported project settings, and CMS in any project your account can reach. It can build pages, update content, and publish. Framer is specific about a few things:

  • Your project doesn’t need to be open, and there’s no local bridge to install.
  • It works on the project you point it to, but it can reach any project your account can. To point it at one, copy the project link from Invite in Framer and tell the agent to connect to it.
  • Nothing goes live until you or the agent publishes. The agent only publishes when you explicitly ask it to.

The older @framer/agent CLI setup still works, and Framer says both stay supported while MCP catches up to the CLI’s features. MCP is now the recommended route for new setups.

Why it matters for designers

The interesting part isn’t another way to prompt a website. It’s that the agent you already use for everything else can now reach your Framer site alongside its other connections. Framer’s own examples lean into that: preparing local case studies and team photos for the CMS, pulling design references from Figma or a client brief from Google Drive, and running scheduled jobs like a weekly keyword check or a pre-meeting analytics report.

If you keep a portfolio or run client sites in Framer, that’s the kind of upkeep that usually slides. A new project write-up sitting in a folder can become a CMS entry without you rebuilding it field by field.

It also pairs naturally with Framer’s branching. Framer’s MCP page suggests asking the agent to work on a branch. That’s a request you make, not an automatic step, so say it every time. Then you can open the branch, click Review Changes, check the diff, and apply it to main. Applying doesn’t publish, either. Your live site only changes when you publish main, and an applied branch can be reverted from version history.

What to watch out for

The agent can delete things. Framer’s page for its older agent setup says plainly that an agent can read and write CMS collections, including deleting items, and the MCP page gives the new server access to your CMS too, so assume the same. Ask it to work on a branch, confirm it did, and actually read the diff, especially on a client site with a full CMS.

Be careful what you say yes to. Publishing is gated behind your request, but a loose “go ahead and finish this” could count. If you want a human look before anything ships, say so in your instructions every time, or save it as a standing rule in your AI tool.

Watch for confusingly similar names. There are unofficial Framer MCP plugins and community servers, including one in the Framer Marketplace that its own listing says isn’t official. The official one lives at Framer’s MCP page and doesn’t need a plugin open in the editor.

Costs land on the other side of the connection. Framer says usage follows your AI tool’s plan or credits, so a long agent session spends from that budget, not Framer’s. The sidebar covers what we could and couldn’t confirm, including whether any Framer plan is required.