Framer MCP: Your AI Agent Can Publish Your Site. Should It?

Framer MCP is Framer's official Model Context Protocol server, released on October 22, 2026. It lets an AI agent such as Claude, Claude Code, Cursor, Codex or ChatGPT work inside your Framer projects: inspect a project, edit the canvas and components, update CMS content and publish the site. It is a remote, hosted server, so there is no local bridge to install and the Framer editor does not need to stay open. You approve the connection with your Framer account, the agent works with your account's permissions, and it only publishes when you explicitly ask. My short take as a design engineer: it is a real shift for marketing sites, and the most important feature is not that the agent can publish. It is that it does not have to.

What Framer MCP is, and what shipped on October 22

MCP is an open protocol that lets AI tools talk to other software in a structured way. The Model Context Protocol documentation explains it well if you have not met it yet. In practice, an MCP server is a door: it decides what an agent can see and what it can change.

Framer's door opens onto four things, according to the official Framer MCP page: the canvas, components, supported project settings and the CMS, in any project your account can access. The Framer MCP update post sums it up as an agent that can work on the canvas, manage CMS content and publish your site, from desktop to phone.

Two details matter more than the headline. First, work can happen on a branch, which you compare with main before anything goes live. Second, Framer says agents publish only when you explicitly ask them to. Otherwise the changes stay in the project.

There was already a way to drive Framer with agents: the older External Agents setup, which runs the local @framer/agent CLI. Framer says both will stay supported until the MCP server reaches feature parity, and localization still needs the CLI for now.

How to connect Framer MCP to Claude, Cursor or ChatGPT

The setup is short, which is part of why I think this will spread fast. From Framer's page, the flow is:

Use the one-click "Add to" link for your client, or follow that client's install guide.
For an app that is not listed, add the server manually with the URL https://mcp.framer.com/mcp.
Approve the connection with your Framer account (OAuth).
Ask your agent to work on a specific project.

The named clients include ChatGPT, Codex, Claude, Claude Code, Cursor, VS Code, Windsurf, Warp, Notion, Devin and several others. Framer's FAQ adds that any client that supports remote MCP servers and OAuth should work. So "is there a Framer MCP for Claude" has a simple answer: yes, in both Claude and Claude Code.

One tip from how I already work with design MCPs: name the project, the page and the branch in your first prompt. Agents are good at doing what you ask and bad at guessing which of your twelve projects you meant.

Framer MCP vs Figma MCP vs the Framer agent CLI

I build most of my AI design experiments with Figma, Claude and the Figma MCP, so the obvious question for me was where Framer's server fits. Here is how I see the three side by side. The Figma column describes how I use it, design to code, not every feature Figma offers.

	Framer MCP	Framer External Agents (CLI)	Figma MCP (my use)
Where it runs	Remote, hosted by Framer	Local CLI on your machine	Connected to the agent I code with
What the agent touches	Canvas, components, supported settings, CMS	Project workflows, including localization	Design context read from Figma files
Output	A live Framer site	A live Framer site	My own code: SwiftUI, WebGL, three.js
Publishing	Only when you ask; branches to review	Through Framer	I ship the code myself
Best for	Marketing sites and CMS-heavy pages	Localization until parity	Custom interactions, native apps

The difference is the destination. The Figma MCP feeds design into code that I own, which is why I could build things like a bank card you shatter to unfreeze in SwiftUI. Framer MCP keeps everything inside Framer. You get speed and hosting, and you give up the deep custom stuff. For most company websites, that is a good trade.

The real feature is the review step, not the publish button

When I led design at ViewStats, the creator platform built by the team behind MrBeast, we took AI thumbnail and title tools from concept to launch in roughly two weeks. The lesson I wrote down in the ViewStats case study was that the hardest part was not the interface. It was designing trust.

Framer seems to have understood the same thing. Look at what their page stresses: you approve the connection, the agent uses your permissions and no more, changes do not reach the live site until someone publishes, and branches let you fix an agent's work before it ships. Those are trust features. They are the reason a team would let an agent near a production website at all.

So my rule is simple. The agent works on a branch. A human compares it with main and presses publish. I would not give that last step away on a company site, even if the tool allows it, because a website is a public promise. A wrong price, a broken form or a CMS entry with placeholder copy is a trust problem, not a typo.

That is the same principle I use for onboarding: show value first, ask for trust second. Let the agent prove itself on drafts and branches before it earns the right to touch anything live.

What Framer MCP changes for product designers and design engineers

At ViewStats my role moved from Lead Product Designer to design engineer. I worked in the codebase with Cursor, Claude Code, React, TypeScript and GitHub, built components and opened pull requests. The iteration loop got dramatically shorter, because the person who designed the screen could also ship it.

Framer MCP brings a version of that loop to people who never wanted to open a code editor. A marketer can ask an agent to add ten CMS entries. A founder can ask for a new landing page variant from their phone. A designer can describe a section and review what comes back. That is real leverage.

It also moves the designer's job up a level, which I wrote about in what AI replaced in my product design work. Production gets cheaper. Judgment does not. Someone still has to decide what the page is for, which components are allowed, what good copy sounds like and when the agent's output is wrong. If your Framer project has no clear component structure, the agent will happily make it messier, faster.

So the practical advice is boring and important: clean up your components, name things clearly, keep CMS fields tidy. Agents work best in well-structured projects, exactly like new teammates do. If you want the bigger picture of where this leaves the role, my guide on what a product designer does covers the parts AI does not touch.

Where I would use it, and where I would not

I would use Framer MCP for CMS updates, new landing page variants built from existing components, content edits across many pages and first drafts of sections that a designer then refines. Those are repetitive, well-bounded jobs where an agent saves hours.

I would not use it, yet, for brand-defining hero sections, custom interactions with physics or shaders, or anything involving localization, which Framer says still needs the older CLI. I would also be careful on shared team accounts, because the agent inherits whatever that account can do.

A limit of this post: it is a first look based on Framer's official pages, a few days after release. Framer does not list individual tools or plan requirements on the MCP page, so details may change quickly.

Framer MCP FAQ

Is Framer MCP free?

Framer's MCP page does not state plan requirements or a price for the server itself. It says AI usage is billed according to the plan or credits of the agent you use, so your Claude, ChatGPT or Cursor subscription pays for the thinking. Check Framer's pricing page for your project's plan.

Is there a Framer MCP for Claude and Claude Code?

Yes. Claude and Claude Code are both named as supported clients. Add the server in Claude or Claude Code, approve the connection with your Framer account, then ask the agent to work on a project.

How do I connect Framer MCP if my tool is not listed?

Use the manual setup with the server URL https://mcp.framer.com/mcp. Framer says any client that supports remote MCP servers and OAuth should work.

Will Framer MCP replace web designers?

It replaces a lot of production work: building pages from existing components, filling CMS entries, small edits. It does not replace deciding what the site should say, designing the component system the agent works with, or reviewing the result before it goes live. Designers who set that structure become more valuable, not less.

A quick checklist before you connect an agent to your site

Clean component structure and clear names, so the agent reuses instead of inventing.
A branch for every agent session, compared with main before publish.
A human owns the publish button on production sites.
A dedicated account or clear permissions, since the agent inherits them.
Localization still handled through the CLI until Framer reaches parity.

If you want help setting up a design system and workflow that agents can work in safely, or you need someone who designs and ships, work with me as your product designer. You can also browse my product design work to see how I approach it.