Framer MCP: Letting an AI Agent Edit Your Business Website

Framer now has an official MCP server. MCP, short for Model Context Protocol, is the standard way an AI tool connects to another app. With it, an AI tool you already use, such as Claude, Codex or Cursor, can open your Framer project and build pages, edit layouts, manage CMS collections, review SEO and publish. Framer updated its help articles on the feature on October 22, 2026.

For a small business with a Framer site, this changes who can make website edits and how fast. It also means an AI tool now holds a key to your public website. This post covers what the connection can reach, how to keep a bad edit off the live site, the workflows we would set up, and when it is not worth the effort.

We read Framer's MCP page, the help articles Getting started with Framer MCP and What you can do with Framer MCP, and Framer's agents page on October 23, 2026. We have not run Framer MCP on a client site. Everything below about how it works comes from Framer's own pages.

Framer MCP is a remote server at mcp.framer.com/mcp. You add it to an AI tool, sign in to Framer and approve access. After that, the agent works with the canvas, components, supported project settings and the CMS in every project your Framer account can reach, using your account's permissions. The Framer editor does not need to be open. By default, changes stay in the project and do not reach the live site until someone publishes. You can also ask the agent to work on a branch, so you review its changes before they are merged or published.

Framer says any AI tool that supports remote MCP servers over Streamable HTTP with OAuth sign-in can connect. It names Codex, Claude Code, Cursor, Pi, Windsurf, Warp and Grok, and it has one-click listings for ChatGPT, Claude and Cursor. For other tools you add a remote server named framer with the URL above, then sign in when Framer opens in your browser.

One point from Framer's FAQ is worth knowing before you budget. When you work through MCP, your AI usage follows the plan or credits of the agent you connect, not Framer's built-in agent. So the cost of the work sits on your Claude, ChatGPT or Cursor plan.

Framer groups the work into five areas.

1. 1Build and edit pages. The agent can create pages or change layouts across desktop, tablet and phone sizes, reusing the project's styles and components.
2. 2Manage content and the CMS. It can create collections and items, update structured content, and import content from sources the AI tool can reach.
3. 3Review SEO and site quality. It can inspect metadata, headings, alt text, page structure and content issues. If the AI tool is also connected to an SEO tool, Framer gives Ahrefs as the example, the agent can combine the two.
4. 4Use outside tools and data. Files, scripts, APIs and research available in the AI tool can feed the work.
5. 5Create a branch for review. The agent makes changes on a Framer branch, and you review them before merging or publishing.

There is one limit at launch. Localization is not supported through MCP. Framer's older local setup, the @framer/agent command line tool, still handles localization, and Framer says both setups stay supported until MCP catches up.

Framer is clear that an agent can change your live site if you ask it to publish. Nothing in the pages we read describes a setting that limits a connection to drafts only. The connection gets your account's permissions, across every project that account can reach.

That makes three choices matter more than any prompt.

1. 1Which Framer account you connect. If the account that owns billing and every client site is the one you sign in with, the agent can reach all of it. A team member account with access to one project is a smaller key.
2. 2Whether work happens on a branch. Branches give you a clean place to compare the agent's edits against the main version. Make "work on a branch" part of every request, so it still happens on busy days.
3. 3Who presses publish. Keep that a person for anything a customer sees: prices, offers, legal text, contact details and booking pages.

This is the same idea we cover in AI agent permissions and human approval: decide what the agent may do by how much damage a mistake would cause, and put a person in front of the steps you cannot undo.

These are the workflows we would build first. Each one keeps a person on the publish step.

1. 1Weekly SEO pass. The agent checks every page for missing meta descriptions, duplicate titles, empty alt text and broken heading order, fixes what it can on a branch, and lists what it changed. A person reviews the branch and publishes.
2. 2CMS updates from a source of truth. A team page, location list, menu or price list that lives in a spreadsheet gets synced into a Framer collection on request. The agent reports which items it added, changed or left alone.
3. 3Landing pages for a campaign. Marketing writes the offer and the agent builds the page from existing components, so it matches the brand. Copy and the form are checked before publishing.
4. 4Content fixes after a review. When someone spots an old phone number or a dead link, they describe the fix in chat and the agent finds every place it appears.

The connection is the easy part. The value comes from the rules around it.

1. 1A written brief the agent reads each time: brand voice, words to avoid, pages it must never edit, and the rule that it works on a branch.
2. 2A change summary at the end of every job, listing pages and CMS items touched, so the reviewer does not have to hunt.
3. 3A check after publishing that the live pages load, forms still submit and tracking tags still fire. If you run ads, a broken form costs money fast.
4. 4A plan for the spreadsheet or system that owns the content, so the agent copies from one source instead of guessing.
5. 5A short test on a copy of the project before you point the agent at the real one. Our post on testing AI agents before production covers what to try.

If your site has five pages and changes twice a year, the setup and review time will cost more than the edits it saves. Make the change by hand.

If the site is not in Framer, this does not help. Webflow has its own server, covered in our post on Webflow MCP with Gemini and Claude, and many other site builders have no agent connection at all.

If nobody on the team will review branches, do not connect it. An agent that edits and publishes with no one looking is a risk to the one page every customer sees.

And if the real need is moving leads from your site forms into your CRM, that is an automation job, not a website editing job. A plain workflow does it more cheaply and more reliably.

Connecting the MCP server is a few steps. Most of the work is deciding which account connects, writing the brief, setting up the review step, and checking the live site afterward. Those are the same pieces we put around any agent that touches a customer-facing system.

If you want help planning this, our MCP development and AI agents work covers connecting agents to the tools you already use, with permissions and review built in. Where the job is really moving data between systems, workflow automation is usually the better fit. Start with a free automation audit and we will tell you whether Framer MCP is worth setting up for your site, or whether a simpler fix covers it. OrchestriAI is not a Framer partner, and nothing here is paid placement.

This post is based on Framer's public pages as of October 23, 2026. Framer MCP is new, and the supported features, clients and limits may change. We have not tested it on a live site. Check Framer's help center before you connect a production project.