Webflow MCP Server: Connect Webflow to Claude, ChatGPT & Cursor

The Webflow MCP server is Webflow’s official Model Context Protocol server. It exposes Webflow’s Designer and Data APIs as a toolset that AI assistants like Claude, ChatGPT and Cursor can call directly, so “build a landing page from this brief using our existing components” or “which pages lost traffic this week?” becomes a conversation rather than a ticket for the web team.

If you have read our guides to the GA4 MCP server and the PostHog MCP server, the idea will be familiar. The difference is that those servers mostly read data. The Webflow server writes to your live website. That makes it more useful and more dangerous, which is why most of what changed in MCP 2.0 is about guardrails.

“Most agentic work happens in a sandbox, or in internal tools. When an agent edits your website, it’s live the second it publishes to customers, competitors, and every model crawling the web.”

“When the stakes are this high, there is no room for guesswork. MCP 2.0 gives agents the brand rules and governance that production work demands. This is the line between agentic experiments and agentic infrastructure.”

Linda Tong, CEO, Webflow, MCP 2.0 press release, 21 July 2026

Why this matters for marketing teams

Most marketing teams already draft copy, mock up pages and take a first pass at code in Claude or ChatGPT. That first step is easy. The wall comes at step two: getting the draft onto the live site. Someone copies it over by hand, or it sits in a ticket waiting for a developer, or it goes live and turns out not to match brand guidelines.

That gap is what Webflow is targeting. In The agentic web is here, Emma Genesen, Senior Product Marketing Manager at Webflow, frames the trade-off most tools force on you:

“Unfortunately, most tools make you pick between speed, control, and governance. Over time, this results in a build up of design drift, unmaintainable code, security vulnerabilities, and tool sprawl.”

Emma Genesen, Senior Product Marketing Manager, Webflow

Webflow’s numbers suggest teams are already past the “should we” stage. According to the company’s own press release (so treat them as vendor figures), over 30% of its enterprise customers actively use the MCP, usage grew more than 4x between January and July 2026, and nearly 90% of MCP users connect through Claude.

See it in practice

The best introduction is Webflow’s own webinar, “Building AI workflows with Webflow MCP 2.0”, recorded in August 2026. Casey Katalbus, a solutions architect at Webflow, hosts three agency partners who walk through the workflows they use on real client projects:

• Steve, VZA Digital built a Claude agent that takes a Figma file and a Webflow project URL, catalogues every section as a component, imports variables for colour and typography, converts pixels to rem, then builds the page. The page went live in a day, after roughly 20 prompts and about 90 minutes of agent time.
• Marcel, South runs a deliberately restricted skill on top of his Small Medium Large framework. Claude gets a screenshot, reads a component registry stored in Webflow’s Agent Instructions, shows a plain-text plan for approval, then assembles the section using only existing components. No new classes, no new styles, no custom code.
• Matt Ward, Craft and Crew built an entire R&D site end to end with AI: brief, strategy, brand identity and mockups in Claude Design, then variables, three CMS collections, atomic components, modules and a 50-item content import through the MCP. Between eight and 12 hours from idea to done.

The Q&A at the end is worth your time if you are nervous about letting an agent near production. None of the three publishes through MCP without review, all of them test on a duplicate site first, and they ask the agent for a step-by-step plan before it executes anything complex.

What changed in MCP 2.0

Webflow launched the first version of its MCP server in 2025 with solid support for site metadata and CMS content. MCP 2.0, released on 21 July 2026, goes much further. The headline changes:

Your design system is agent-ready

Agents can now build with your existing Webflow styles, elements and components, including props, slots and variants, and bind elements to CMS data. Webflow’s own example: give Claude a folder of customer case studies and it can build a CMS Collection page using your design system, then generate and publish each page.

The bridge app is (mostly) gone

MCP 1.0 needed a companion app running inside the Designer for almost anything visual. In 2.0 most work runs through the Data API, so agents can build across multiple pages or sites at the same time with the Designer closed. The Bridge App is only needed for a handful of live-session capabilities (more on that below).

Agent Instructions

Site-specific rules and skills, written in markdown and stored on the site itself. Point them at your design system, CMS collections, legal docs or tone of voice guidelines, and every agent connecting through MCP receives them automatically. Since Webflow Conf in September 2026, Webflow can generate them for you from what it already knows about your site.

Webflow Analyze data over MCP

Ask about traffic trends, top pages, engagement events, time on page and traffic sources without leaving the chat, and pass the results to other MCP connectors like Slack or Asana in the same conversation.

Governance

Agents work in page branches (an Enterprise feature) before changes touch the live site, Webflow’s roles and permissions are fully enforced (including custom roles and per-site, per-collection, per-page and per-locale access), and every MCP action lands in the site activity log with AI attribution.

MCP 2.1, announced at Webflow Conf, adds GSAP-powered Interactions, better CMS querying (sort and filter on custom fields, search across reference fields, paginate results) and Webflow Cloud deploy logs surfaced to agents. Webflow said it would roll out to everyone in September 2026.

What can you do with it?

The server exposes dozens of tools grouped into families. Here is what each family means for a marketer, based on the live server’s own tool guide (version 2.0.1 at time of writing):

There is also a Webflow Guide tool, exposed as an MCP resource, that explains to the agent how to use everything else. In Claude and Cursor you can @-reference it in a prompt to load it into context before the agent starts.

Webflow MCP server vs the Data API vs the Designer

Not sure which route fits how you work? Here is how the three compare:

The MCP server earns its keep on the jobs that are tedious in the Designer and not worth writing code for: the 200-item CMS update, the site-wide metadata audit, the fifth landing page variant this month. For the other servers worth knowing about, see our MCP servers hub.

Real prompts you can try

Webflow’s own prompt library is a good starting point. Some practical prompts for marketers:

• “List all my collections and show me their field structures”
• “Audit my site for broken links, missing alt text and incomplete meta descriptions, then fix what you can”
• “Which pages have missing or weak SEO metadata? Suggest improvements and apply them with my approval”
• “Build a responsive hero section with a headline, description and CTA button using our existing components”
• “Take these ten customer case studies and create a CMS item for each, with a featured image and meta description”
• “Update every blog post that mentions our old product name to use the new one”
• “How did the site perform this week? Flag pages with dropping traffic and summarise the top traffic sources”
• “Find all hardcoded colours and convert them to design variables”
• “Create a branch for the pricing page, update the plan names, and show me the diff before I merge”

Setting it up

The server runs remotely at https://mcp.webflow.com/mcp and authenticates with OAuth, so there are no API keys to store and nothing to install locally for the main clients. Each authorisation grants access to one Workspace, and you pick which sites within it the agent can touch.

Claude Desktop

1. In the Claude chat window, click the plus icon and choose Add connectors.
2. Search for Webflow (switch from Featured to All if it does not appear) and click Connect.
3. Sign in to Webflow when prompted.
4. Select the sites or Workspaces you want Claude to access and click Authorize App.

If you prefer a custom connector, use the URL https://mcp.webflow.com/mcp (or https://mcp.webflow.com/beta/mcp for the beta server).

Claude Code

claude mcp add --transport http webflow https://mcp.webflow.com/mcp

Start Claude Code, run /mcp, select Webflow and authenticate in the browser window that opens. If you manage several sites, add the server with -s project inside a dedicated folder for each site so each project keeps its own authorisation. Our Claude Code guide for marketers covers the basics if this is new to you.

Cursor

Install the Webflow plugin from the Cursor Marketplace, or add this to .cursor/mcp.json:

{
  "mcpServers": {
    "webflow": {
      "url": "https://mcp.webflow.com/mcp"
    }
  }
}

Then go to Settings, Cursor Settings, MCP & Integrations and click Connect next to Webflow to authorise. Webflow recommends a per-project config rather than the global one to avoid repeated auth prompts.

ChatGPT

Open the Webflow app in ChatGPT, click Connect, sign in with Webflow and pick your sites. Then mention @Webflow in a conversation. By default ChatGPT asks you to approve every action before it takes it in Webflow.

Older configs

Earlier documentation and the GitHub README use https://mcp.webflow.com/sse with the mcp-remote package. That still works, but the /mcp endpoint is the current one. There is also a local install via npx with a static API token if you need to run the server on your own machine.

Add the Webflow skills

The MCP server gives the agent tools. Webflow’s skills repo gives it workflows that chain those tools together: bulk CMS updates with validation and preview, a site audit with a 0 to 100 score, a link checker, an accessibility audit, image compression, a Figma-to-Webflow build and a safe-publish flow that previews and confirms before anything goes live.

Install them in Claude Code with:

claude plugin marketplace add webflow/webflow-skills
claude plugin install webflow-skills@webflow-skills

Or run the interactive installer for any agent that supports the Agent Skills standard:

npx skills add webflow/webflow-skills

If the idea of skills is new to you, we cover agent skills for marketers and when to use MCPs, CLIs and skills separately.

Guardrails before you touch production

Everything the agency panel said in the webinar comes down to one point: the tool is capable, so the process has to be careful. Their advice, plus Webflow’s own:

• Duplicate the site first. Test on a copy or a staging site until the agent and its instructions behave predictably. Steve at VZA Digital ran about 15 tests before trusting his agent on a client build.
• Write Agent Instructions early. Component naming, class conventions, brand voice, legal language. Marcel’s whole approach depends on the agent reading a component registry from Agent Instructions and using nothing else.
• Ask for a plan, then approve it. For anything with several steps, have the agent explain step by step what it will do before it executes. Matt Ward works one piece at a time so a mistake in step two does not cascade through step seven.
• Do not publish through MCP without review. All three panellists build with MCP and publish by hand. Webflow’s safe-publish skill exists for exactly this reason.
• Use branches if you have them. Page branching is an Enterprise feature. If you are on Enterprise, have the agent work on a branch and merge after review.
• Check the activity log. Webflow logs every MCP action with who triggered it and what changed. You can query the log through MCP too.
• Limit authorised sites. Grant access per project rather than to the whole Workspace. Webflow also recommends it for performance.

Matt’s framing is the one to remember. Treat the agent like a junior developer: you operate at the strategic level and hand it the small, well-defined tasks it is good at.

Limitations to know about

• No Interactions yet. The server cannot create or apply Webflow Interactions (IX3). MCP 2.1 adds GSAP-powered interactions, but check the changelog for your rollout.
• Localisation is secondary locales only. You can read and update existing content in secondary locales, but not create new localised CMS items, and the primary locale is read-only through the localisation tool.
• Analyze needs the add-on. The analytics tools return nothing unless your site has Webflow Analyze, an optional paid add-on on paid Site plans.
• One Workspace per authorisation. Switching Workspaces means re-authenticating. For connector-based clients like Claude and ChatGPT that means removing and reinstalling the connector.
• The Bridge App is still needed for live Designer work. Visual snapshots of elements and reading or changing the current selection, page, mode, branch or breakpoint require the Designer open with the Bridge App connected. Everything else works without it.
• Rate limits are real. Marcel hit the Data API rate limit during his live demo. Big bulk jobs may need to be batched.
• It cannot change access. The server cannot add users or assign roles. It works inside your existing permissions.

What does it cost?

Nothing extra. MCP 2.0 is available on all Webflow site plans at no additional cost, and feature access follows your existing plan allowances, roles and permissions. The two exceptions are Webflow Analyze (a paid add-on) for the analytics tools and an Enterprise plan for page branching.

The bigger picture

Webflow’s roadmap, in its own words, “is pointed at full parity: everything you can do in Webflow, your agents will be able to do too.” Webflow Conf added Agent Presence (agents visible on the canvas like any other collaborator), site-wide Releases and AEO content agents, with more MCP access to AEO data promised later in the year.

That fits what we are seeing across the stack. Figma’s MCP handles the design side, Webflow’s handles the build, and analytics servers from GA4 and PostHog close the loop on results. Alex Mejias, Head of Global Design Technology and Interactive Experience at Amazon Brand Innovation Lab, put it this way in Webflow’s launch post:

“Our workflows with Figma, Claude Code, and Webflow’s MCP have been an absolute game changer. I really see this as our future workflow.”

Alex Mejias, Head of Global Design Technology and Interactive Experience, Amazon Brand Innovation Lab

Whether MCP itself is the protocol that wins is a fair question. An official server from Webflow, with nearly 90% of its users arriving through Claude, says that for now it is.

We are on-track to deliver a 43% increase in inbound leads this year. There is no doubt the adoption of Growth Method is the primary driver behind these results.

Laura Perrott, Colt Technology Services

About Growth Method

If an agent building your landing page from a brief feels like where marketing is heading, that is the bet we are making at Growth Method too. We are the agentic marketing platform for B2B teams. Connect Webflow once, alongside PostHog, GA4, Search Console and the rest of your stack, and both your team and your AI agents can plan campaigns, launch them and see which ones worked.

The Webflow MCP server is how an agent changes your site. Growth Method is where you decide which campaign to run next, and track it from plan to result with the plan, launch, analyse workflow.

Get started to see it running on your own data.

Frequently asked questions

What is the Webflow MCP server?

The Webflow MCP server is Webflow’s official Model Context Protocol server. It exposes Webflow’s Designer and Data APIs as tools that AI assistants like Claude, ChatGPT and Cursor can call, so you can build pages, manage CMS content, update SEO metadata and read Webflow Analyze data by describing what you want in plain English.

How do I connect Webflow to Claude?

In Claude Desktop, click the plus icon, choose Add connectors, search for Webflow and click Connect. Sign in to Webflow, pick the sites or Workspaces you want Claude to access and click Authorize App. In Claude Code, run the claude mcp add command with the server URL https://mcp.webflow.com/mcp and authenticate through the browser when prompted.

Does the Webflow MCP server cost anything?

No. Webflow MCP 2.0 is available on all Webflow site plans at no additional cost. Feature access follows your existing plan, roles and permissions. Two exceptions: the analytics tools require the Webflow Analyze add-on, and page branching requires an Enterprise plan.

Do I still need the Webflow MCP Bridge App?

Mostly no. Since MCP 2.0, creating and editing elements, components, styles, variables, CMS content, pages and assets all work through the Data API without the Designer open. The Bridge App is only needed for capabilities tied to a live Designer session, such as visual snapshots of elements or reading and changing the current selection, page or breakpoint.

Can an AI agent publish my Webflow site through MCP?

It can, within your existing Webflow permissions. Whether it should is a different question. The agency teams on Webflow’s MCP 2.0 webinar all said they do not publish through MCP without careful human review, and Webflow’s own safe-publish skill adds a preview and confirm step before anything goes live.

What are Agent Instructions in Webflow?

Agent Instructions are markdown rules and skills stored on your Webflow site that tell agents how to work on it: which components to use, what your brand sounds like, which legal language is non-negotiable. The MCP server hands them to any connected agent automatically, and Webflow can now generate them from what it already knows about your site.

Resources and links

• Webflow MCP server documentation
• How the Webflow MCP server works
• Webflow MCP prompt library
• Webflow skills on GitHub
• Webflow MCP server on GitHub
• The agentic web is here. Webflow is built for it.
• Webflow MCP 2.0 press release
• The biggest announcements from Webflow Conf 2026
• Building AI workflows with Webflow MCP 2.0 (webinar)