Webflow MCP in 2026: An Honest Review From an Agency That Left Webflow

Between July and October 2026, Webflow shipped three major updates to its MCP server, the piece that lets AI assistants like Claude, ChatGPT and Gemini work on a Webflow site directly. According to Webflow, more than 30% of its enterprise customers already use it, usage has grown more than fourfold since January, and nearly 90% of those users connect through Claude.

We should say up front that we have a stake in this. We built client sites on Webflow for years, and today we move them to Astro with our own CMS. So rather than trade opinions, this review sticks to three things: what Webflow's own documentation says the server can and can't do, how that approach differs from an agent working in a real codebase, and what we measured on a site we migrated ourselves.

The short answer

• For content operations, Webflow MCP is genuinely good now. Bulk CMS edits, SEO metadata, audits, new pages built from your existing components, questions about traffic: it handles all of that, mostly without the Designer open.
• It is still an API window onto an editor built for people. The agent can do what Webflow chooses to expose, on Webflow's runtime and within Webflow's plans. Webflow's own roadmap talks about reaching "full parity" with the Designer, which also tells you it isn't there yet.
• "Bots waste tokens on jQuery" is mostly a myth. The JavaScript cost for real visitors, on the other hand, is very real. Our numbers are below.

What Webflow MCP actually is

MCP (Model Context Protocol) is an open standard that lets an AI assistant call tools inside other software. Webflow's MCP server turns Webflow's Data and Designer APIs into tools an agent can use: read and edit CMS items, create pages, apply styles, publish. You add it to Claude, ChatGPT or Gemini as a connector, or to Claude Code, Cursor and similar clients as a remote server, and authorize one Webflow workspace. The agent works within the roles and permissions your account already has. The server itself is available on every site plan at no extra cost.

What changed in 2026

Where it works well

Version 2.0 changed the character of the tool. The first release was essentially a CMS and metadata API. Now an agent can assemble a page from your existing components and bind it to CMS data. Webflow's own example: hand Claude a folder of customer case studies, and it builds a Collection page in your design system, then generates and publishes every entry. For a marketing team that already lives in Webflow, that's a real time saver.

The jobs we'd hand to it without hesitation:

• Bulk CMS work: importing, rewriting, tagging or translating existing items.
• SEO hygiene: titles, meta descriptions, alt text, spotting thin or duplicate pages.
• New pages built from components you already have.
• Questions like "what happened to traffic on these pages last month?" without opening a dashboard.

The governance features (roles enforced per site, collection, page and locale, an activity log of what the agent changed, branches for testing before going live) are aimed squarely at larger teams, and they're well thought out.

Where it stops

Webflow's developer documentation lists the current limits plainly. As of October 2026:

• No new sites. A project has to be created in the Webflow dashboard before an agent can touch it.
• Classic interactions are invisible. The agent can build interactions on the new GSAP engine, but on a site that still uses Classic interactions it sees an empty list. Those have to be edited by hand.
• No new localized CMS items. It can update existing items in secondary languages, but not create new ones.
• Some tasks still need the Designer open with the Bridge App connected: image uploads, visual snapshots, and anything tied to the current selection, page, breakpoint or canvas ("change this heading").
• One workspace per authorization. Switching means re-authenticating, which in Claude or ChatGPT means removing and reinstalling the connector.
• Fonts and secrets go through side doors. Google and Adobe Fonts live in site settings, not in MCP. Webflow Cloud deploys require an app connected to GitHub, and secret environment variables go through the CLI.
• Some features depend on your plan. Branch merging and team administration need Enterprise, and Analyze, Optimize and AEO data are paid add-ons.

None of these is a dealbreaker on its own. Together they describe the shape of the tool: the agent works through a list of operations Webflow has wrapped for it. When the operation you need isn't on that list, you're back in the Designer.

The real difference: an editor built for people, or a codebase

Webflow solved a genuine problem. Someone in marketing could change a heading, swap an image or adjust a button colour without filing a ticket and waiting for a developer. That's exactly why we used it for years. The visual editor is the product.

MCP teaches that same editor to take instructions from an agent. But an AI agent isn't a person who needs a visual editor. It reads and writes text, and code is text. Give it a codebase and it doesn't need an API translating its intent into editor operations: it opens the file and changes it, the way a developer would.

That's how our Astro + CMS sites work. Content lives in the CMS, where the client edits it in a panel. The site itself is plain code in a Git repository that Claude Code or any other agent can work on directly. We explained what that means in practice, even if you're not a developer, in this article on connecting Claude Code to your website.

None of this makes Webflow MCP pointless. It's a different bet. Webflow is betting that agents should work inside a managed platform with guardrails around them. The code-first approach bets that agents are now good enough to work on the real thing, with Git as the guardrail.

What bots actually read: we measured it

A popular argument against Webflow is that its jQuery and runtime scripts waste AI crawlers' tokens and Google's crawl budget. We wanted to check that against real numbers, so we compared the homepage of Kwalitet, a cleaning business we migrated, as it was on Webflow (an archived copy from 2025) and as it is today on Astro.

Analytics and reCAPTCHA scripts are excluded from both columns. Images are not counted.

The numbers lead to two honest conclusions.

1. For AI crawlers, the jQuery argument doesn't hold. Crawlers like GPTBot and ClaudeBot generally fetch the HTML and don't execute JavaScript, so Webflow's scripts never reach them. Webflow also publishes complete, server-rendered HTML, so the content is all there. In raw HTML the Webflow version was actually lighter than ours, because our build inlines all the CSS into the page. Strip the markup, which is what most AI pipelines do before reading a page, and both versions come down to a rough 1,400 to 1,800 tokens of actual text. Anyone telling you Webflow sites are invisible to AI is wrong.

2. For people, and for Google, JavaScript is the real cost. A phone has to download, parse and run 1.75 MB of script before the Webflow page is fully interactive. The Astro version has almost nothing to run. That's where the drop from 810 ms to 0 ms of Total Blocking Time comes from, and that's what Core Web Vitals measure. Google does render JavaScript, but rendering is a separate and more expensive step than fetching. For a site with a few dozen pages, crawl budget is rarely the bottleneck. Page experience for real visitors usually is.

So the case for lightweight pages isn't that bots choke on jQuery. It's that every visitor shouldn't have to pay for a runtime most pages don't need. If you want the wider picture, our Webflow vs Astro comparison covers speed and cost side by side.

Who should stay on Webflow, and who shouldn't

Webflow + MCP makes sense if your team designs visually in the Designer every week, you're on Enterprise and need its governance, and most of the work you'd give an agent is content operations: CMS, SEO, landing pages from existing components.

A code-first site makes more sense if you want an agent to be able to change anything, not just what an API allows; you care about page weight and mobile performance; you'd rather not pay for per-site plans and add-ons to unlock features; or you simply want to own your site outright.

Frequently asked questions

Does Webflow have an MCP server?

Yes. Webflow runs an official MCP server that works with Claude, ChatGPT, Gemini, Claude Code, Cursor and other MCP clients. It's included on every site plan at no extra cost.

How do I connect Webflow MCP to Claude?

In Claude, add the Webflow connector from the connector directory and authorize the workspace you want the agent to access. In Claude Code, Cursor or another MCP client, add Webflow's remote MCP server using the address in Webflow's developer documentation. Each authorization covers a single workspace.

Can Webflow MCP build a whole website?

It can build pages, components, CMS content and interactions inside an existing project, but it can't create a new site. That still starts in the Webflow dashboard. Image uploads and anything tied to what's selected on the canvas also still need the Designer open.

Is Webflow MCP free?

The server itself costs nothing extra. Some capabilities depend on your plan: branch merging and team administration require Enterprise, and Analyze, Optimize and AEO data are paid add-ons.

If you're weighing whether to stay and wire up MCP, or move to a codebase an agent can work on fully, see our four fixed-price ways to move off Webflow.