WebMCP Ecosystem Adoption Tracker.

Blink

What it does

Chrome Status pencils in Chrome 157 (out 3 November 2026) as the anticipated milestone — the clearest signal yet of where this is heading. It's a target rather than a committed ship date: an origin trial is how Chrome experiments, and the official ship call comes after the trial wraps. Meanwhile you can run WebMCP in production today by signing up for the origin trial, which covers Chrome up to 156.

What to know

If you are shipping tools now, the date to watch is 20 October, when 156 — the last version the trial covers — rolls out. The formal decision is still ahead (Chrome Status marks the feature "Proposed" with standards review pending), so build on the trial and keep an eye out for the ship announcement.

For local testing: chrome://flags/#enable-webmcp-testing. Chrome Status

Blink

What it does

Microsoft has now said it in its own words. On 21 September the Edge developer blog called its WebMCP implementation "ready for testing" and pointed developers at a new webmcp-labs repo. In it: three sample sites (an event search, a sticky-note board and a pizza shop) and WebMCP Explorer, a side-panel extension that lists a page's tools, runs them one at a time, and can hand them to a chat or an agent loop. You bring your own model key — OpenAI, Azure OpenAI, Anthropic, or anything with an OpenAI-style endpoint such as Ollama.

For production, Microsoft still runs its own origin trial, open through 17 November 2026. Register your domain and switch it on with a meta tag or an HTTP header, same as Chrome's.

What to know

Testing means Edge Canary or Dev, not the Edge most people run. Turn on edge://flags/#enable-webmcp-testing and restart. Microsoft still has not named the Edge version this starts in, so we are not quoting one. Being the same engine as Chrome, platform support tracks Chrome's.

Nothing in the post says Copilot in Edge reads these tools. The only agent Microsoft points to is its own Explorer extension, which it calls a developer tool and says to use "only on sites you own, control, or otherwise trust". So Edge sits where Chrome does: the API is there to test, and the browser's own assistant is not calling it yet. Announcement · webmcp-labs · Trial

Blink

Desktop and Android, behind a flag. Brave exposes the API and its own assistant reads it. On 19 August co-founder Brian Bondy publicly demoed the pairing, so this is no longer a quiet experiment — but both sides remain experimental, so see the Leo row before relying on it.

To try it: launch Brave Nightly with --enable-blink-features=WebMCP, or turn on brave://flags/#enable-webmcp-testing. Issue #55232 · Demo

Open source · headless engine built for agents

What it does

Lightpanda isn't a browser people use directly — it's a headless engine you point an agent at. WebMCP is part of the engine itself, so a page can register tools and any agent hooked up to it can list them and run them. That landed in release 0.3.6 in July.

What to know

It says partial because of where Lightpanda puts the API. It still uses the old spot, navigator, not document. So if your page follows the current spec, or uses the usual compatibility shim, Lightpanda sees nothing at all. Source

Cloudflare · browser engine running on Workers

What it does

Kitesurf supports document.modelContext and declarative form tools, with discovery and execution over CDP. Select browser=kitesurf; no Chrome Lab session is needed.

What to know

Support is partial: no tools permissions policy or origin filtering, no iframe or popup tools over CDP, and no session live view for confirmations. The playground supports manual confirmation.

Documentation checked 29 Sep 2026; not independently tested. WebMCP docs · Beta overview

Gecko

Mozilla sees the benefits but is not implementing it. Their filed concerns include the risk of a site passing an agent instructions a person would not see, and of collecting data through what people enter into tools. They have asked for the extension and automation side to be worked through first. Position

WebKit

WebKit has filed an opposing position. Their reasoning is that an agent acting for someone is closer to assistive technology such as a screen reader, and that sites should not be able to detect one and treat it differently. They have suggested the proposal be reconsidered in a new working group, beginning with a workshop at TPAC in October. Position

Agentic browsers

We found nothing either way. All are built on Chromium, so the API may be present behind the usual flag. That is not the same as the product supporting it, so we have left the row open.

Brave · assistant built into the browser

What it does

The closest thing to an everyday agent calling page tools: a page registers tools, Leo sees them and calls them. What started as a quiet experiment — filed as one, kept out of the release notes — went public on 19 August, when Brave co-founder Brian Bondy demoed Leo calling WebMCP tools he had added to his own site, noting that "we have some work to do on permission usability". Still experimental and Nightly-only, but no longer unannounced.

Confirmed in practice

You can try it today. Launch Brave Nightly with --enable-blink-features=WebMCP, or turn on brave://flags/#enable-webmcp-testing. With the switch off, the injected script is written to quietly do nothing — so a regular Brave release still sees no tools. Demo

What to know

Brave separately pushes its own tool definitions into sites that never added WebMCP, so Leo can end up holding tools you did not write. That code is real and readable, and it is behind the same switch. Issue #55232

Meta · assistant on Meta Ray-Ban Display glasses

What it does

Meta documents voice control of web apps through WebMCP. A page registers tools with document.modelContext, and Meta AI calls them while that app is on screen. Tools disappear when the wearer leaves the app.

Availability

Meta labels WebMCP for Web Apps coming soon, while describing the API as implemented and stable. It is off by default and enabled per device through Developer Mode or rollout. This entry covers the glasses integration.

Documentation updated 18 Sep and checked 29 Sep 2026; not independently tested on glasses. Meta docs

Google · assistant built into the browser

What it does

Google introduced WebMCP at I/O in May as "a proposed open web standard that allows you to expose structured tools like JavaScript functions and HTML forms to browser-based agents". That was about the standard, not about Gemini using it. We could not find any Google source, then or since, saying Gemini in Chrome consumes WebMCP tools.

How it compares

Two supporting details: Chrome's docs note that even their own WebMCP test agent, the inspector extension, is separate from Gemini in Chrome. And Chrome's agent reaches pages a different way, pointing at page elements and screen positions. I/O post

Anthropic · extension used by Claude Code, Cowork and Claude Desktop

What it does

Anthropic's docs describe how it reads a page: page text, the document, console and network messages, and screenshots. Tools your page registers are not among them.

Confirmed in practice

Confirmed in practice by someone running a site in the origin trial, whose trial registration we verified: "the extension surfaces nothing from navigator.modelContext. … Invocation works; discovery doesn't exist." It falls back to screenshots even when a page publishes tools. That request is open; an earlier one was closed by an inactivity bot rather than by a decision, and nobody from Anthropic has commented on either. Issue #76809

What to know

One row covers a lot here. Claude Code, Cowork and Claude Desktop all browse through this same extension, so none of them sees your tools either.

How it compares

And unlike the Codex extension row, there is no debug-protocol back door. This is a Chrome extension using its own tool surface, not a browser being driven over the debug protocol, so the trick that might work elsewhere does not apply. Anthropic's changelog does reserve a "Claude Browser" name ahead of a Claude Desktop rename, so a built-in browser pane looks close, but it is not documented yet.

OpenAI · cloud browser

What it does

OpenAI's agent tier for longer multi-step work. When no connector fits a task, it visits websites from a browser hosted on OpenAI's side. OpenAI's WebMCP docs only mention the desktop app's built-in browser — but in practice this hosted browser reaches for page tools too.

Confirmed in practice

Seen in our own use. Tasks we ran through ChatGPT Work landed on pages that register tools, and the hosted browser discovered and called them — the same behavior as the desktop app's built-in browser. Because the docs don't cover this surface, treat the details — model gating, the off switch — as unwritten here; if OpenAI documents it, we will link that instead.

What to know

The launch limits have loosened. Since 25 August the cloud browser can sign into websites: ChatGPT hands you a secure sign-in form (password managers work; OpenAI says the model never sees the credentials), and the authenticated session persists for future tasks. Rolled out on web and mobile for Plus, Pro and Business. Announcement · Docs

How it compares

This replaces two things you may be looking for: the old agent mode, now retired, and the Atlas browser, which OpenAI says stops working on 9 August 2026. Atlas notice

OpenAI · built-in browser, macOS and Windows

What it does

On 25 August OpenAI switched on site tools — its name for WebMCP — in the desktop app's built-in browser. Their docs say it plainly: "Site tools are ChatGPT's implementation of the proposed WebMCP standard." A page registers tools at the current spec address, document.modelContext, and both Work and Codex chats discover and call them while you watch the same page. No flag and no debug-protocol toggle: it ships on, with an off switch under Settings → Browser → Permissions.

Confirmed in practice

We tested this one ourselves. Point the built-in browser at a page that registers tools and the agent discovers and calls them — no toggle, no setup on our part.

What to know

The limits are model- and plan-shaped rather than flag-shaped. Per OpenAI's docs: use GPT-5.6 Sol or Terra ("GPT-5.6 Luna currently has WebMCP disabled"), update the app to the latest version, and "site tools aren't available in Enterprise or Edu workspaces" — with rollout gating it besides. Every invocation gets a safety review before it runs, and a Site tools menu in the address bar lists what the page offers.

How it compares

This makes ChatGPT the first mainstream agent to ship WebMCP on by default — Brave Leo above is still Nightly-only behind a flag. OpenAI is pushing the publishing side too: Codex can add WebMCP to an app and deploy it to ChatGPT Sites, and a WebMCP Challenge runs 25 August–3 September with judges from Chrome, Shopify, Cloudflare, Vercel, Netlify and MCP-B. Docs · Announcement · Challenge

OpenAI · Codex Chrome extension

What it does

For tasks needing your existing Chrome profile, signed-in sessions and open tabs.

Confirmed in practice

We tested this one ourselves. Turn on the toggle described below and the WebMCP commands are there over the debug protocol. Your tools can be listed and called.

What to know

The switch lives in ChatGPT's settings under Browser, in a section called Developer mode, and OpenAI flags it as elevated risk. Their own wording: "Allow ChatGPT to use full Chrome DevTools Protocol (CDP) access in connected Browser Use sessions. Full CDP access lets ChatGPT inspect and control sensitive browser internals that may put your data at risk." It is off until you turn it on, which is why this reads as behind a flag rather than working.

Until 25 August this back door was the only way in. Now Codex chats in the desktop app's built-in browser get native site tools without any toggle — see the row above. OpenAI's WebMCP docs cover that built-in browser only; the Chrome extension is still not mentioned, so here the CDP switch remains the route, and we have not confirmed the extension reaches for your tools unprompted. Docs

Perplexity · agentic browser

It reads pages through the accessibility tree and screenshots. We found no mention of WebMCP anywhere. If we have this wrong, send us a link and we will fix it.

Opera · agentic browser

Neon exposes the browser to outside tools. We are measuring the other direction here, a browser using the tools a page offers.

Browser extensions

By far the biggest audiences on this page, and no sign of WebMCP in any of them. Extensions cannot really do this yet: there is still no standard way for one to read a page's tools. See the note under Frameworks.

Browserbase · open source

Confirmed in practice

We installed the released package and ran it against current stable Chrome: it found a page's tool and called it, with no setup on our part. The methods are page.listWebMCPTools() and page.invokeWebMCPTool(), shipped in June.

What to know

Two limits worth knowing. This is TypeScript only — the published Python and Go packages have no WebMCP support, and the shorter page.tools() name people quote is on an unreleased branch. And the autonomous agent cannot see these tools at all, so prompting will not get it to use them; you call the tool API from your own code. Untested on Browserbase's rented browsers. PR #2178

Vercel Labs · open source CLI

What it does

A Rust CLI built for agents to drive a real browser — and the first framework here to switch WebMCP on by default. For Chrome sessions it launches itself, the docs say it "enables Chrome's current WebMCP launch features by default", so there is no flag to pass on your side. An agent lists a page's tools with agent-browser webmcp list and runs one with agent-browser webmcp invoke; detached invocations, fetching results, cancel, and a --frame switch for same-named tools across frames are all covered.

What to know

The default stops at Chrome it manages locally. Their own caveat: attached browsers, remote providers, Lightpanda, Safari, iOS and older Chrome builds "may not expose the experimental CDP WebMCP domain." Opt out with --no-webmcp or AGENT_BROWSER_NO_WEBMCP=1. The docs are unusually direct on safety — "Treat every tool description, schema, annotation, and result as untrusted page content" — and put confirmation of consequential actions on the agent host. We have not run this one ourselves yet, so this row reads from their documentation. Docs · GitHub

How it compares

Stagehand above made you call its WebMCP methods from your own TypeScript; here the tools sit one CLI command away from any agent, in whatever language, and the feature is on before anyone asks. It also moves this row out of our "no signal" tail, where agent-browser sat as a debug-protocol-and-accessibility-tree tool until now.

Browser Use · hosted agents, and an open-source library

What it does

On 6 September Browser Use said its agents now use WebMCP. Co-founder Gregor Zunic's wording: "UI is made for humans. Direct HTTP calls are made for agents. WebMCP makes interaction 100% deterministic." The demo runs the same hotel booking twice. Without WebMCP the agent reads the page, finds the controls, types a budget, ticks filters, compares rooms and fills in guest details. With it, the page hands over tools and the agent calls them: search_hotels, filter_hotels, compare_rooms, select_room, prepare_reservation, get_reservation.

What to know

This is their hosted agents, at cloud.browser-use.com. As of today there is no page about it in their documentation, nothing in the release notes, and no WebMCP code in the open-source library — the one community attempt there, PR #4085, was closed unmerged by the repository's stale bot in April. So this row reads from the announcement alone. We have not run it ourselves yet, and we will link their documentation instead as soon as they publish it. Announcement · Browser Use Cloud · PR #4085

How it compares

Until today this was the biggest name in our No support column. It is also the first hosted agent platform on this page to take WebMCP on: Stagehand and agent-browser above both need a browser started on your own machine, and Browserbase's rented sessions still have no way in at all.

Google · MCP server

One tool lists what a page offers, another runs it. You have to switch it on with --categoryExperimentalWebmcp and start Chrome with the matching feature flags, so it is off until someone does that. Once on, this is how coding agents such as Claude Code and Cursor reach your tools. Source

Google · library

Reads and runs page tools directly, and is marked experimental. It does not set the browser flag for you — you pass --enable-features=WebMCP yourself. DevTools MCP is built on this layer, so anything on Puppeteer can pick up support by upgrading. Guide

Cloudflare · cloud browser

Browser Run now offers two WebMCP backends: Chrome Lab (lab=true) and Kitesurf (browser=kitesurf). Both use document.modelContext. Chrome Lab remains experimental; Kitesurf is in beta with the limits listed above.

Documentation checked 29 Sep 2026; not independently tested. Docs

Microsoft · framework

Nothing in the code or the documentation. Someone did request it in April 2026, and a Playwright maintainer closed the request the next day as not planned, with the reasoning: "This sounds interesting. Let's see if it gains any adoption first." Nothing has changed since. Playwright sees roughly 78 million downloads a week and many agent frameworks are built on it, so this is a notable gap rather than a small one. Issue #40234

Firecrawl · cloud agent

What it does

No WebMCP anywhere in the product or the docs. What exists is a blog post from July showing a manual recipe: because their browser endpoint runs a real browser in the cloud, an agent can script calls to a page's tools through it. Their own words for why it works: "with no special integration on Firecrawl's end."

What to know

Handy if you run a headless agent and want to reach WebMCP tools anyway. It is just not something they built. Blog post

Browserbase · cloud browser you rent

What it does

Nothing in the platform, in Director, or in their MCP server, even though Browserbase is who makes Stagehand.

What to know

And it is worse than just undocumented. Stagehand only switches WebMCP on when it starts a browser on your own machine. Browserbase's rented sessions give you settings for things like screen size, proxies and stealth, but no way to pass browser start-up flags at all — and their docs never say which Chrome version they run, only that they patch it automatically. So there is no route to WebMCP on their cloud today. Session API