WebMCP vs browser MCP servers: what each does and why agents need both
By the pilotfish team · Last updated
WebMCP and browser MCP servers both let an AI agent use the web, from opposite sides. With WebMCP a website registers its own tools for agents inside the page. A browser MCP server gives the agent a browser to work any page, tools or not. Most sites offer no tools yet, so agents need the browser; where a page does offer them, calling the tool beats clicking. pilotfish uses both.
Key takeaways
- WebMCP is the website's side: a page registers tools for agents (a name, a description, a JSON input schema and a function), or marks up a form as one.
- A browser MCP server is the agent's side: it gives an AI app a browser to read and work any page, whether the site offers tools or not.
- Most sites offer no tools yet, so an agent still needs a browser; where a page does, calling its tool is faster and less brittle than clicking.
- pilotfish does both: it lists a page's WebMCP tools with the page and lets the agent call them, under the same approval rules as a click.
What is WebMCP?
WebMCP is a proposed web standard from the W3C Web Machine Learning community group (draft), with an early implementation in Chrome. A page uses it to tell an agent what it can do, in the agent's terms: a tool such as searchFlights with a description and an input schema, which runs the page's own code when called. A form can be offered the same way with a few attributes. The site decides which tools exist and what they do.
What is a browser MCP server?
A browser MCP server, such as Playwright MCP, Chrome DevTools MCP, Browserbase or pilotfish, gives an AI app tools to drive a whole browser: open a page, read it, click, type and so on. It works on any site, because it uses the page the way a person does. We compare them in MCP servers for browser automation.
How do WebMCP and a browser MCP server differ?
| WebMCP | Browser MCP server | |
|---|---|---|
| Who builds it | The website | The agent's side (a browser or automation vendor) |
| Where it runs | Inside the page | Beside the browser, on your machine or in a cloud |
| What it can do | Only the tools the site registered | Anything a person can do in the page |
| Works on | Sites that adopted it | Any site |
| Breaks when the layout changes | No: the tool stays the same | It can: a button moves or is renamed |
Do you need both?
An agent that does real tasks does. Few sites offer WebMCP tools today, so the browser is still how an agent reaches the rest. Where a page does offer one, calling it is the better step: one call with a typed input rather than several clicks that can miss. The two meet in the browser: the page's tools are only there while the page is open.
How pilotfish uses WebMCP
pilotfish reads a page's tools through Chrome's own WebMCP support, without running any of the page's script itself, and lists them with the page as PAGE TOOLS. The agent calls one with page_tool in hosted tasks and browser_page_tool in the open MCP, which appears only once a page offers tools, so it costs nothing elsewhere. A tool's call is held to the same rules as a click: one whose name reads as paying, sending or deleting, or that the page marks as consequential, needs your yes, and saved secrets are filled in only for that page's own host. The tool's answer comes back marked as the site's words.
For site owners: should you add WebMCP?
If agents already visit your site to search, book or fill in forms, offering those as tools gives them a reliable path and you a say in how it is used. Start with read-only tools (search, look up) and forms that only fill; mark anything that commits the user as consequential so agents ask first. The draft and Chrome's documentation are the place to start.
Frequently asked questions
What is WebMCP?
A proposed web standard from the W3C Web Machine Learning community group, with an early implementation in Chrome. A website uses it to register tools for AI agents, each with a name, a description, a JSON input schema and a function the page runs, or to offer a form as a tool.
Is WebMCP the same as MCP?
No. MCP connects an AI app to servers over a protocol. WebMCP lets a web page offer tools to an agent that is visiting it. They share ideas (named tools with schemas) but run in different places.
Does WebMCP replace browser automation?
Not while most sites offer no tools. An agent still needs a browser to reach those, and uses a page's WebMCP tools where it has them.
Does pilotfish support WebMCP?
Yes. Hosted pilotfish tasks and the open pilotfish MCP list a page's WebMCP tools and can call them, with the same approval rules as clicks: paying, sending, deleting, or anything the page marks as consequential, waits for your yes.
Which browsers support WebMCP?
Chrome has an early implementation, behind a feature flag and in an origin trial for sites. The open pilotfish MCP starts Chrome with WebMCP switched on; hosted pilotfish tasks see the tools of pages that take part in the origin trial.
Sources
- WebMCP draft (W3C Web Machine Learning community group)
- pilotfish MCP repository (browser_page_tool, @gaiventures/pilotfish-mcp 0.1.58)
- pilotfish open MCP tool reference and how it works