<- all tokdocs

WebMCP and Chrome's Prompt API: Can You Build a Free Site Chatbot?

By , founder of Mavric and Apso

View on TikTok ->

WebMCP is a draft web standard that lets a page register tools, such as search_products or add_to_cart, that an AI agent can call. A TikTok from @agentic.conversations, cut from an interview with Google's Dominic Farolino, says sites can feed those tools into Chrome's Prompt API to get a "free chatbot" that acts on the page. The syntax is real, but the tools option is marked experimental and WebMCP is an origin trial from Chrome 149, so this does not run on stable Chrome today.

What is WebMCP?

WebMCP is published by the Web Machine Learning Community Group as a "Draft Community Group Report, 20 October 2026." That is an incubation venue, not the W3C Recommendation track. Chrome's documentation describes it as "A standard way for pages to register tools with agents."

In the current draft, each Document has its own ModelContext, and pages register tools with document.modelContext.registerTool() and list them with document.modelContext.getTools(). Much material from earlier in 2026 uses navigator.modelContext, so start from the current draft if you are building against it.

Can the Prompt API use WebMCP tools?

Yes, in experimental contexts. The video shows the line LanguageModel.create({ tools }), and that is genuine syntax. MDN documents tools as "An array of objects representing tools available to the AI." Each tool carries a name, a description, an inputSchema (a JSON Schema for its parameters) and an execute callback that returns a promise resolving to a string.

The Prompt API Draft Community Group Report of 16 October 2026 defines the option this way:

dictionary LanguageModelCreateCoreOptions {
  // Tools that the language model can use.
  // **EXPERIMENTAL**: Only available in experimental contexts.
  sequence<LanguageModelToolDeclaration> tools = [];
};

Chrome's own Prompt API documentation covers monitor, signal, samplingMode, initialPrompts, expectedInputs and expectedOutputs, and does not list tools at all. MDN notes the feature "is not Baseline because it does not work in some of the most widely-used browsers." The base Prompt API is stable in Chrome 148 on the web and Chrome 138 in extensions, but the tools option on it is not.

Is WebMCP available in Chrome?

Only through an origin trial. Chrome's WebMCP page tells developers to "Join the 'WebMCP origin trial' from Chrome 149" and warns that "WebMCP is under active discussion and subject to change in the future." For local testing it points to the flag chrome://flags/#enable-webmcp-testing. The video describes the setup in the present tense as something "people are using," and leaves this out.

Is an on-device chatbot really free?

Free means no API key and no per-token bill for the site owner, because the Prompt API runs Chrome's on-device model. The cost moves to the visitor's machine. Chrome's getting-started page lists the requirements:

  • At least 22 GB of free space on the volume with the Chrome profile. If free space drops below 10 GB after download, the model is removed.
  • More than 4 GB of VRAM, or 16 GB of RAM and 4 CPU cores.
  • An unmetered connection for the first model download. Later use works offline.

Platform support is the bigger limit for a shop. The model APIs run on Windows 10 or 11, macOS 13 or later, Linux, and Chromebook Plus devices. Chrome's page states that "Chrome for Android, iOS, and non-Chromebook Plus ChromeOS devices are not yet supported." A store that relies on this as its only assistant leaves out its mobile visitors and visitors on lower-spec desktops.

How do in-page agents and the browser agent work together?

The first half of the clip is about coordination between the browser's own agent and agents embedded in page iframes. Its diagram, an "in-page agent" inside an iframe connected to a separate "browser agent," matches what the spec models.

Because each framed document has its own ModelContext, an iframe registers tools in its own context. A tool's exposedTo member is "An array of origins that control which documents this tool is exposed to, in the current document's tree." Execution across separate top-level browsing contexts is not allowed, and an origin mismatch rejects with a NotFoundError.

The caption "Sub-agents" frames this as Farolino's idea, but the phrase "It's almost like a sub-agent" comes from the interviewer, and Farolino answers "Yeah, exactly." "Sub-agent" is not a term the WebMCP spec uses. The clip uses drawn illustrations and a reserved example domain, trailshop.example, and never shows a working demo.

Who is building WebMCP?

The video's description credits "Dominic Farolino's team at Google." Farolino is billed on the AGNTCon + MCPCon Europe 2026 schedule as "Editor of the WebMCP Specification & Software Engineer, Google." The draft lists four editors, though: Brandon Walderman (Microsoft), Khushal Sagar (Google), Dominic Farolino (Google) and Samuel Dallstream (Microsoft). The repository acknowledges more contributors from both companies and notes that the draft "has evolved significantly, primarily driven by Dominic Farolino."

The interview was recorded at AGNTCon + MCPCon Europe in Amsterdam on 17 and 18 September 2026, where Farolino gave the talk "Three Doors to One Tool: MCP vs WebMCP vs CLI." The clip went up 37 days later.

Bottom line

Feeding a page's WebMCP tools into LanguageModel.create({ tools }) is real and spec-defined. It is experimental, WebMCP itself requires an origin trial from Chrome 149 or a testing flag, and the on-device model excludes Android, iOS and lower-spec machines. Treat it as something to prototype now, not something to ship as a site's only chatbot.

Sources

Published October 25, 2026. Based on a TikTok by @agentic.conversations. Full video notes (markdown)