WebMCP integration

What is WebMCP?

WebMCP (Web Model Context Protocol) is a browser API announced by Google in Chrome 146 that lets websites expose structured tools directly to AI agents. Instead of an agent parsing the DOM or reading accessibility trees to figure out what a page does, the website tells the agent exactly what actions are available.

Google and Microsoft co-developed it. It’s being standardized through the W3C’s Web Machine Learning community group.

Two APIs make up WebMCP:

Declarative API - Websites add a toolname attribute to HTML forms. The agent reads the form schema and submits structured data. No JavaScript needed.

Imperative API - Websites register tools via navigator.modelContext.registerTool() with a name, description, input schema, and execute function. This handles complex multi-step workflows.

Early benchmarks show a 67% reduction in computational overhead compared to vision-based browsing, with task accuracy around 98%.

How this relates to MCP

WebMCP and Anthropic’s Model Context Protocol (MCP) are complementary, not competing:

Mudabbir already supports MCP servers for backend connectivity. WebMCP adds the frontend equivalent: structured interaction with websites through Chrome.

Together, they give Mudabbir agents structured access to both backend services and website UIs.

How Mudabbir will use it

Mudabbir currently browses the web using Playwright with accessibility tree snapshots. This works with any website but requires the agent to interpret page structure and figure out what to click.

With WebMCP, when a website exposes structured tools, Mudabbir can call them directly:

Without WebMCP (current):
  Navigate to page -> read accessibility tree -> interpret elements
  -> guess which button to click -> execute action -> hope it worked

With WebMCP (planned):
  Navigate to page -> ask Chrome what tools the site exposes
  -> call add_to_cart({product_id: "123", quantity: 2})
  -> get structured response

The plan is a hybrid approach:

1. Check if the website exposes WebMCP tools
2. If yes, use them (faster, more reliable)
3. If no, fall back to the existing accessibility tree approach

This gives Mudabbir the widest compatibility: structured when available, smart fallback when not.

Privacy angle

WebMCP runs client-side in the browser. Mudabbir runs locally on your machine. Your browsing data stays on your computer, same as it does today with Playwright.

Google’s own agent integration will likely route through Gemini (their cloud). Mudabbir keeps everything local.

Current status

• WebMCP is in early preview in Chrome 146 Canary (behind a flag)
• No general availability date announced
• Websites need to implement WebMCP tools for it to work
• Mass adoption is likely 12-24 months out
• Mudabbir will add support once the API stabilizes

Track progress on GitHub issue #156.

Further reading

• Chrome WebMCP early preview announcement
• W3C Web Machine Learning community group
• Mudabbir browser automation docs
• MCP server integration