What is WebMCP?

WebMCP is a browser API standard that lets websites expose structured tools to AI agents natively — without scraping, without hallucination.

The problem with current AI browsing

When an AI agent visits a webpage today, it reads HTML like a human would read a book. It sees text, infers structure, and guesses what actions are possible. This leads to:

  • Hallucinated actions — the AI thinks a button does something it doesn't
  • Fragile automation — selectors break when the page updates
  • No discoverability — the agent has to try things to learn what's possible

How WebMCP works

WebMCP adds a navigator.modelContext API to the browser. Websites register tools on this object — structured definitions of what actions are available, what parameters they accept, and how to execute them.

// What happens behind the scenes when your snippet loads
navigator.modelContext.registerTool({
  name: "addToCart",
  description: "Add a product to the cart",
  inputSchema: { ... },
  execute: (args) => { /* your executeJs */ }
})

When a WebMCP-compatible AI agent visits the page, it queries navigator.modelContext.tools and sees exactly what's available — no guessing required.

Agentable's role

Agentable is the management and analytics layer on top of WebMCP. Instead of manually writing JavaScript to register tools on every page, you:

  1. Define tools in the Agentable dashboard
  2. Paste one script tag
  3. The script fetches your tool definitions and registers them automatically

The polyfill

Native navigator.modelContext support is still rolling out across browsers. The Agentable snippet automatically loads @mcp-b/webmcp-polyfill for browsers that don't support it natively yet. This means your tools work everywhere today.

Further reading