Site2WebMCP

Inspiration

Nobody will hand-write WebMCP for every website. Zero site setup.

WebMCP lets agents use a site as tools instead of clicking through the UI. That only works if someone builds the tools. Every site is different — hand-writing and maintaining that for the whole web is the real bottleneck. We flipped it: the person who needs agent tools shouldn’t wait on every company to ship WebMCP.

What it does

Site2WebMCP is a Chrome extension (also available as an agent skill) that reads the forms and buttons already on a live page and exposes them as WebMCP tools. An agent can then run real workflows — search, add to cart, checkout, and so on — without browser control and without the site owner implementing anything.

How we built it

A simple webpage with two demo modes and WebMCP for auto-setup of demo 1 — a playground only. The product generates WebMCP for any external site.
Chrome MV3 extension: discover forms and primary actions from the live DOM, plus registerTool where needed.
An agent skill that does the same job as the extension — inject into the page, discover DOM actions, register them as WebMCP.

Challenges

Getting from “tools appear” to “an agent actually completes a workflow” was the hard part — especially validating the agent-side path, not just the discovery UI. The runtime still leans on e-commerce-shaped heuristics; proving this beyond that shape is the next fight. Hand-written WebMCP will always be sharper for a given site — this is WebMCP for sites that never shipped it, replaceable when they do.

ChatGPT Desktop’s built-in browser is read-only: no extensions, no injected scripts. That blocks dynamically generated WebMCP there.

Chrome Web Store review is slow, so the extension isn’t installable from the store yet.

What’s next

Use site JavaScript to find server requests and expose them as tools, with descriptions of what those functions do.
Crawl support centres and docs for how to use the site, then turn that into tools and skills.
Use llms.txt to extract tools or workflows (chains of tool calls).
Use an external or browser-built-in LLM to identify workflows on a site and turn them into tools and skills.
Measuring the speed and tokens savings compared to a vanila browser use that does clicks.