Trellis, a Personal WebMCP Marketplace

A catalog of tested WebMCP tools for sites that have not shipped any, plus the extension that installs them and the agent that writes them for you.

Why I built it

I have spent this year putting WebMCP tools onto a production site. The problem is who has to ship it.

WebMCP only works if the site adds it. That is the right design, but it splits the web in two. A few sites will offer tools. The rest will still use actuation. We created a way to use the protocol on sites that have not adopted it, and an argument aimed at those sites: here are the tools for your pages, here is proof they work, here is a user already calling them.

What it does

Point the extension at a site. It asks a public catalog whether anyone has published tools for that hostname, shows you what they would touch, and on your say-so installs them into the page's real WebMCP surface. From there they are ordinary tools. Chrome's own agent sees them, my extension sees them, and I drive them today with a browser extension with an agent-in-browser on Chrome and with Codex.

Nobody uploads code. A tool is a name, a description, an input schema, and steps from a fixed set of nine operations. One interpreter runs them, and it cannot reach the network, touch storage, or point at anything the tool did not declare. Every tool is run against the real page before publishing, and both its result and everything it touched are recorded.

What I learned

Single page apps were the hard part. A site like that never reloads, so tools installed on the home route are still in the agent's list after the user clicks into a product page, pointing at elements that are gone.

Browsers only announce a real page load. A route change inside a single page app arrives as a separate event, and if you do not listen for it you never learn the page changed. So the extension listens for both: a real load reinstalls everything, a route change only re-decides which tools still apply. The catalog makes that cheap, because the answer for a hostname is one lookup it can repeat on every route change with no help from the site.

The other lesson was about who gets to write a tool. Use agents to write the code, host them locally, and inject them via an extension you own. Then any webmcp enabled site you have crawled can use the webmcp api and tools for now.

The honest limit is that as pages change, you'll need to recrawl.