BrowserMatic

Add WebMCP to any site instantly. No install, no extension simply enter a web address, we analyse the site, and open it in a browser with WebMCP applied in real time.

Comment

Inspiration

Every one of these is a tab-juggling exercise:

  • Find the best deal across four stores
  • Book dinner and movie tickets together
  • Plan and price a trip in one shot
  • Fill in council forms and applications

Four stores, two booking sites, three travel sites, the same thirty fields on the twentieth government form. None of it is hard. It's just tedious, which is exactly what you'd want an assistant for.

They all hit the same wall. Sites are starting to publish a machine-readable list of actions an AI is allowed to take, but those tools only exist on that site's own page. An agent can help you on one store, and that's the whole conversation. Four stores means four conversations, and nobody wants that.

What it does

Type a web address. BrowserMatic opens it in a browser that installs the API before the site's own code runs, so the site registers its tools and we carry them into your assistant, origin-qualified, across as many sites as you want.

Four stores in one conversation. Flights and hotels side by side. And for forms, the same name and address on the twentieth site without typing it again.

BrowserMatic Screenshot

How we built it

A Cloudflare Worker with one Durable Object per session, each with its own Browser Rendering Chromium, consent list and audit log.

BrowserMatic Architecture

It also works directly in ChatGPT's in-app browser, no declarative form and no iframe discovery, so tools register imperatively on the top-level page; no session cookie, so a session is a capability URL. Names are origin-qualified because its permission model is per-site, and each tool holds its own AbortController, so granting a fourth store never disturbs the first three.

The injection is one line of Playwright's addInitScript, a minimal WebMCP implementation installed before any page script runs. The same Worker also speaks MCP at /mcp over JSON-RPC with OAuth 2.1 and PKCE, verified against the official SDK, so any compliant client can drive the same session headlessly.

Challenges

A Chrome extension would have been the easy route. We used the hackathon to find out how far embedded remote browsers could go instead, since that's the lower-friction path for users. Most of the pain came from rendering a session a human and an AI can both drive at the same time. We're happy with where it landed in the time we had.

We are still getting stuck in some "honey pots" so need to do a better job at detecting and mitigating.

What's next

A registry of the sites we've driven and the patterns we generate for them, so a success on one visit gets reused on the next. Then usage data for site owners so they can see how agents behave on their pages and improve accordingly, something like Google Analytics / Datadog for agentic commerce. We also want to try Cloudflare's agent payments and pay-per-access content as a business model.

Built With

Share this project:

Updates