# Pinnacles

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Harsh Kumar](https://cerebralvalley.ai/u/harsh_kumar), [roshan pradhan](https://cerebralvalley.ai/u/roshan1800), [Ankush Kapoor](https://cerebralvalley.ai/u/ankushhkapoor)
- **GitHub:** https://github.com/Harsh-Kumar-6/webmcp-bridge
- **Demo video:** https://drive.google.com/file/d/1Ma6nIxpGmyw8tQRLSDBw4aohiX8R5O6F/view?usp=sharing
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3351

WebMCP Bridge turns websites into tested MCP tools, without waiting for the site owner.

Meta's WebMCP and AI Connectors let agents call a website's functions, but only when its owner adds tools, and most sites haven't. The usual workaround is a browser agent that runs a model on every step: look at a screenshot, decide, click, repeat. That is slow and costs money on every run.

WebMCP Bridge does that work once. You give it a URL and a goal in plain English. Muse Spark drives a real browser toward the goal and captures the requests behind the page. Spark then proposes tools, inputs and tests. Plain code runs the tests, and only tools that pass are shipped. Canvas pages with no usable HTML or API take a second route: Spark reads the legend, SAM 3.1 finds the controls in a screenshot, and code reads pixels and clicks saved coordinates.

Each compiled site gets an MCP server URL and a generated WebMCP script. After compiling, a direct tool call uses no model and still fetches live data.

We built and tested it on Hacker News search, Wikipedia search and a canvas seat map we made for the demo. SAM 3.1 found all 60 seats on that map, and every label matched the page's real layout. In a five-task benchmark with one run per task, direct tool calls were about 8 to 50 times faster than a Spark screenshot agent, with no model cost per call, and both completed all 5 tasks correctly.

Safety limits are enforced in code: only read-only API tools ship, passwords are never filled, exploration stops at captchas, clicks such as buy, pay, delete and log in are blocked, and site credentials are hidden from agents.

It focuses on public, read-only actions, so it stays safe by design.

## More from Meta Global AI Developer Hackathon

- [almona.eg](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3348)
- [Derm Forge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3349)
- [Farfaz Learning](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3350)
- [Ahoo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3352)
- [Blockparty](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3353)
- [Steel AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3354)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3351. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt