Chrome for Developers reposted this
Ive been having a lot of fun going through the WebMCP hackathon winners. One idea showed up in several of them: do the math in your tool code, and give the model a simple answer. WebMCP tools run inside the web page, so they can use the GPU, draw on a canvas, or run plain JavaScript before replying to the agent. Some examples: 🧠 Faraday (MRI viewer) uses WebGPU to crunch brain scan data. The agent asks for sizes in milliliters, and the page does the conversion. The scan stays on the users machine. 🏠 Alza (floor plan editor) checks for furniture collisions and replies with short codes like door-blocked. The agent moves things around until no issues are left. 💐 Bouquet Studio measures the arrangement and returns numbers, like how visible each flower is and how much space is still empty. The tool description sets targets for both. 🖌️ MASIL (calligraphy studio) draws each brush stroke off-screen first and checks that it will actually be visible. What I like about this: the model is good at deciding what to do next, and code is good at math. Splitting the work that way plays to both. Numbers cant capture everything about how something looks, but they give the model a clear target. Great work from all four teams! 👏 Repo links are in the comments. Have you tried anything like this in your own tools? Id like to hear how it went. Also, thank you to all the Hackathon judges for surfacing all the cool app! Sarah Drasner Alex Nahas Kasper Kulikowski #WebMCP #WebDev
