Wave Security Browser MCP Protection: Pages hand agents tools. Wave decides what runs.

I've spent most of my career at the boundary where security meets the browser. Six years in Cisco's Cloud Security CTO office building zero trust, then running product at a browser security company. Every product in that world, including the ones I helped build assumes a person is doing the clicking. Those days are over. There's a second actor in the tab now. Gemini in Chrome, Claude, the ChatGPT extension, Copilot in Edge. And with WebMCP (a W3C spec co-edited by Google and Microsoft), a web page can hand those agents real, callable tools. Book the trip. Move the money. Export the customers. It all happens client-side, inside the page, where network and endpoint controls see exactly nothing. So I started building earlier in the year, before the origin trials even opened. Then the world caught up fast. Chrome shipped the trial in June, researchers published 94–100% hijack rates against these tools the same month, and on August 25 OpenAI turned on the demand side: ChatGPT and Codex now call WebMCP tools out of the box. The approach evolved twice along the way. First, I thought I was building a WebMCP firewall. I wasn't. I as building the boundary itself, the page-to-agent handoff, Second, while agent traffic ramps, the same install turned out to solve a problem teams already have today: on-device redaction and blocking of PII and secrets in ChatGPT, Claude, and Gemini before anything leaves the machine. Wave decides what's allowed to register, what's allowed to run, and what data is allowed to leave, with a block reason the agent can actually read. Built solo over five months. I'd love your take, especially the skeptical one. Ask me anything.