📰 Story

Show HN: AG2B – Run the agent loop in the browser, expose your tools via WebMCP

Hello everyone, TL;DR Live demo: https://ag2b-example.vercel.app Working on different projects, especially in B2B, I am getting the same request more and more often - "Add an AI feature, yesterday!" Most agent frameworks I found run the runtime on the server - and that pulls in a whole stack. Both backend and frontend have to learn how that runtime works and adopt its SDKs. The server turns into an orchestrator carrying most of the logic. For a feature that's really just "automate what the app already does, and add a brain," I kept asking the same thing: do we actually need all of that and how to ship these AI features faster? AG2B (Agent to Browser) is my research on this question. The agent loop runs in the browser, where your app already is. Tools (the first primitive) are just your existing client functions (store actions, click handlers, whatever you already wrote), so there's no second tool registry and no contract to negotiate between the frontend and backend. The server shrinks (but not limited) to a thin proxy that keeps your API key out of the bundle plus optionally gates with user's permissions. Because the loop runs where the UI lives, each turn can carry a fresh snapshot of what the user sees and can do right now. Scopes (the second primitive) re-inject live context on every iteration - into the system or user prompt, your choice. The model sees what's actually on screen right now. Nothing stale, zero roundtrips and clear context for the LLM. On security take, do