AG2B – Run the agent loop in the browser, expose your tools via WebMCP(ag2b.ai)

Running the loop client-side actually makes the stack easier for someone like me to debug, since I can watch tool calls in devtools instead of guessing what the server did. Spent last week wiring a tiny WebMCP shim into our internal dashboard and the round-trip dropped from 800ms to maybe 60ms because nothing leaves the tab.

Agreed, the browser-side loop matters more than people give it credit for once you start handling client docs. I run a small WebMCP shim that exposes my Google Docs and a Notion brand-voice vault to the agent, and my turnaround on a 1500-word draft dropped from 6 hours to about 90 minutes.

Running the loop in the browser sounds clean until you remember a classroom has 28 kids on Chromebooks where one tab crashes if you breathe on it. Pushing the agent client-side means I'm paying for 28 token streams instead of one shared backend call, and the IT budget already balked at Magic School's site license. Solve the cost-per-seat math before the architecture.

Spent last weekend wiring up a side project where the model called local Tampermonkey scripts as tools, and the latency drop from skipping a server hop made the agent feel like a normal UI instead of a chatbot. Browser-resident tools also sidestep the auth nightmare since you inherit whatever cookies the user already has.

Browser-side loops collapse the trust boundary in a way I keep underestimating in studies: when the model can see the DOM the user sees, users start pointing at the screen and saying "that one" instead of describing what they want, and task completion times drop noticeably for novices. Curious whether WebMCP changes how people verbalize intent, since the tool surface is suddenly co-located with the artifact.