AI Pulse · Under the Radar

A 12-megabyte Go binary called PinchTab is quietly gaining traction among developers building AI agents. It gives you full Chrome control through plain HTTP calls, no framework required. The pitch is simple: it’s a lighter, cheaper alternative to Playwright-MCP.

The efficiency claim matters. Instead of sending screenshots to your language model, PinchTab serves accessibility-tree snapshots. That cuts token cost from roughly 10,000 per page down to 800. Over hundreds of agent runs, that difference compounds fast. The approach showed up in an X thread, a dev.to writeup, and an open feature request on NousResearch’s Hermes agent repo, all in the same week. That kind of convergence usually signals something people actually want.

Why it spreads

Framework-agnostic tools travel well. If you’re running agents in Python, Node, or anything else that can make HTTP requests, you can drop PinchTab in without rewriting your stack. The binary runs locally, you point your agent at it, and it handles the browser. No ceremony.

The accessibility-tree method also sidesteps a common problem: vision models are expensive and slow when you’re just trying to parse a form or click a button. Trees give structure. They’re smaller, faster, and often enough for navigation tasks.

This is the kind of practical tooling that ends up inside an AI command centre, the sort of thing we wire into the Omni platform when a client needs reliable, repeatable browser automation without burning budget on tokens. PinchTab won’t make headlines, but it solves a real cost problem for anyone running agents at scale. If you’re building or buying agent workflows that touch the web, it’s worth watching.