« All posts

hwatu: A Warm-Daemon Verification Browser for AI Coding Agents

hwatu offers a warm-daemon verification browser for AI agents, running one-call, sub-40ms visual and behavioral page checks.

hwatu is a new tool that lets AI coding agents verify their own output visually and behaviorally instead of guessing. Unlike Playwright or chrome-devtools-mcp, which are cold-starting automation libraries, hwatu runs as a persistent warm daemon: engine, GPU context, and a prewarmed WebView stay resident, ready to serve checks instantly. It ships as a single static binary plus the system's webkitgtk, speaking MCP, CLI, or a one-line JSON socket protocol.

The tool bundles verification primitives directly into the browser: pixel-diff scoring with heatmaps, animations expressed as numeric values, deterministic animation seeking, DOM state as structured tokens, and real input events, all reachable through a single 'check' call. According to the project's benchmarks, a full verification pass — open, load, evaluate, screenshot — completes in roughly 35 ms, far faster than a warm Playwright server driven the same way, which measured 341 ms per call.

For engineers, windows are headless by default but can be handed to a human instantly via 'focus', preserving cookies and session state inside their tiling window manager. This makes agent-to-human handoff seamless for CAPTCHAs or judgment calls, since headless is treated as a switchable window property rather than a fixed launch-time decision.

This synthesis was produced from its source by AI; there is no human editor or manual review step. How we work