# PawBrowse - I made my agent's browser automation ~2x faster by giving it an element table instead of screenshots (open source, reproducible benchmark) — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/b4aefa0889c025f4

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-09-27 11:47 ingested 2026-09-28 02:37

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1wri6sn/pawbrowse_i_made_my_agents_browser_automation_2x/)

## retained text



```
Built with Claude Code, for Claude Code users who want their agent to drive the browser faster.

The mechanism: most browser drivers work off screenshots, so every action is perceive-then-act = two round-trips, and each round-trip is a full model inference. PawBrowse instead hands the agent a compact table of the clickable things in view (real accessible names, values, state) and returns the fresh table with every action, so an action is one round-trip.

I measured it against Claude-in-Chrome on the same task, same model, both filmed live (book a 4-star Lisbon hotel on Booking.com): 5 agent round-trips vs 10. ~2x fewer, so ~2x faster in the round-trip-bound loop and cheaper in tokens. One run each, fully reproducible from the repo (the exact task + the render script are included; that's the gif). I'm not claiming more than the round-trip count shows.

It's an open-source Chrome extension + a one-file, zero-dependency MCP server, and it's free (MIT). Drives your real, logged-in Chrome with no remote-debug port, no relaunch, no second model, and no API key. Live on the Chrome Web Store + npm.

Repo: [github.com/ItaiZeilig/pawbrowse](https://github.com/ItaiZeilig/pawbrowse) - I'm the author, happy to answer anything.
```



document fetched 2026-09-28 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1wri6sn/pawbrowse_i_made_my_agents_browser_automation_2x
