# browser-rs

> Browser automation for AI agents, in Rust. A small MCP server that drives a
> bundled Chromium (Chrome for Testing) with Playwright MCP's tool names and
> parameters, plus tools agents asked for: pages as Markdown instead of big
> snapshots, many steps per call, direct HTTP with the browser's cookies,
> parallel page reads, and a hand-off that shows a person the exact live page.
> Over four real pages it used 535 MB where @playwright/mcp used 1,496 MB. MIT.

Made by Luka Löhr — https://lukaloehr.com

    git clone https://github.com/luka-loehr/browser-rs
    cd browser-rs && cargo build --release
    claude mcp add browser-rs "$PWD/target/release/browser-rs"

Tested on macOS. Linux uses the same code but is untested; Windows is not
supported.

## Setup

- [Agent setup runbook](https://browser-rs.lukaloehr.com/setup.txt): written for a coding agent. Checks the toolchain, builds, registers the server, installs Chromium and verifies with a real page. Start here.

## How to drive it fast

- Read with browser_text (Markdown, `lead`, `heading`), browser_links, browser_table, browser_extract, browser_read or browser_fetch instead of snapshots.
- Act with refs or selectors (`text=`, `role=`, `link=<regex>`, `href=`, `a >> b` into iframes and shadow roots), several steps per browser_batch.
- Pass `snapshot: "none"` on actions you do not need to see; the default reply holds only changed lines.
- browser_scrape reads many URLs in parallel; browser_fetch with `fields` or `transform` returns only the JSON you need.
- browser_handoff when a person must log in, pass 2FA or a CAPTCHA; the page does not reload.
- Every reply ends with elapsed_ms.

## Source

- [GitHub repository](https://github.com/luka-loehr/browser-rs): source, README with the full tool list, benchmark script. MIT.