browserlisted

# browser The agent writes a short Python script; the script does everything that needs no fresh look at the page in **one** call and prints a snapshot. Speed comes from the number of calls, not from fast clicks: each call costs about a second of overhead, a snapshot costs 0.01-0.13 s. Engine: [browser-harness](https://github.com/browser-use/browser-harness) (MIT), the Python package `browser_harness`, installed separately as a uv tool (references/setup.md). The snapshot is ported from [jev-ultrafast](https://github.com/browser-use/jev-ultrafast) (`snapshot.js`, MIT): one CDP call instead of the accessibility tree and screenshots; for them it cut browser protocol calls from 1,092 to 101. Pagination follows the ideas of [maxun](https://github.com/getmaxun/maxun) (AGPL-3.0: ideas only, the code is ours). Racing a click against a native dialog follows playwright-core's `_raceAgainstModalStates` ([playwright-cli](https://github.com/microsoft/playwright-cli), Apache-2.0). Install, telemetry, profile: [references/setup.md](references/setup.md). ## Script first, the browser is reconnaissance The rule: use a script wherever one can work; the browser walks a site to find how its data and actions go (internal API, headers, anti-bot), so that a script or skill can be written from that. Before opening a site, look for an existing script: `sites.md` for the site, memory, the project's own tools. Kuper shows why: the browser hit a ServicePipe CAPTCHA and then a "turn off your VPN" page