browser-use

browser-use

Facts

Openness
Open source
Licence
MIT
Language
Python
Pinned version
5c892e0 (2026-09-04)
Topics checked
31 of 33
Kind of agent
Browser/GUI

What it does notably well

  • Budgets

    Step budget made visible to the model: from 75 % of max_steps a BUDGET WARNING context message tells it to consolidate and call done, and on the last allowed step the output…

    Final recovery attempt after the failure cap: at 5 consecutive failures the next step is forced to a done-only schema with a message to report what was found, and only one…

  • Compaction

    The summary is fed back as <previous_compacted_memory> on the next compaction and rendered to the model inside <compacted_memory> with an inline comment telling it to treat the…

  • Concurrency

    DOM tree build and screenshot capture start as two parallel tasks in one browser-state request, the screenshot waits only for the time left in a shared budget, and a failed…

  • Config and routing

    Each auxiliary model task has its own plain setting: page_extraction_llm, judge_llm and fallback_llm are Agent constructor arguments, and compaction_llm is a field of…

  • Message model

    Three fixed message slots (system, one state message, per-step context messages) in one canonical OpenAI-shaped type; tool results are text-folded into the next state message, so…

See how browser-use does on each topic

The topic-by-topic verdicts, the code lines and a side-by-side comparison are in the app. Sign in with Google; new accounts are approved by hand.

Independent analysis, not affiliated with browser-use.