🪴 Brain

Browser Use

3 min read

Browser Use

🚀 Description

browser-use/browser-use — open-source Python framework that lets an LLM drive a real browser. Instead of feeding screenshots, it surfaces a structured DOM the agent can read and act on. Pair with any LLM (Browser Use Cloud, Anthropic, Google) and either run open source locally or use the fully-hosted cloud agent for stealth, proxy rotation, and 1000+ integrations.

The umbrella project for the Browser Harness (thin CDP harness, self-healing) and Video Use (same “give the LLM a structured surface, not raw frames” pattern, applied to video).

Version status (checked 2026-07-16)

Latest pip package: 0.13.4 — versioning still 0.x, so there is no v4. The headline release is Browser Use CLI 3.0 (2026-07-01), now powered by Browser Harness: the agent executes arbitrary Python in the browser instead of a fixed action menu (click/type), so it can inspect, adapt, and recover in the same coding loop it was trained on. Install shifted to uv tool install browser-use. Core deps now pin browser-harness==0.1.5 and browser-use-sdk==3.4.2. This is the note’s original “Browser Harness is the engine” thesis landing in the product.

🧩 Features

  • Python ≥3.11 (CLI 3.0 wants 3.12); uv add browser-use for the SDK, uv tool install browser-use for the CLI; one-call Agent(task=..., llm=..., browser=...).run()
  • Multi-LLM: ChatBrowserUse, ChatGoogle('gemini-3-flash-preview'), ChatAnthropic('claude-sonnet-4-6')
  • CLI 3.0 (2026-07-01): agent writes and runs Python in the browser via Browser Harness — no fixed action list; inspect / adapt / recover in one coding loop. Drop it into Claude Code, Codex, or any coding agent. (Legacy fixed-action commands open/state/click <idx> predate 3.0.)
  • Templates: uvx browser-use init --template default|advanced|tools
  • Cloud free tier: 3 concurrent browsers, captcha solving, proxies, no card
  • LLM Quickstart: point any coding agent (Cursor, Claude Code) at Agents.md — no manual onboarding
  • Benchmark suite open at browser-use/benchmark — 100 real-world tasks

🎨 Why it matters

The “DOM as context” insight maps directly to how Video Use reads transcripts instead of frames. Same author, same philosophy: structured surface > pixel dump. Aligns with Spec-driven SEO and GEO thinking — give the model the schema, not the rendered noise.

Reasoning for

For tasks where a coding agent needs to actually click around: form filling, scraping with login, account onboarding flows, e-commerce purchases. The hosted Cloud Agent is the easy path for production; OSS for custom-tool integration. See related Hermes Agent and Paperclip for orchestration around such agents.

Alternatives considered

  • Playwright/Puppeteer directly — lower-level, no agent reasoning layer
  • Browser Harness — same org, but a thinner CDP layer where the agent writes its own helpers
  • Hosted Cloud Agent — strongest stealth + scale, paid

🔐 Security note

litellm was dropped from core dependencies after the 2026-03-24 supply-chain attack (backdoored litellm 1.82.7 & 1.82.8 published to PyPI). Provider SDKs (anthropic, openai, google-genai, groq) are now pinned directly instead of routed through litellm. Pin your browser-use version and audit transitive deps before running any agent that touches credentials or a logged-in session (as in Running Browser Use on Windows via Edge CDP).

🔗 Links

📖 Further reading

Template: tool

Graph View

Backlinks