Updated on August 17, 2026
Browser Harness is an open-source harness from Browser Use that wires a language model straight into your Chrome browser through the DevTools protocol. When a helper is missing mid-task, the agent writes it in Python, so the harness gets sharper with every run. The project is free on GitHub and has passed 16,000 stars. It plugs into coding agents like Claude Code, Codex CLI or Cursor in a couple of minutes.
The idea fits in one line. A single websocket connects the language model to Chrome, with no framework in between. You ask your agent to submit a form, the file-upload step has no ready-made helper, so the agent writes one in its workspace and finishes the job. That agent_helpers.py file grows run after run, while the harness source folder stays locked (so the agent cannot saw off the branch it sits on).
There is more under the hood. Domain skills, switched off by default, teach the agent site-specific shortcuts it can reuse later. A recording feature can also condense hours of agent work into short videos with subtitles and honest mistake-and-fix notes, triggered by simply asking the agent to show what it did.
Browser Harness drives the Chrome you already use, cookies and open sessions included. That is the big split with classic automation scripts, which break the moment a site moves a button or reshuffles its markup. Here the agent looks at the new page and adapts, a different logic from most agent automation tools.
Three setups coexist, depending on whether you value personal sessions or raw scale.
| Mode | What you get |
|---|---|
| Local logged-in Chrome | Work on your accounts, no API key required |
| Isolated Chrome with remote debugging | Automation kept away from your personal profile |
| Browser Use Cloud | Parallel browsers, proxies, CAPTCHA handling |
Installation happens through a short text pasted into your coding agent, which installs the package with uv on Python 3.12, registers the skill and connects to your browser. One checkbox in chrome://inspect grants access, and you are off. Running locally costs nothing.
For dozens of parallel browsers, Browser Use Cloud starts free, then bills usage at roughly $0.02 per browser hour and $5 per GB of proxy data. Those numbers shift over time, so treat the Browser Use pricing page as the only source of truth.
Yes, Browser Harness is completely free and open source. It runs on your local Chrome with no account or API key. Only the optional Browser Use Cloud, handy for parallel browsers with proxies and CAPTCHA handling, moves to usage-based billing once you go past its free tier.
Caution is warranted, as with any autonomous agent. The project narrows the risk, recordings are off by default, the harness source is write-protected, and Chrome asks for manual permission before any connection. For sensitive work, launch an isolated Chrome with a dedicated profile and review the action log afterwards.
Two opposite philosophies from the same team. The browser-use library wraps Chrome in tens of thousands of lines of deterministic heuristics, element extractors, iframe and download handling. Browser Harness bets the other way, a minimal bridge where the model keeps full freedom and writes whatever is missing itself.
Not for repeatable test suites, where a deterministic script stays faster and cheaper. For one-off tasks on changing, protected or login-gated sites, the harness pulls ahead, because the agent corrects itself when the page evolves instead of failing on a stale selector.
Verdict: Sixteen thousand GitHub stars back the minimal-harness bet. Developers already living inside a coding agent will get a driven browser that improves with every task, with Browser Use Cloud ready the day they want scale.
Manus AI Agent 
OpenClaw 
Hermes Agent 
