Browser Harness lets the agent write its own helpers
The browser-use project wires language models to a real Chrome over CDP, and has the agent write any helper function it finds missing along the way.
Illustrative image: a workbench at night with an open laptop and a toolbox that still has one empty slot.
Browser Harness attaches a language model to a locally running browser through the Chrome DevTools Protocol and lets the agent write the helper functions it lacks mid-task, saving them into agent-workspace/agent_helpers.py.
At a glance
- Connects to a local browser over a Chrome DevTools Protocol (CDP) websocket.
- Agent-written helpers go to agent-workspace/agent_helpers.py; the core under src/browser_harness/ is off limits to the agent.
- Installed with uv on Python 3.12; additional clients attach through an MCP server.
- As of August 31, 2026 the project page lists about 17,300 stars, 1,700 forks, 36 open issues and 568 commits.
- License: MIT. Every figure above is self-reported by the project.
Browser Harness, published by the browser-use group, connects a language model to a locally running browser through a Chrome DevTools Protocol websocket. The project page bills it as a harness that lets language models finish tasks in the browser. The connection is the ordinary part; what happens when the agent runs into a gap is not.
A protected core and a writable workspace
The repository splits into two zones. The core sits under src/browser_harness/ and is meant to stay untouched by the agent, while a separate workspace is open for it to edit.
When the model needs a helper function that does not exist, the project says it writes one and saves it to agent-workspace/agent_helpers.py. The next run picks it up. That is the self-healing part of the tagline: the toolkit accumulates instead of having to be complete up front.
Setup and clients
Installation runs through uv on Python 3.12, driven by a setup prompt the project supplies for Claude Code or Codex. The browser link is made at chrome://inspect/#remote-debugging, and an MCP server lets several clients attach; the page names Claude Code, Devin and Cursor. For running many browsers in parallel it points to Browser Use Cloud.
Numbers, all self-reported
As of August 31, 2026 the project page shows about 17,300 stars, 1,700 forks, 36 open issues and 568 commits on the main branch, under an MIT license. These are counters on the repository itself, not an independent measurement.
What could not be checked
Exactly one source was available for this article: the project page. Nothing here is corroborated by a second, independent newsroom.
Two questions stay open in particular: how dependable self-written helpers remain across many tasks, and whether the protected core stays protected in practice. The description we read cites no benchmark figures and no comparison against other browser agents. Answering either would require independent testing.
FAQ
What does Browser Harness do?
It is an open-source harness from browser-use that links a language model to a real browser over the Chrome DevTools Protocol so the model can carry out tasks in that browser.
What makes Browser Harness self-healing?
According to the project, when the agent is missing a helper function it writes one itself and stores it in agent-workspace/agent_helpers.py, so the helper is available on later runs.
Which coding agents can connect to Browser Harness?
The project page names Claude Code, Devin and Cursor through its MCP server, with the setup prompt aimed at Claude Code or Codex.
