Browser Harness

Overview

Browser Harness is a thin, editable browser-control harness from Browser Use. It connects an LLM-powered coding agent to Chrome over CDP, then gives the agent a browser-harness command plus a SKILL.md workflow for screenshots, coordinate clicks, raw CDP calls, helper edits, cloud browsers, and optional domain-specific skills.

Use it when a coding agent needs to operate a real browser session rather than only call a hosted scraping API or deterministic Playwright script. It is most relevant for Claude Code, Codex, and similar agents that can read the included skill instructions and then call the harness from a terminal.

Install

The upstream install guide recommends cloning the repository into a durable location and installing the checkout as an editable uv tool:

git clone https://github.com/browser-use/browser-harness
cd browser-harness
uv tool install -e .
command -v browser-harness

Then register the repository's SKILL.md with the agent host. The install guide documents a Codex path under $CODEX_HOME/skills/browser-harness/SKILL.md and a Claude Code path through an import in ~/.claude/CLAUDE.md.

The repository has a pyproject.toml for the browser-harness package, version 0.1.0, with Python 3.11+ and dependencies on cdp-use, fetch-use, pillow, and websockets. At verification time, the PyPI browser-harness name was a reserved placeholder rather than the documented install path, so prefer the repo checkout and editable uv install unless upstream publishes a package.

Browser Connection Modes

Agent Capabilities

Use Cases

• Attach Codex or Claude Code to a local Chrome session for supervised web tasks.
• Give an agent a cloud browser when local Chrome is unavailable or parallel sub-agents need isolated browser sessions.
• Build reusable browser automation playbooks for sites such as GitHub, LinkedIn, Amazon, Shopify Admin, Gmail, Salesforce, Reddit, PubMed, SEC EDGAR, and other domains represented in the domain-skill tree.
• Let an agent use screenshots and coordinate clicks when selectors are brittle or cross-origin iframes make normal DOM automation awkward.
• Add task-specific helpers in agent-workspace/agent_helpers.py while keeping the protected core package small.
• Compare Browser Harness with Browser Use, Playwright MCP, Hyperbrowser, Browserbase, Stagehand, Chrome DevTools MCP, and computer-use agents.

Source Review

Verified on 2026-06-18:

• The upstream repository is browser-use/browser-harness, MIT-licensed, and GitHub metadata showed more than 15,000 stars.
• The README describes Browser Harness as a thin, editable CDP harness that connects an LLM directly to a real browser.
• The README includes a setup prompt for Claude Code or Codex and points agents to install.md.
• The install guide documents editable installation with git clone and uv tool install -e ..
• The install guide documents Codex and Claude Code skill registration paths for the repository's SKILL.md.
• The install guide documents local Chrome remote debugging, isolated Chrome profiles, Browser Use Cloud browsers, BROWSER_USE_API_KEY, BU_CDP_URL, BU_CDP_WS, BU_NAME, and cloud browser shutdown behavior.
• SKILL.md documents the browser-harness heredoc workflow, new_tab usage, screenshot-first operation, coordinate clicks, HTTP fetches, raw CDP calls, remote browsers, and optional domain skills.
• AGENTS.md states that agents operating the harness edit inside agent-workspace/, including agent_helpers.py and domain-skills/.
• The repository includes many domain-skill examples and interaction skills for browser mechanics such as downloads, uploads, dialogs, tabs, iframes, shadow DOM, profile sync, screenshots, and cookies.
• pyproject.toml declares package name browser-harness, version 0.1.0, Python >=3.11, and a browser-harness console script.
• The GitHub releases endpoint did not expose a latest release at verification time, so this entry does not claim a tagged release.

Safety and Privacy

Browser Harness is powerful because it lets an agent act inside the same browser surfaces a human uses. That is also the risk. Treat a real Chrome profile as account access, not as a disposable test browser. For sensitive accounts, start with an isolated Chrome profile or a dedicated cloud browser, then add explicit approval checkpoints for purchases, messages, profile changes, file uploads, form submissions, and destructive actions.

The domain-skill model is useful but public by default when contributed back to the repository. Keep generated skills generic, avoid screenshots or private data, and do not encode secrets, customer records, private internal routes, or credential-bearing examples.

Duplicate Check

Checked current content/tools/, content/mcp/, content/skills/, guides, README entries, open pull requests, and repository-wide content for Browser Harness, browser-use/browser-harness, Browser Use Cloud browser harness, CDP browser agent, browser-harness, and Browser Harness domain skills. The directory already has a Browser Use tools entry and several browser automation MCP entries, but no dedicated Browser Harness content entry, exact source URL duplicate, target file, or open duplicate PR was found.

Disclosure

Editorial listing. No paid placement or affiliate link is used. Browser Harness is MIT-licensed open-source software, while Browser Use Cloud, proxies, cloud profiles, captcha-related services mentioned by upstream, API keys, and remote browser sessions may have separate billing, terms, privacy controls, and usage limits.