| name | browser-harness |
| description | Direct browser control via CDP. Use when the user wants to automate, scrape, test, or interact with web pages through their real Chrome (not the claude-in-chrome MCP). Invokes the `browser-harness` CLI which pipes Python with helpers preloaded into a daemon connected to the user's running browser. Triggers on "browser-harness", "use the harness", "drive my chrome", "scrape this site", "automate this page", or any browser-automation task where compositor-level coordinate clicks, raw CDP, or self-healing helpers are wanted. |
browser-harness
Self-healing CDP browser harness. The repo lives at ~/dev/browser-harness/ and the CLI browser-harness is on $PATH (installed via uv tool install -e .).
Before doing anything
- Read
~/dev/browser-harness/SKILL.md in full โ that is the canonical usage doc and must be in context before calling the CLI.
- Read
~/dev/browser-harness/helpers.py โ that is where the functions live; the agent edits these.
- If this is a first-time setup, attach is failing, or daemon is stale, also read
~/dev/browser-harness/install.md.
- Search
~/dev/browser-harness/domain-skills/ and ~/dev/browser-harness/interaction-skills/ before inventing approaches.
Invocation shape
browser-harness <<'PY'
new_tab("https://example.com")
wait_for_load()
print(page_info())
PY
- First navigation is
new_tab(url) โ goto(url) clobbers the user's active tab.
- Helpers are pre-imported, daemon auto-starts, no
cd or uv run needed.
- Verify visible actions with
screenshot().
Contribute back
If a non-obvious site behavior is discovered (private API, stable selector, framework quirk, URL shortcut, wait, trap), open a PR adding to ~/dev/browser-harness/domain-skills/<site>/. The harness improves only because agents file what they learn.
