| name | agent-browser |
| description | Automate browser interactions using the agent-browser CLI, and keep its sessions and processes from leaking. Use when navigating pages, clicking elements, filling forms, taking screenshots, extracting page data, running browser tests, verifying UI behavior, or cleaning up leftover browser processes. Use when the user mentions browser, agent-browser, test in browser, navigate, click, screenshot, interact with page, automate, leaked or runaway browsers, Chrome eating memory, or killing processes. |
Agent Browser
Browser automation CLI for AI agents. Installed globally as agent-browser.
Core Workflow
Every browser automation session follows this pattern:
agent-browser open <url> --session <name>
agent-browser snapshot -i --session <name>
agent-browser click @e5 --session <name>
agent-browser eval "document.title" --session <name>
agent-browser close --session <name>
Session Management
Run every command with --session <name>. Use ONE session for the whole run, named after the lane it belongs to, reused across every navigation and interaction — not a session per step or per surface. Close it before reporting results, so the report describes a machine already cleaned up.
agent-browser open https://example.com --session lane-7
agent-browser close --session lane-7
A session left open keeps its Chrome processes and the agent-browser daemon alive for days, and they accumulate across runs until they exhaust the machine's memory. Long before that they poison unrelated test runs with timeouts. There is no cap on how many browser sessions may run at once — verified teardown, not rationing, is what keeps the machine healthy.
