hongbietcode/qa-kit/agent-browser

agent-browser Skill

Fast browser automation CLI for AI agents. Chrome/Chromium via CDP with accessibility-tree snapshots and compact @eN element refs (~280 chars/snapshot vs 8K+ for Playwright MCP).

Use agent-browser for browser testing, screenshots, form fills, scraping, exploratory QA, bug hunts, cloud browsers, Electron apps, Slack automation, and flows where a fresh or tool-managed browser is fine. Prefer it over generic browser tools for profile-independent browser work.

Use a browser tool attached to the user's real Chrome profile instead (Claude in Chrome, Chrome DevTools MCP, or the Codex Chrome plugin) when the task needs the user's actual Chrome profile: existing cookies, logged-in sessions, a specific Google account, a tenant/workspace already open in daily Chrome, or deterministic targeting across multiple Chrome profiles.

Install / Upgrade

npm install -g agent-browser     # install (or upgrade) to latest
agent-browser install            # download Chromium (one-time)
agent-browser install --with-deps  # Linux: include system deps
agent-browser upgrade            # self-upgrade the binary
agent-browser --version          # verify

Re-run npm install -g agent-browser (or agent-browser upgrade) periodically — new commands and skills ship with the binary.

Start here — load live workflow content

This file is a discovery stub, not the usage guide. Before running any agent-browser command, load workflow content from the installed CLI:

agent-browser skills get core             # start here: workflows, common patterns, troubleshooting
agent-browser skills get core --full      # include full command reference and templates

The CLI serves skill content from the installed binary, so instructions always match the installed version. This stub stays intentionally small and points at skills get core instead of duplicating command details that can change between releases.

Specialized skills

Load a specialized skill when the task falls outside browser web pages:

agent-browser skills get electron          # Electron apps (VS Code, Slack, Discord, Figma, ...)
agent-browser skills get slack             # Slack workspace automation
agent-browser skills get dogfood           # Exploratory testing / QA / bug hunts
agent-browser skills get vercel-sandbox    # agent-browser inside Vercel Sandbox microVMs
agent-browser skills get agentcore         # AWS Bedrock AgentCore cloud browsers

Run agent-browser skills list to see everything available on the installed version.

When to use

For web-testing from-manual, explore only the supplied case scope and return route/state, account role, locator names/uniqueness, expected-vs-observed behavior and evidence per case. Refresh snapshot handles after page changes; they cannot serve as Playwright locators. The owning workflow decides coverage and writes tests. In its --dry-run mode, do not start or interact with a browser.

Default for browser automation that does not depend on the user's real Chrome login state: autonomous sessions, ad-hoc navigation, screenshots, form fills, scraping, multi-tab work, self-verifying build loops, Electron desktop apps, Slack automation, and Browserbase/cloud browsers.

For low-level Chrome DevTools Protocol diagnostics, use the configured chrome-devtools-mcp bridge or client when one is available. Reason first: if the task does not need a specific real Chrome profile, Chrome DevTools MCP may use its normal navigation tools. If it does need profile/cookie/account state, use the real-profile browser tool: open a new tab for the task and confirm its URL and account before using MCP inspection tools. See references/agent-browser-vs-chrome-devtools.md for the trade-off.

Why agent-browser

• Native Rust CLI rather than a Node.js wrapper
• Works across AI agents such as Claude Code, Codex, Cursor, Continue, and Windsurf
• Chrome/Chromium via CDP without a Playwright or Puppeteer dependency
• Accessibility-tree snapshots with stable element refs for reliable interaction
• Sessions, authentication vault, state persistence, and video recording
• Specialized workflows for Electron apps, Slack, exploratory testing, and cloud browsers

Cloud browsers

For CI/CD or environments without a local browser:

export BROWSERBASE_API_KEY="..."
export BROWSERBASE_PROJECT_ID="..."
agent-browser -p browserbase open https://example.com

See references/browserbase-cloud-setup.md for detailed setup. For AWS Bedrock AgentCore or Vercel Sandbox, run agent-browser skills get agentcore / agent-browser skills get vercel-sandbox.

Observability dashboard

Agent Browser exposes an observability dashboard independently of browser sessions on port 4848. It can also be opened through a proxied or forwarded URL such as https://dashboard.agent-browser.localhost. Stay on the dashboard origin; session tabs, status, and stream traffic are proxied internally, so individual session ports do not need to be exposed.

Troubleshooting

Resources

• Upstream: https://github.com/vercel-labs/agent-browser
• Browserbase: https://docs.browserbase.com/