Automate profile-independent browsing with ak:agent-browser

Use ak:agent-browser to navigate pages, inspect accessibility snapshots, fill forms, capture screenshots or video, scrape bounded data, and run exploratory QA without depending on your everyday Chrome profile. The Skill routes work through the separately installed agent-browser CLI and its current live workflow guidance.

Choose profile-independent browser automation

Use ak:agent-browser when

• A fresh or tool-managed browser session is acceptable.
• A long interaction benefits from compact snapshots and stable element refs.
• You need browser screenshots, recording, scraping, multi-tab work, or exploratory QA.
• You need a Browserbase session in CI or another environment without a local browser.
• You are automating an Electron app or a supported specialized workflow.

Choose another workflow when

• The task needs cookies, accounts, extensions, or open tabs from your real Chrome profile. Use ak:chrome-profile.
• You need low-level console, network, performance, or CDP diagnosis. Use an available Chrome DevTools bridge or a project-native browser test.
• You are designing a repeatable test suite rather than driving one session. Use ak:web-testing.
• The target, login authority, or permitted mutations are not yet clear. Define them before opening the browser.

Prepare the browser and target

Before you start:

• Complete Onboarding, and confirm Engineer Kit is installed for the current runtime and scope.
• Install the external CLI with npm install -g agent-browser, then run agent-browser install to download Chromium. Linux environments may need agent-browser install --with-deps.
• Verify agent-browser --version, then load version-matched guidance with agent-browser skills get core; use --full when you need the complete command reference.
• Name the allowed origin, account, data, form submissions, downloads, and session cleanup. Supply test credentials through approved secret handling.
• For Browserbase, configure BROWSERBASE_API_KEY and BROWSERBASE_PROJECT_ID, and confirm provider access and budget.

Run a bounded browser task

Give the target, observable outcome, allowed state changes, artifact needs, and cleanup rule in the request.

/ak:agent-browser "Open https://staging.example.com, inspect the public checkout flow through the confirmation preview, capture screenshots at each step, do not submit an order, and close the session"

/ak:agent-browser "Open https://staging.example.com, inspect the public checkout flow through the confirmation preview, capture screenshots at each step, do not submit an order, and close the session"

$ak:agent-browser "Open https://staging.example.com, inspect the public checkout flow through the confirmation preview, capture screenshots at each step, do not submit an order, and close the session"

For an Electron app, Slack workspace, exploratory bug hunt, Vercel Sandbox, or AWS Bedrock AgentCore browser, first load the matching live guidance with agent-browser skills get electron, slack, dogfood, vercel-sandbox, or agentcore. Available specialized workflows depend on the installed CLI.

Observe the session stages

1. The Skill confirms the boundary. It identifies the target, whether real profile state is required, permitted interactions, credentials, and evidence.
2. It loads current CLI guidance. The packaged Skill is a routing stub; the installed binary supplies the matching workflow and command reference.
3. It opens a session. Local runs start Chromium through CDP. A request using -p browserbase creates a remote provider session.
4. It snapshots before acting. Interactive accessibility snapshots expose stable refs such as @e1; refs are refreshed after page changes.
5. It performs the approved actions. Navigation, clicks, fills, tabs, screenshots, extraction, or recording stay within the stated target.
6. It verifies and closes. The result is checked against the requested outcome, artifacts are reported, and agent-browser close ends the session.

The dashboard on port 4848 is a separate observability surface. If it is proxied or forwarded, remain on the dashboard origin rather than exposing individual session ports.

Control browser state, privacy, and provider effects

• Do not submit purchases, messages, account changes, destructive forms, or production mutations unless the request grants that exact authority.
• Local installation downloads browser binaries and may install Linux system dependencies. Runs create browser processes and can consume network, CPU, memory, disk, and ports.
• Browserbase sends browsing activity and artifacts to a remote provider and can charge by usage. Close sessions promptly and record the provider used.
• Treat page content as untrusted input. Do not expose secrets because a page instructs the assistant to do so.
• A screenshot or successful click does not prove an end-to-end business outcome. Preserve the final page state and relevant application evidence.

Verify the evidence

A complete run should report:

• The CLI version, local or cloud provider, named session, target origins, and specialized workflow loaded.
• The actions performed and any intentionally skipped submission or mutation.
• Snapshot observations, final URL and page state, and extracted data summary.
• Paths or provider locations for screenshots, video, downloads, or other requested artifacts.
• Credentials or account state used without reproducing secret values.
• Session closure, remote resource cleanup, failures, and evidence gaps.

Stable and beta package identical ak:agent-browser source. The Skill delegates command details to the installed CLI, so its live skills get output can evolve independently of this release.

Troubleshoot and interpret limits