Skills
Automate isolated browser work with ak:agent-browser
Use ak:agent-browser for snapshots, interactions, screenshots, scraping, and exploratory QA without relying on your everyday Chrome profile.
Use ak:agent-browser when the outcome requires a browser or supported desktop
app that the assistant can drive through the agent-browser CLI. The Skill
routes profile-independent navigation, accessibility snapshots, element
interaction, screenshots, scraping, exploratory QA, and cloud-browser work to
the workflow content shipped by the installed CLI.
Choose ak:agent-browser for isolated sessions
Use ak:agent-browser when
- A fresh or tool-managed browser session is acceptable.
- You need compact snapshots and stable element references for a longer flow.
- The task involves screenshots, form preparation, scraping, exploratory QA, or a supported Electron or cloud-browser workflow.
- You can verify the result without depending on cookies or accounts already open in your everyday Chrome profile.
Choose another workflow when
- The task requires your real Chrome cookies, signed-in account, tenant, or a specific existing profile. Use the profile-aware browser workflow available in your installation.
- You need a repeatable test suite rather than an exploratory run. Use a web testing workflow and the project's test runner.
- You only need low-level Chrome DevTools diagnostics. Use an available DevTools integration and keep its selected profile explicit.
Prepare the browser environment
Before you start:
- Complete Onboarding and install Marketing Kit for the runtime and scope you are using.
- Install or update
agent-browserwithnpm install -g agent-browser, then runagent-browser installonce to download Chromium. Linux environments may needagent-browser install --with-deps. - Run
agent-browser skills get corebefore browser commands so the workflow matches the installed CLI. Use--fullwhen you need its command reference. - Decide whether the session is local or cloud-hosted. Browserbase requires
BROWSERBASE_API_KEYandBROWSERBASE_PROJECT_ID; keep them out of prompts, logs, screenshots, and version control. - Define the allowed origins, account boundary, permitted form actions, and whether any submission or upload is authorized.
| Runtime | Invocation | Availability boundary |
|---|---|---|
| Claude Code | /ak:agent-browser ... | Native delivery is the default; explicit plugin delivery is also supported. The Skill can call the installed CLI within Claude Code's permissions. |
| Cursor | /ak:agent-browser ... | Slash invocation is user-verified. Cursor projection does not establish identical Hook, tool, or approval behavior. |
| Codex | $ak:agent-browser ... | The Skill uses native Codex discovery. CLI execution still depends on the current sandbox, network, and command permissions. |
See Runtime adapters for delivery and component-level differences.
Run a bounded browser task
/ak:agent-browser Open https://example.com in an isolated session, capture an interactive snapshot and a screenshot, report the page title, and do not submit forms or sign in/ak:agent-browser Open https://example.com in an isolated session, capture an interactive snapshot and a screenshot, report the page title, and do not submit forms or sign in$ak:agent-browser Open https://example.com in an isolated session, capture an interactive snapshot and a screenshot, report the page title, and do not submit forms or sign inFollow the session stages
- Confirm the state boundary. Decide whether an isolated browser is enough. If existing Chrome state is required, stop and switch workflows.
- Load installed guidance. The Skill reads the live
coreworkflow, or a specializedelectron,slack,dogfood,vercel-sandbox, oragentcoreworkflow when the task needs it. - Open and observe. Start the session, navigate to the approved URL, take a current snapshot, and identify elements from that evidence.
- Act narrowly. Click, fill, upload, or navigate only within the approved task. Refresh the snapshot after page changes instead of reusing stale refs.
- Verify and close. Capture the requested evidence, report unresolved browser state, and close the session. Closing a Browserbase session also ends the remote session.
Keep account and external effects explicit
A browser action can be a real-world action
Opening and inspecting a page is different from signing in, submitting a form, uploading data, sending a message, accepting terms, purchasing, or changing an account. Approve those mutations separately before they occur.
Installing the CLI or Chromium changes the local environment and may require network access. A cloud-browser run sends traffic and session data to its provider and may incur provider usage. Use local Chromium when appropriate, close remote sessions promptly, and never place provider keys in command text that will be retained.
Verify the evidence
A complete run should identify:
- The session type and approved origin or app.
- The snapshots, screenshots, extracted data, or QA findings produced.
- Each mutation performed, including any form submission, upload, or account change.
- The final observed state and whether the session was closed.
- Any provider, permission, authentication, or page-state limitation.
Troubleshoot and understand limits
| Symptom | Safe next step |
|---|---|
| The runtime cannot find the Skill | Confirm target and scope, restart the runtime, then use Runtime cannot find a Skill or Agent. |
agent-browser is missing | Install it with npm, verify agent-browser --version, then load agent-browser skills get core. |
| Chromium or Linux libraries are missing | Run agent-browser install or agent-browser install --with-deps. |
| An element ref no longer works | Take a new interactive snapshot after the page update. |
| The site needs an existing Chrome login | Stop. An isolated session is the wrong state boundary; switch to the profile-aware workflow. |
| Browserbase authentication or quota fails | Preserve the provider error without exposing keys, close any open session, and check provider access outside the Skill. |
The Skill depends on the installed CLI, available browser, current site behavior, network access, and provider limits. It does not guarantee that a site permits automation or that a flow is safe to submit. Continue with Marketing Hooks for background guardrails or Marketing Kit lifecycle for installation recovery.
Route complex work with ak:agentkit
Use ak:agentkit to classify ambiguous or multi-step work, select installed capabilities, build the shortest safe chain, and time verification.
Inspect and generate media with ak:ai-multimodal
Use the pinned Multix CLI through ak:ai-multimodal for media analysis, extraction, transcription, conversion, and provider-backed generation.