AgentBrowse Sandbox

Use this skill when the task is browser-only and does not require MagicPay payment/session/secret flows:

• open or attach to a browser;
• inspect the visible page state with observe/act loops;
• navigate directly when the destination URL is known;
• extract structured data from the current page;
• capture screenshots for evidence or debugging;
• close the browser cleanly when the task is done.

Do not use this skill for protected credential or payment-card completion. If the page reaches a login, identity, or payment form that needs approved stored secrets, hand off to the magicpay or magicpay-agent skill instead.

Prerequisites

• agentbrowse CLI must be available on PATH. If missing, install with npm i -g @mercuryo-ai/agentbrowse-cli.
• Goal-driven observe and all extract usage require MagicPay gateway configuration. Before using those commands, either:
  • run agentbrowse init <apiKey> to save the key in ~/.magicpay/config.json, or
  • set MAGICPAY_API_KEY and, when needed, MAGICPAY_API_URL.
• The environment must allow browser launch or provide a CDP endpoint for agentbrowse attach.

Workflow

1. Confirm agentbrowse is available by running agentbrowse --version. If the command is not found, run npm i -g @mercuryo-ai/agentbrowse-cli, then verify with agentbrowse --version again.
2. If the task needs goal-focused observe or any extract, make sure MagicPay gateway access is configured with agentbrowse init <apiKey> or the corresponding environment variables before continuing.
3. Establish browser context with agentbrowse launch or agentbrowse attach <cdp-url>. If the correct browser and page context already exist, continue without restarting.
4. Drive the visible-state loop: agentbrowse observe -> decide next action -> agentbrowse act -> agentbrowse observe.
5. Use agentbrowse navigate <url> when the destination URL is already known.
6. Use agentbrowse extract '<schema-json>' [scopeRef] when structured data is needed from the page.
7. Use agentbrowse screenshot [--path <file>] when you need a captured image or diagnostic artifact.
8. Use agentbrowse browser-status only for diagnostics and recovery.
9. Use agentbrowse close only for browser teardown, not as a substitute for confirming task success.

Ask-User Boundary

Ask the user only when:

• the task reaches a protected login, identity, or payment boundary that this skill must not cross;
• the correct next action is genuinely ambiguous after re-observing the page;
• the environment does not allow browser startup and no attachable browser is available.

Safety

• Do not claim progress unless the visible page state confirms it.
• Re-observe after meaningful page changes instead of reusing stale target refs.
• Do not improvise protected-value entry or approval logic inside this skill.

References

Load additional detail only when needed:

• references/workflow.md for resume and recovery guidance;
• references/commands.md for the browser-only CLI reference;
• references/statuses.md for common browser/runtime failure categories;
• references/guardrails.md for surface boundaries and escalation rules.