Agent Browser

Select the MyClaw agent that should use browser automation, then click Install. MyClaw adds the skill to that hosted instance; no separate VPS or manual Agent Browser server setup is required.

Open your MyClaw chat and describe the result you want. For example: “Use Agent Browser to open the product page, compare the visible plans, and return a table with plan names, prices, and source links.” Add the target site, success criteria, and any actions the agent must confirm before completing.

• You need an agent browser skill for OpenClaw that can navigate, read page state, and interact across several pages.
• You want an agent browser skill for Claude Code or another terminal agent that produces structured, machine-readable browser output.
• You need deterministic element targeting on dynamic sites where pixel coordinates, XPath, or hand-written CSS selectors are unreliable.
• You want isolated browser sessions, reusable login state, network controls, or repeatable form and testing workflows.

• Your task is primarily visual and depends on interpreting screenshots, charts, canvas content, or pixel-level layout.
• You need to load a browser extension; Agent Browser does not support extension-based workflows.
• You only need a single static HTTP request and do not need JavaScript rendering or page interaction.
• The target site prohibits automation or requires a human-only verification step. Always follow the site's terms and access rules.

Agent Browser turns the current page into an accessibility-tree snapshot and assigns stable @ref IDs to actionable elements. OpenClaw or Claude Code can inspect that compact state and target the correct button, field, or link without relying on screen coordinates.

Click, fill, type, hover, select, check, drag, scroll, and press keys through one agent-friendly CLI. Read text, HTML, attributes, visibility, and enabled state so the agent can verify each action before continuing.

Run named browser sessions with separate cookies and storage. This is useful for admin-versus-user testing, multi-account workflows, and concurrent research tasks that must not share authentication state.

Save cookies and local storage after a successful login, then load that state in later runs. Your agent can resume authenticated workflows without repeating the same sign-in steps every time.

Inspect network activity, block unwanted requests, and mock responses when testing or extracting data. These controls help make browser runs faster, more focused, and easier to reproduce.

Wait for an element, text value, URL pattern, network-idle state, or JavaScript condition before taking the next snapshot. This reduces race conditions in React, Vue, and other single-page applications.

Ask your MyClaw-hosted OpenClaw agent to search the web, follow relevant results, extract structured facts, and return sources. The agent can re-snapshot after each navigation so every action uses current page state.

Use the Agent Browser skill in Claude Code-style terminal workflows to reproduce a bug, inspect a rendered page, complete a form, or validate a user journey while keeping browser actions scriptable and reviewable.

Keep an admin session and a user session open at the same time, each with isolated cookies and saved state. Test permissions, invitations, approvals, and account-specific UI without cross-session leakage.

Save an authenticated session once, then let the agent return to dashboards, fill recurring forms, collect permitted data, or run regression checks without rebuilding the login flow for every task.