Claude Chrome

Use Claude Code with Chrome browser extension for web browsing and automation tasks. Alternative to OpenClaw's built-in browser tools.

Security Scan

The skill is coherent with its stated purpose (using the Claude Code binary + Chrome extension for browser automation) but its runtime instructions advise bypassing browser permission checks and running arbitrary commands on nodes, which is high-risk and worth caution.

Purpose & Capability

Name/description match the requested artifacts: SKILL.md requires the 'claude' binary and describes using the Claude Code Chrome extension/Chrome on the node. Requiring the claude binary and checking for the native host process is proportionate to the stated goal of driving a local Chrome extension.

Instruction Scope

Instructions tell the agent to run nodes.run to execute arbitrary 'claude' commands on a target node, including using the flag --dangerously-skip-permissions (auto-approve all actions). That flag plus advice to be 'aggressive' with automation gives the agent broad ability to click, fill forms, read page contents, and interact with authenticated sessions — potentially exposing cookies, local storage, or other sensitive data. The instructions do not constrain what pages or data are accessed and encourage bypassing user approvals, which increases risk.

Install Mechanism

No install spec or code files — instruction-only skill. Nothing will be written to disk by the skill itself; lowest install risk.

Credentials

No environment variables, no unrelated credentials requested. Requiring the 'claude' binary and an active Chrome + extension is consistent with the functionality. No unexpected secrets or config paths declared.

Persistence & Privilege

always is false and the skill does not request persistent system-wide changes or access to other skills' configurations. Agent invocation is allowed (default) — combined with the instruction scope concerns this increases blast radius but is not anomalous by itself.

What to consider before installing

This skill is internally consistent with its description, but its runtime guidance to use --dangerously-skip-permissions and run arbitrary 'claude' commands on a node is risky. Before installing, consider: 1) Only run on nodes you fully control and trust; isolate the Chrome profile used for automation (no personal accounts, no saved passwords/cookies you care about). 2) Avoid or audit use of --dangerously-skip-permissions — prefer manual approval of domain permissions when possible. 3) Limit the node's privileges and network access, and do not expose sensitive credentials or services to that node. 4) If you only need scraping or public-page reads, prefer safer tools (e.g., web_fetch) that don't access a full browser profile. 5) If you proceed, monitor activity and be prepared to rotate any secrets that might be exposed. If you want a lower-risk alternative, ask the skill author to remove the recommendation to auto-skip permissions and to provide guarded examples that limit pages/actions and demonstrate safer defaults.

Claude Chrome — Browser Automation via Claude Code

Use Claude Code's Chrome integration to browse the web, interact with pages, and automate browser tasks. This is an alternative to OpenClaw's built-in browser tools (Chrome Relay, OpenClaw profile).

Prerequisites

1. Claude Code installed on the node (e.g. /opt/homebrew/bin/claude)
2. Claude Code Chrome extension installed and enabled in Chrome
3. Chrome running on the node

How It Works

Claude Code can connect to Chrome via its built-in browser extension (MCP server). When enabled, Claude Code gains browser tools — it can navigate pages, click elements, fill forms, read content, and more.

Step 1: Check if Chrome Extension is Active

Look for the native host process to confirm the Chrome extension is running:

nodes.run node=<your-node-id> command='["bash", "-lc", "pgrep -f \"claude --chrome-native-host\""]'

If this returns a PID, the Chrome extension bridge is active and ready.

Step 2: Run Claude Code with Chrome

Use nodes.run with your node to execute browser tasks:

nodes.run node=<your-node-id> commandTimeoutMs=120000 command='["bash", "-lc", "claude --dangerously-skip-permissions --chrome -p \"Go to example.com and read the headline\""]'

Flags:

• --dangerously-skip-permissions — auto-approve all actions (required for automation)
• --chrome — enable Chrome browser integration
• -p / --print — non-interactive print mode (required for automated use)
• bash -lc — login shell to ensure PATH is loaded

Timeout: See benchmarks below for guidance. Recommended defaults:

• Simple tasks (single page read): commandTimeoutMs=30000 (30 seconds)
• Medium complexity (multi-step navigation): commandTimeoutMs=120000 (2 minutes)
• Complex workflows (multiple pages + summarization): commandTimeoutMs=180000 (3 minutes)

Performance Benchmarks

Task Type	Example	Duration	Recommended Timeout
Simple	Read button text on Google	13s	30s (30000ms)
Medium	Wikipedia search + navigate + summarize	76s	2min (120000ms)
Complex	Multi-page navigation + external links	~90s+	3min (180000ms)

Gateway timeout note: OpenClaw's gateway has a hardcoded 10-second connection timeout. Commands will error immediately but continue running in the background. Results arrive via system messages when complete.

Limitations

• Domain permissions: Claude Code's Chrome extension may require user approval for new domains (cannot be automated)
• Gateway timeout: Initial connection times out at 10s, but commands continue running
• Desktop required: Only works on nodes with a desktop environment, Chrome, and the extension active

Tips

• Always use --dangerously-skip-permissions for automated runs
• Always use -p / --print for non-interactive output
• Always use bash -lc for login shell (PATH loading)
• Be aggressive with timeouts - commands complete in background even after gateway timeout
• Claude Code can combine coding and browsing in a single session
• Check the native host process before attempting browser tasks
• For simple data scraping, consider web_fetch instead (faster, no domain permissions needed)