browser-use-integration

# Browser-Use Plugin Integration ## Overview The `browser-use` plugin provides headless Chromium screenshot capability as a fallback when the claude-in-chrome extension is unavailable. When installed alongside `designer`, it lets `designer:ui` screenshot its own artboards and lets the commands capture a page the user names, without requiring the user's Chrome instance to be open. This skill is **optional** — all designer plugin functionality works without browser-use installed. browser-use is the fallback for Tier 2 of the three-tier browser capture chain: 1. claude-in-chrome (preferred — real Chrome, CSS snapshot, JS execution) 2. browser-use (fallback — headless Chromium, screenshot only) 3. Error with install options `designer:review` takes image files only and never captures; the capture happens before it is dispatched, in the command or in `designer:ui`. ## When to Use Use this skill when: - `designer:ui` has written artboards and needs `screens/*.png` for the judge - A command was given a page URL as a reference or implementation and must capture it - claude-in-chrome extension is NOT installed or unavailable - User needs automated screenshot capture without opening Chrome manually ## Detection Pattern Probe the MCP tools directly — no subprocess, no plugin listing. A tool that answers is the only presence check this skill makes: ``` chrome_available = try mcp__claude-in-chrome__tabs_context_mcp() → success: use Tier 1 (claude-in-chrome) → error: chrome not available browser_use_available = try mcp__plugin_browser-use_browser-use__browser_list_sessions() → success: use Tier 2 (browser-use) → error: neither available → Tier 3 (error) ``` The tool names carry the plugin namespace: `mcp__plugin_browser-use_browser-use__<tool>`. An agent that wants them must list them in its `tools:` line by that full name. ## Screenshot Capture Pattern When browser-use is available (Tier 2), use this sequence. It needs `browser_save_screenshot`, the browser-use tool that writes a PNG file; if the tool is missing, the installed browser-use predates it: say so, and name `claude plugin update browser-use@magus`. **A local HTML artboard must be served, not opened.** browser-use reports "Navigated to" for a `file://` URL but stays on `about:blank`, so the screenshot is blank. Serve the artboard directory on loopback for the capture: ``` Step 1: Serve local files (skip for an http(s) TARGET_URL) Bash (background): python3 -m http.server <port> --bind 127.0.0.1 --directory "<artboard dir>" TARGET_URL = http://127.0.0.1:<port>/<artboard>.html Step 2: Navigate, then record the session id mcp__plugin_browser-use_browser-use__browser_navigate(url=TARGET_URL) mcp__plugin_browser-use_browser-use__browser_list_sessions() → the new entry is SESSION_ID Step 3: Write the PNG mcp__plugin_browser-use_browser-use__browser_save_screenshot(output_path="${OUTPUT_DIR}/<name>.png", full_page=false) → {"path": "…", "size_bytes": N, "width": W, "height": H} `output_path` must be absolute; the tool refuses a relative one. An "Error:" line means no file was written: close the session, then stop with that error. Step 4: Close (ALWAYS — even on error) mcp__plugin_browser-use_browser-use__browser_close_session(session_id=SESSION_ID) Stop the Step 1 server. ``` ## Limitations vs claude-in-chrome | Capability | claude-in-chrome | browser-use | |------------|-----------------|-------------| | Screenshot capture | Yes | Yes | | Full-page screenshot | No (viewport only) | Yes (full_page=True) | | CSS snapshot (computed styles) | Yes (javascript_tool) | No | | JavaScript execution | Yes (javascript_tool) | Limited (retry_with_browser_use_agent) | | User's real Chrome session | Yes | No (separate headless instance) | | Authenticated sessions | Yes (user's cookies) | No (fresh session) | | Real DOM state | Yes | DOM index map only | | Animation disable (before capture) | Yes (CSS injection) | No | ## User Notification Template If NEITHER browser capture method is available, stop with this exact message: ``` ERROR: No browser capture method available. Options: 1. Install Claude-in-Chrome extension (preferred — full CSS snapshot support) → https://github.com/anthropics/claude-in-chrome 2. Enable browser-use plugin (headless screenshot fallback) → /plugin install browser-use@magus 3. Provide an image file reference instead → Re-run with REFERENCE_SOURCE=/path/to/screenshot.png ``` If browser-use IS available but claude-in-chrome is not, log this info message (do not stop): ``` INFO: Using browser-use for headless screenshot capture (Tier 2 fallback). ```