Playwright

Playwright Skill (Sidecar)

Browser automation via the browsercontainer Docker sidecar running chrome-devtools-mcp (Google, 40+ tools) over MCP SSE. Chrome Beta 149+ with hardware-accelerated WebGPU and WebGL via NVIDIA Vulkan/ANGLE on the RTX 6000. VNC monitoring on port 5903.

Connection

All browser automation routes through the external sidecar. No local browser is used.

Claude Code (MCP Client)
  |  MCP Protocol (SSE)
  v
server.js SSE bridge (browsercontainer:8931)
  |  spawns chrome-devtools-mcp per session
  v
Chrome Beta 149+ (WebGPU, Vulkan/ANGLE)
  |  Xvfb :2
  v
x11vnc → VNC port 5903

MCP Registration

Auto-registered at boot by entrypoint-unified.sh as browser-gpu. Manual registration:

claude mcp add browser-gpu --transport sse http://browsercontainer:8931/sse

Or in .mcp.json:

{
  "mcpServers": {
    "browser-gpu": {
      "url": "http://browsercontainer:8931/sse"
    }
  }
}

Quick Verification

# Health check
curl -s http://browsercontainer:8931/health

# List CDP tabs
curl -s http://browsercontainer:9223/json/list | jq '.[].url'

# VNC into Chrome desktop
vncviewer localhost:5903

When To Use / Not To Use

Use this skill when:

• You need full browser interaction (click, type, navigate, fill forms)
• WebGPU/WebGL rendering validation (Three.js, R3F, Babylon.js)
• Visual regression testing with screenshots
• Multi-tab workflows
• Debugging web apps running on the host or inside Docker

Consider alternatives:

• browser -- lightweight AI-optimised snapshots, smaller context, faster
• qe-browser -- QE-grade typed assertions, visual-diff baselines, injection scanning
• chrome-cdp -- raw CDP protocol access for performance profiling and low-level debugging
• web-summary / gemini-url-context -- summarising page content without interaction
• curl / httpx -- API testing without a real browser

Available Tools (40+)

Key Tools

browser_snapshot (preferred over screenshots for LLM interaction) Returns an accessibility tree of the page -- structured, deterministic, no vision model needed.

browser_take_screenshot Capture viewport or full-page screenshot as PNG/JPEG. Returned as image content.

browser_evaluate Execute JavaScript in the page context.

WebGPU

The sidecar runs Chrome Beta 149+ with full WebGPU support via NVIDIA Vulkan/ANGLE:

• --enable-features=Vulkan,VulkanFromANGLE,DefaultANGLEVulkan,UseSkiaRenderer,SharedArrayBuffer,WebGPU
• --enable-unsafe-webgpu
• GPU: Quadro RTX 6000 (24GB VRAM, Turing architecture)

WebGPU requires a secure context. localhost and 127.0.0.1 are secure by default. For other HTTP origins, add them to the TREAT_AS_SECURE env var in docker-compose.browsercontainer.yml. about:blank and data: URIs do NOT have WebGPU access — always navigate to a real HTTP/HTTPS URL first.

Network

The sidecar is on visionclaw_network and discoverable by hostname browsercontainer.

Visual Access

vncviewer localhost:5903   # no password, Display :2

Legacy note

There is no local Playwright install. Older Node scripts that require('playwright') and chromium.launch() a local browser are retired and kept only as historical examples under references/legacy-local-scripts/. Drive all new automation through the sidecar MCP tools above.