open-browser

name	open-browser
description	Browser automation via Playwright MCP (official Microsoft headless browser)
version	1.0.0
metadata	{"tags":["browser","automation","scraping","web","playwright","mcp"],"category":"automation","related_skills":["computer-use","open-code"]}

Open Browser (Playwright MCP Integration)

Automate web interactions using Playwright, the official Microsoft headless browser with native MCP server support. Playwright provides robust browser automation with multi-browser support (Chromium, Firefox, WebKit).

When to Use

• Web scraping and data extraction
• Form filling and submission
• Browser-based testing
• Web automation tasks requiring JavaScript execution
• PDF generation from web pages
• Screenshot capture
• Multi-browser testing

When NOT to Use

• Desktop GUI automation (use computer-use skill)
• Simple HTTP requests (use REST API directly)
• File operations (use open-code skill)

Prerequisites

• Playwright Docker container running (docker-compose.mesh.yml includes playwright-mcp)
• MCP server accessible at http://playwright-mcp:3100 (container) or http://localhost:3100 (host)
• Agent profile with browser_automation capability

Architecture

OpenOrchestrator (plan with browser step)
  → OpenAgents (browser-automation profile)
  → MCP call to playwright-mcp-server
  → Playwright container (headless browser)
  → Results → OpenRec (audit) → OpenBrain (observation)

Playwright MCP Tools

Navigation

browser_navigate(url: string, waitUntil?: "load" | "domcontentloaded" | "networkidle")

Navigate to URL and wait for specified state.

Page Interaction

browser_click(selector: string)

Click element matching CSS selector.

browser_fill(selector: string, value: string)

Fill input field with value (triggers change events).

browser_type(selector: string, text: string)

Type text into input field character by character.

browser_press_key(key: string, selector?: string)

Press keyboard key (optionally targeting specific element).

browser_select_option(selector: string, value: string)

Select dropdown option by value or text.

Page State

browser_snapshot()

Returns: { url, title, text } - current page location, title, and text content.

browser_evaluate(script: string)

Execute JavaScript in page context, returns result.

browser_wait_for(selector: string, timeout?: number)

Wait for element to appear (timeout in seconds).

Advanced Features

browser_screenshot(options?: { fullPage?: boolean, path?: string })

Capture screenshot of current page.

browser_generate_pdf(options?: { path?: string })

Generate PDF from current page (Chromium only).

Debugging

browser_network_requests()

Returns all network requests made by the page.

browser_console_messages()

Returns console logs from the page.

Session Management

browser_close()

Close current session and reset environment.

Procedure

Basic Web Scraping

1. browser_navigate(url="https://example.com", waitUntil="networkidle0")
2. browser_snapshot()  # verify page loaded
3. browser_evaluate(script="document.querySelectorAll('.item').length")  # count items
4. browser_evaluate(script="Array.from(document.querySelectorAll('.item')).map(e => e.textContent)")  # extract data
5. browser_close()

Form Submission

1. browser_navigate(url="https://example.com/form")
2. browser_fill(selector="#name", value="John Doe")
3. browser_fill(selector="#email", value="[e-mail]")
4. browser_select_option(selector="#country", value="US")
5. browser_click(selector="#submit")
6. browser_wait_for(selector=".success", timeout=5)
7. browser_snapshot()  # verify success
8. browser_close()

Anti-Detection

Obscura automatically applies:

• Per-session fingerprint randomization
• Tracker blocking
• Real V8 JavaScript execution
• CDP protocol (indistinguishable from real Chrome)

No additional configuration needed for basic anti-detection.

Decision Rules

Scenario	Action
Need to interact with web page	Use browser_* tools
Need desktop GUI automation	Use computer-use skill
Need to execute code in sandbox	Use open-code skill
Need simple HTTP request	Use REST API directly
Page requires JavaScript	Obscura handles automatically (V8 engine)
Anti-detection required	Obscura applies automatically

Pitfalls

• Don't use browser_evaluate for simple DOM queries - use browser_snapshot + parse
• Don't forget browser_close() - sessions persist until closed
• Don't assume elements exist - use browser_wait_for before interacting
• Don't use pixel coordinates - Obscura uses CSS selectors (more reliable)
• Don't ignore waitUntil parameter - networkidle0 ensures all resources loaded

Verification

• Obscura container running (docker ps | grep obscura)
• MCP server accessible (curl http://localhost:3100/health)
• Agent can call browser_navigate successfully
• Results logged to OpenRec (check correlation_id trace)
• No security bypass (approval required for external requests)

Security

• Browser automation requires browser_automation capability in agent profile
• External requests logged to OpenRec for audit
• Approval gate for sensitive operations (configurable per org)
• No credential storage in Obscura sessions (ephemeral)

Integration with Mesh

// OpenOrchestrator plan step
{
  "goal": "Scrape product data from example.com",
  "required_skills": ["open-browser"],
  "approval_required": false
}

// OpenAgents dispatch
POST /v1/runs
{
  "profile": "browser-automation",
  "goal_id": "...",
  "parameters": {
    "url": "https://example.com/products",
    "selector": ".product-item"
  }
}

// OpenRec audit
{
  "type": "browser.automation.completed",
  "payload": {
    "url": "https://example.com/products",
    "items_scraped": 42,
    "duration_ms": 3500
  }
}

Related Skills

• computer-use - Desktop GUI automation (non-web)
• open-code - Code execution in sandboxed worktrees
• open-mcp-scaffold - Add new MCP tools to mesh