Browser automation framework for AI agents. A Claude Code plugin that teaches agents how to navigate, map, extract, and authenticate on the web using Playwright MCP.
manucraft gives your Claude Code agent reliable browser superpowers — from clicking buttons to scraping entire sites — using proven patterns from production web automation.
- Install the Playwright plugin for Claude Code (if not already installed)
- Install manucraft:
claude plugin add herman181920/manucraft
- Try it:
/browse https://example.com
manucraft provides four skills that auto-activate when your agent encounters browser tasks:
The fundamental primitive: navigate -> snapshot -> interact -> screenshot
Every browser task follows this loop. Snapshot the DOM to "see" the page structure, use ref labels to interact with elements, take screenshots for visual evidence.
"Map first, download second."
Crawl a website's structure before interacting with it. Builds a reusable JSON sitemap with page types, link relationships, and pagination detection. Prevents missed pages and infinite loops.
Extract structured data into JSON, CSV, or Markdown.
Handles tables, product listings, article content, directories, and more. Works with sitemaps for multi-page extraction. Includes quality checks (item count, null fields, dedup).
Manage browser authentication flows.
Login form detection, credential entry, session persistence via Chrome profiles, MFA handling, and session expiry detection. Never stores credentials in project files.
| Command | Description |
|---|---|
/browse <url> |
Navigate to a URL and interact |
/map-site <url> |
Crawl and map a site's structure |
/extract-data <url> |
Extract structured data from a page |
/login <url> |
Authenticate to a website |
auth-handling (if login needed)
|
v
core-navigation <-- the primitive everything uses
|
v
site-mapping (optional: map structure first)
|
v
data-extraction (extract from mapped pages)
Example: "Scrape all products from this store"
/loginto authenticate/map-siteto discover all product pages/extract-datato pull structured product data
- Skills are prompt templates (markdown) that teach agents browser automation patterns
- Primary backend: Playwright MCP tools (zero setup, part of Claude Code ecosystem)
- Fallback: Python Playwright with headless Chromium (for Chrome conflicts)
- Hook: PreToolUse detection warns when Chrome may conflict with MCP
Playwright MCP sometimes conflicts with an open Google Chrome browser. Manucraft handles this:
- A PreToolUse hook detects if Chrome is running and warns the agent
- If MCP tools fail, skills guide the agent to use the Python fallback script
- The fallback uses its own headless Chromium — no conflicts
The patterns in manucraft are proven in ILLYSCRAPE, a production web scraping platform. The "map first, download second" philosophy and the navigate-snapshot-interact-screenshot loop come directly from real-world automation experience.
See CONTRIBUTING.md for how to add new skills, test locally, and submit PRs.
MIT