SchemaSnap

HTML → typed tables, forms, lists as JSON. 30ms, no key, no browser.

Born from shadow-web's SchemaSnap module. A parse primitive — not a scraper, not a crawler. Feed it HTML, get structured data.

pip install "schemasnap[server]"
uvicorn schemasnap.server:app --port 8000 &

curl -X POST http://127.0.0.1:8000/v1/compress \
  -H "Content-Type: application/json" \
  -d '{"html":"<table><thead><tr><th>Product</th><th>Price</th><th>Stock</th></tr></thead><tbody><tr><td>Widget</td><td>$19.99</td><td>150</td></tr></tbody></table>"}'

{"tables":[{"columns":["Product","Price","Stock"],
  "types":["string","currency","integer"],
  "rows":[["Widget","$19.99","150"]]}]}

Self-hosted API

There is no hosted endpoint — the Cloud Run deployment was retired in October 2026. Run it yourself; it is two commands:

pip install "schemasnap[server]"
uvicorn schemasnap.server:app --port 8000

Quick install

pip install schemasnap

# For MCP (Cursor, Claude Desktop)
pip install "schemasnap[mcp]"

# For URL extraction (requires Playwright)
pip install "schemasnap[browser]"
playwright install chromium

Usage

Python SDK

from schemasnap import extract_from_html, parse_page

# From HTML string (no browser, 30ms)
result = extract_from_html(open("page.html").read())
print(result.tables)      # [{columns, types, rows, total_rows}]
print(result.forms)       # [{action, method, fields}]
print(result.lists)       # [{type, items, total}]
print(result.token_budget) # {raw_bytes, compressed_bytes, compression_ratio}

MCP (Cursor, Claude Desktop)

pip install "schemasnap[mcp]"

Add to Cursor Settings → MCP:

{
  "mcpServers": {
    "schemasnap": {
      "command": "schemasnap-mcp"
    }
  }
}

Tool: compress_html(html, max_rows_per_table=50, url="") → typed tables, forms, lists, actions.

curl (no Python needed)

curl -X POST http://127.0.0.1:8000/v1/compress \
  -H "Content-Type: application/json" \
  -d '{"html": "<table>...your html...</table>"}'

Benchmarks

Measured July 2026 on a 1 vCPU / 256 MB instance (Cloud Run, us-central1), Python 3.12. Cold start ~250 ms — irrelevant when you call the library in-process.

Features

• Tables → columns, types (string|integer|number|currency|percentage|date|email|url), rows — from <table>, ARIA role=table|grid, Tailwind grids, or .row pseudo-tables
• Forms → action, method, fields with name, type, label, required, placeholder, pattern, minlength, select options
• Lists → type (unordered|ordered|select), items, total
• Semantic groups → "Login Form", "Navigation", "Checkout", "Search" auto-detected
• Deterministic — same input → same output. No LLM guess, $0 per call

Competitive positioning

SchemaSnap is a parse primitive, not a scraper:

Firecrawl / Jina     =  URL → content (crawl + render)
Crawl4AI             =  URL → markdown (self-host crawl)
pandas read_html()   =  HTML → strings (dev tool)
SchemaSnap           =  HTML → typed structure (agent utility)

Composable: Firecrawl fetch → SchemaSnap parse. Or curl → SchemaSnap parse.

Honest limits: HTML only (no JS render), layout tables skipped, no crawl, no proxy rotation.

Relationship to shadow-web

SchemaSnap started as schema_snap.py inside shadow-web — a broader agent-facing web interaction suite. This standalone version:

• Adds MCP server (compress_html tool) for Cursor / Claude Desktop
• Inlines semantic grouping (no external dependency)
• Fixes tree mutation bug during tag removal
• Adds deterministic type inference (8 types)
• Hardened for public API (SSRF guard, rate limiting, oversize protection)

License

MIT

#web-scraping #data-extraction #ai-agents #html-parser #json-api #python #llm-tools