Sub-150ms Web-to-Markdown Engine & Model Context Protocol (MCP) Server
Engineered for LLM context window efficiency, RAG pipelines, and autonomous AI coding agents.
When autonomous agents or RAG pipelines scrape the web using Puppeteer, Jina, or Firecrawl, they dump 25,000+ tokens of navigation trees, cookie popups, tracking scripts, and inline SVGs directly into the prompt.
- High Cost: $0.08+ per page in Claude 3.5 Sonnet token fees.
- Slow Latency: 4 to 8 seconds per scrape.
- Context Poisoning: LLMs hallucinate on irrelevant footer links and cookie consent disclaimers.
TokenMarkdown compresses web pages by 95% into dense, clean Markdown in sub-150ms, preserving your context budget and boosting reasoning accuracy.
| Feature | Raw HTML / Scrapers | Firecrawl / Jina | TokenMarkdown.com |
|---|---|---|---|
| Average Latency | 6,200ms | 2,100ms | ⚡ 148ms |
| Tokens Consumed | ~28,500 tokens | ~3,200 tokens | 🎯 ~940 tokens (95% Savings) |
| Prompt Cost (Sonnet) | $0.0855 / call | $0.0096 / call | 💰 $0.0028 / call |
| Zero-SDK URL Proxy | ❌ No | ❌ No | ✅ Yes (curl tokenmarkdown.com/https://...) |
| Native MCP (Stdio + SSE) | ❌ No | Partial | ✅ Yes (npx -y tokenmarkdown-mcp) |
You don't even need an API key for quick testing. Prepend https://tokenmarkdown.com/ to any URL:
curl -s "https://tokenmarkdown.com/https://stripe.com/docs/api"Add to your Claude Desktop configuration (claude_desktop_config.json) or Cursor MCP settings.
🎁 Free Sandbox Bonus: Pass
--email you@company.cominargsto instantly double your free sandbox quota from 50 to 100 requests! Zero credit card required.
{
"mcpServers": {
"tokenmarkdown": {
"command": "npx",
"args": ["-y", "@quitegoodproject/tokenmarkdown-mcp", "--email", "you@company.com"]
}
}
}Or run via CLI:
npx -y @quitegoodproject/tokenmarkdown-mcp --email you@company.comcurl -X POST "https://tokenmarkdown.com/v1/extract" \
-H "Authorization: Bearer <TOKENMARKDOWN_API_KEY>" \
-H "Content-Type: application/json" \
-d '{
"url": "https://linear.app/blog/rethinking-issue-tracking",
"include_images": false
}'{
"url": "https://linear.app/blog/rethinking-issue-tracking",
"title": "Rethinking issue tracking — Linear Blog",
"markdown": "# Rethinking issue tracking\n\nMost issue trackers are built as databases...",
"tokens": 820,
"savings_percent": 96.2,
"execution_ms": 142
}- Clay.com Blueprint: Drop-in HTTP enrichment column to cut Clay AI credits by 90%.
- n8n Workflow Blueprint: Instant webhook extractor node for automated research agents.
TokenMarkdown is the client-side engine for agents scraping the web. For website owners who want their own sites to be natively ingestible by AI agents without scraping latency:
- RFC-2602 Specification: Open Inbound AI Ingestion Protocol (AIP) for context-ceiling preservation and deterministic content negotiation (
Accept: text/markdown). - Empirical SaaS Truncation Benchmark: Research report analyzing why 84% of top SaaS applications are truncated at 25,000 tokens by Perplexity and ChatGPT Search.
- AnswerRail (Reference Edge Gateway): Deploy RFC-2602 at your Cloudflare Anycast edge in <60s to serve sub-15ms clean Markdown, byte-0 schema tables, and Ed25519 provenance with zero origin code changes.
TokenMarkdown is an official developer primitive engineered by The Quite Good Project.
- Parent Specification: RFC-2601 Developer Primitives Spec & RFC-2602 AI Ingestion Protocol
- Reference Inbound Rail: AnswerRail.com
- Companion Tool: TokenEnrich.com (Sub-140ms ~180-Token Firmographics)
MIT © 2026 The Quite Good Project.