GitHub - firecrawl/firecrawl: The context API to search, scrape, and interact with the web at scale. 🔥
The context API to search, scrape, and interact with the web at scale. 🔥 - firecrawl/firecrawl
AI Agent Tool
Summary
jevcrawl is an MCP server and CLI that uses a lightweight decision model to select links and passages during web crawling, returning only verbatim answer-relevant text to AI agents. It achieves 54% answer-page accuracy (2x Crawl4AI) at roughly 10x lower cost per correct answer compared to LLM-driven crawling.
Highlights
- Uses a small decision model to pick links and passages, so the agent's LLM only reads answer-relevant content at a few cents per crawl
- 54% answer-page accuracy vs 33% for LLM-driven crawling and 2x Crawl4AI, with ~10x lower cost per correct answer
- Prompt injection resistant by design: page text is only scored, never followed as instructions
- Returns verbatim passages with source URLs, heading paths, and character offsets — the decision model cannot invent text
- Integrates via MCP with Claude Code, Claude Desktop, Cursor, or standalone CLI; optional Playwright-based JavaScript rendering
auto-generated
Context
Audience
Developers building AI agents that need to browse websites, particularly those using Claude or other MCP-compatible tools who want to reduce token costs and improve answer retrieval accuracy
DomainAI Agent Infrastructure
Formatopen source tool with documentation
Accessopen source, requires Jev API key from console.typesafe.ai
Source Overview
jevcrawl is an MCP server and CLI that uses a small decision model to select links and passages, returning only verbatim answer-relevant text with source URLs and character offsets.
Benchmarks show 54% answer-page accuracy versus 33% for LLM-driven crawling and 2x Crawl4AI, at roughly 10x lower cost per correct answer. It integrates with Claude Code, Claude Desktop, Cursor, or standalone CLI via a Jev API key, with optional Playwright-based JavaScript rendering.
Topics
Related
Crawl4AIMCP (Model Context Protocol)Claude CodePlaywrightToken efficiency for LLM agentsPrompt injection defense
