FATSEARCH

FATSEARCH (command fats) is an open-source CLI and stdio MCP server for AI agents that reads pages an agent's built-in fetch can't handle well, such as JS-rendered pages and PDFs. It returns a short outline that the agent opens one section at a time, with pages cached locally as clean Markdown. The latest GitHub release listed is v0.2.5, published 2026-11-06.

What It Is

FATSEARCH is a web-reading tool for agents, described by its makers as the ripgrep of web searching. A single binary acts as a command-line tool and, via fats mcp, as a local MCP server exposing one tool, fats(cmd). Users run fats <url> to fetch, clean and outline a page, then read a section (fats u1#3), filter within it, list or follow links, run a web search that returns ids, search GitHub repositories, or read local PDF, HTML, Markdown, CSV and JSON files.

How the fetch ladder works

Every page starts with a free plain HTTP request from the user's own connection. If the page is a JS shell, captcha or wall, FATSEARCH escalates to a provider the user has configured: Oxylabs, Bright Data or Decodo, cheapest first and hedged so a slow provider doesn't stall the agent. Without a provider key it reads only what plain HTTP returns, and web search needs an Oxylabs or Bright Data key. A dry:true modifier shows what would run and its estimated cost, and fats stats lists each fetch's provider, time, cost and tokens.

Output and safeguards

• Menus, banners, navigation and footers are stripped, while prices, dates and numbers are kept; a loss guard counts anything dropped.
• Pages are cached on disk for up to an hour.
• fats setup opens a one-time local form on 127.0.0.1 so API keys stay out of the chat, the model context and agent logs. Keys are saved to auth.json with mode 0600.
• The project says it shares anonymous usage stats, never URLs, searches, page text or keys, and that this can be turned off.

Benchmarks and limits

The vendor reports that on one Wikipedia page FATSEARCH put 1,384 tokens into context versus 112,517 for Firecrawl, and that it answered 35 of 38 hard pages within 60 seconds versus 24 for Firecrawl and 19 for Tavily. It also notes Firecrawl and Tavily are faster on the pages they do read. Documented gaps: no OCR for scanned PDFs, no DOCX/XLSX, no semantic matching (needs a local model), MCP over stdio only, and no pages behind a login.

Setup path

Install via a curl script, npm, Homebrew or Cargo, optionally add a provider key with fats setup, then register with an agent host such as Claude Code, Codex or Cursor. An agent skill is published for the Agent Skills standard.