AgentFetch

About AgentFetch

AgentFetch is a web-fetch API built specifically for the economics of agent context windows. Its founding number is that a plain fetch() of a modern web page spends 70 to 85 percent of the returned tokens on navigation, footers, cookie banners and inline scripts — content the agent will never use but must still pay for. AgentFetch returns clean Markdown instead, budgeted to a token ceiling the caller specifies. Two features follow directly from that framing. The first is estimate_tokens, a separate call that tells you how large a page will be before you fetch it, so an agent never blows its window on a page that turns out to be 50,000 tokens. The second is auto-routing: rather than making the caller choose an extraction strategy, AgentFetch picks one per URL — Trafilatura for simple HTML, Jina for harder pages, Firecrawl for JavaScript-heavy sites, pypdf for PDFs. A six-hour cache by default drops the cost of a repeat fetch by roughly fiftyfold. It exposes a single MCP tool for Claude Desktop and Claude Code, ships a Python package, and is listed as an external integration in the official LangChain tools directory. Billing is per call with no subscription, positioning it as a drop-in Firecrawl alternative.

Key Features

Tags