Show HN: PageIndex MCP – Chat with Long PDFs on Claude or Cursor (github.com)

PageIndex MCP is a new Claude Desktop extension that lets you chat with long PDFs without hitting token or chunking limits. It supports both local and online PDFs, offers a free 1,000-page allowance and unlimited conversations, and installs via a one‑click .mcpb bundle (double‑click to add to Claude Desktop). The desktop extension handles OAuth automatically; you can run a local MCP server (requires Node.js ≥18) with a simple npx pageindex-mcp command, connect to PageIndex’s hosted MCP endpoint, or use mcp-remote as a bridge for clients that don’t support HTTP MCP servers. Note: local server enables local file uploads, while the hosted endpoint only processes URLs. Technically, PageIndex uses a “vectorless” retrieval‑augmented generation (RAG) approach based on multi‑step reasoning and tree search rather than similarity vectors and top‑k chunking. That design promises higher relevance (retrieval beyond raw similarity), transparent, traceable reasoning paths, preserved document context (no chunking), and lower infrastructure overhead (no vector DB). For developers this means easier installation/configuration, more human‑like document navigation, and reproducible retrieval trajectories—appealing for workflows that need accuracy and auditability when querying long, complex PDFs. The project is MIT‑licensed.