rangefind-mcp

rangefind-mcp

An MCP server over rangefind static search indexes. Point it at any index — a local directory or an http(s) URL; single, generational, or sharded — and agents get search as tools. No search server behind it: the index is plain files, read with HTTP range requests and cached on disk.

Run

# open mode: tools accept any index path or URL
npx rangefind-mcp

# configured mode: named indexes only
npx rangefind-mcp --index docs=./public/rangefind --index osm=https://cdn.example.com/rangefind/

Claude Code:

claude mcp add rangefind -- npx rangefind-mcp --index docs=./public/rangefind

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "rangefind": {
      "command": "npx",
      "args": ["rangefind-mcp", "--index", "docs=/path/to/rangefind"]
    }
  }
}

Tools

Tool	Purpose
rangefind_search	BM25F full-text search with typo correction, facets, filters, sorting, geo (radius / bounding box / nearest-first / distance boost), shard scoping, pagination
rangefind_suggest	search-as-you-type autocomplete
rangefind_count	exact match counts without fetching results
rangefind_info	index metadata: totals, build time, provenance/license, features, shard list with groups and coverage bboxes
rangefind_list_indexes	the configured index names (configured mode only)

All tools are read-only and return structured content. Errors are actionable tool results (unknown index → the available names; unknown shard → the available shards/groups), not protocol failures.

Programmatic use

import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { createRangefindMcpServer } from "rangefind-mcp";

const server = createRangefindMcpServer({
  indexes: { docs: "./public/rangefind" },
  open: false
});
await server.connect(new StdioServerTransport());

Why this is interesting for agents

A rangefind index is a self-describing corpus: rangefind_info exposes what it contains (including data provenance and, for sharded planet-scale indexes, the geographic shard tree), and every query — text, geo, faceted — runs against static files with no keys, no rate limits, and no per-query cost. Offline corpora work too: point --index at a downloaded directory.