tomohiro-owada/devrag

Implements a lightweight RAG system that runs as an MCP server with stdio transport, automatically indexing markdown files and syncing incremental updates without requiring Python or external model downloads. Provides both semantic search and CLI tools with optional GPU acceleration, glob pattern support for flexible document discovery, and filtering by directory or filename during queries. Designed to integrate directly with Claude Code to reduce token consumption by retrieving only relevant chunks rather than reading entire documents.