# CodeSeeker

An MCP server that fuses keyword search, vector embeddings, directory-level summaries, and knowledge-graph expansion into one retrieval tool, giving Claude Code, Copilot, and Cursor a structural understanding of a codebase beyond plain grep.

CodeSeeker gives an AI coding assistant a four-layer hybrid search over a codebase, keyword matching, vector embeddings, RAPTOR-style directory summaries, and graph expansion, fused into a single MCP tool, so a query like finding the authentication logic returns code by meaning rather than only files containing the literal word. It works with Claude Code, GitHub Copilot, Cursor, Windsurf, and Claude Desktop, and a companion Claude Code plugin keeps the index synced automatically after the initial one-time indexing command.

It is one of three sibling MCP servers from the same team, alongside a tiered-memory server and a coding-standards server, that can be installed independently or together. The project frames itself explicitly as a stopgap: as specification-driven development, better agent navigation, and larger effective context reduce how much an assistant needs to search a codebase to understand it, the project expects its own usefulness to shrink over time.

Website: https://github.com/jghiringhelli/codeseeker