semantic-code-mcp

[](https://www.npmjs.com/package/@smallthinkingmachines/semantic-code-mcp) [](https://opensource.org/licenses/MIT)

MCP server for semantic code search using AST-aware chunking and vector embeddings. Works with any AI coding tool that supports MCP.

GitHub | npm

Traditional search tools like grep, ripgrep, and ag match text patterns exactly. When developers ask conceptual questions like "How is authentication handled?" or "Where do we process payments?", these tools require knowing exact function names or code patterns. This leads to:

• Overwhelming results: Thousands of lines containing search terms, most irrelevant
• Naming convention blindness: "authenticateUser", "login", "validateSession", and "handleAuth" are the same concept—grep doesn't know that
• Lost context: Results show isolated lines without surrounding code structure

AI coding tools inherit these limitations. Claude Code relies on grep/ripgrep for code search—no semantic understanding, just string matching. Aider uses repo maps with graph ranking to select relevant code, but still depends on structural analysis rather than meaning. These approaches work on smaller codebases but struggle at scale, burning tokens on irrelevant results or missing conceptually related code.

Semantic search understands code by meaning, not just text. It can answer "How is user authentication implemented?" by understanding conceptual relationships—regardless of function names or file locations.

Using local embeddings and vector search, it bridges the gap between text search limitations and LLM context constraints, providing more accurate results for navigating large