jgravelle/jcodemunch-mcp

What it solves

AI agents often waste tokens by reading entire files to find a specific piece of code, which increases costs and consumes the context window. jCodeMunch provides a token-efficient way for agents to retrieve only the exact code they need (such as functions, classes, or methods) with byte-level precision, reducing token usage by an average of 96% compared to traditional grep-and-read methods.

How it works

The tool indexes a codebase once using tree-sitter AST parsing to store structured symbol metadata—including signatures, kinds, and byte offsets—alongside raw file content in a local index. Instead of re-reading whole files, agents can query this index to fetch only the specific implementations or structural information they require.

Who it’s for

Developers using AI agents (like Claude Code, Cursor, Windsurf, or Continue) who work with large repositories or unfamiliar codebases and want to reduce AI spend and improve agent reliability during code exploration and refactoring.

Highlights

• Extreme Token Efficiency: Benchmarked at 28.3x fewer tokens than grep-and-read agents.
• Structural Queries: Supports advanced queries that grep cannot answer, such as finding importers, calculating blast radius, and identifying dead code.
• Broad Language Support: Supports over 70 languages via tree-sitter.
• Local-first Design: Indexes are stored locally at ~/.code-index/ to ensure privacy and security.
• MCP Compatible: Works as a Model Context Protocol server for a wide range of AI IDEs and CLI tools.