munchrs

A fast, token-efficient MCP server for codebase indexing and symbol retrieval.

munchrs is a Rust binary that indexes codebases using tree-sitter AST parsing and exposes tools over the Model Context Protocol (MCP) via stdio transport. It gives AI assistants structured access to your codebase — symbols, outlines, and full-text search — without dumping entire files into context.

Install

Download a prebuilt binary from Releases, or build from source:

git clone https://github.com/vihu/munchrs.git
cd munchrs
just build
# Binary at target/release/munchrs

Token Savings

Instead of dumping entire files into an LLM context window, munchrs returns only the structure and symbols you need. Measured on the munchrs codebase itself (22 Rust files, 5,309 lines, ~42k tokens):

A typical "understand the codebase and find a function" workflow uses ~95% fewer tokens compared to reading all files, or ~87% fewer compared to reading just the right files.

MCP Configuration

Claude Code

claude mcp add munchrs /path/to/munchrs

With .mcp.json

Add the following to your .mcp.json (user/project scoped):

  "mcpServers": {
    "munchrs": {
      "type": "stdio",
      "command": "/path/to/munchrs",
      "args": [],
      "env": {}
    }
  }

Tools

Supported Languages

Python, JavaScript, TypeScript, Go, Rust, Java, PHP, Dart, C#, C, C++, Swift, Elixir, Erlang

Environment Variables

Recommended CLAUDE.md Instructions

Add the following to your project's CLAUDE.md to teach it how to use munchrs efficiently:

## Codebase Exploration

When a `munchrs` MCP server is available, prefer it over raw file reads for codebase exploration:

- Use `list_repos` to check if the repo is already indexed; if not, use `index_folder` to index it
- Use `get_repo_outline` for a high-level overview of the codebase (directories, languages, symbol counts)
- Use `get_file_tree` to browse the project structure before diving into specific files
- Use `get_file_outline` to see all symbols in a file with their signatures — avoid reading entire files
- Use `search_symbols` to find functions, classes, or types by name across the repo
- Use `get_symbol` to fetch the full source of a specific symbol instead of reading the whole file
- Use `get_symbols` to batch-fetch multiple symbols in one call when you need several
- Use `search_text` as a fallback for grep-like searches (string literals, comments, config values)
- Use `invalidate_cache` only when the index seems stale or corrupt — it forces a full re-index

Credits

• All credit for the original python implementation goes to: https://github.com/jgravelle/jcodemunch-mcp

Differences from jcodemunch-mcp

• Language: Rust (single static binary) vs Python
• Storage: reads from original files on disk — no file copies in the index
• Index layout: nested ~/.munchrs/<owner>/<name>/ vs flat ~/.code-index/<slug>
• Additional languages: Elixir and Erlang support
• Summarization: OpenRouter as the LLM gateway vs direct OpenAI/Gemini APIs
• Output format: compact toon fmt vs JSON with metadata envelopes

License

MIT