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):

Approach	What's sent to the LLM	Est. tokens
Read all source files	22 files verbatim	~42,000
Read 2 relevant files	server.rs + extractor.rs	~15,000
munchrs tools	repo outline + file tree + 2 file outlines + symbol search	~1,900

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

Tool	Description
index_folder	Index a local folder containing source code
list_repos	List all indexed repositories
get_file_tree	Get the file tree of an indexed repository
get_file_outline	Get all symbols in a file with signatures
get_symbol	Get full source code of a specific symbol
get_symbols	Get source code of multiple symbols in one call
search_symbols	Search for symbols matching a query
search_text	Full-text search across indexed file contents
get_repo_outline	High-level overview of a repository
invalidate_cache	Delete index, forcing re-index

Supported Languages

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

Environment Variables

Variable	Description
MUNCHRS_LOG_LEVEL	Log level (default: info)
MUNCHRS_LOG_FILE	Path to log file (logs to stderr if unset)
CODE_INDEX_PATH	Custom index storage directory (default: ~/.munchrs/)
OPENROUTER_API_KEY	API key for AI-powered symbol summarization (optional)

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