jCodeMunch

jCodeMunch is an MCP (Model Context Protocol) server built by J. Gravelle (jMunch LLC) that replaces file-based code reading with surgical symbol-level retrieval, reducing AI token consumption by an average of 96.5% against a grep-and-read agent baseline. It indexes a codebase once using tree-sitter AST parsing and then lets AI agents fetch exact functions, classes, methods, and constants by name — returning tens of lines instead of thousands. The tool is free for personal use and requires a commercial license for for-profit use.

What It Is

jCodeMunch is a token-efficiency layer for AI coding agents that operate over large codebases. Instead of having an agent open and scan entire source files, jCodeMunch builds a persistent local symbol index and exposes MCP tools like search_symbols, get_symbol_source, find_importers, and get_blast_radius. When an agent needs to understand how dependency injection works in a 214K-token codebase, jCodeMunch returns the relevant function bodies in roughly 480 tokens rather than requiring the agent to read every file. It works with any MCP-compatible client, including Claude Code, Cursor, Windsurf, VS Code, Codex CLI, Continue, and Gemini CLI.

How Symbol Retrieval Works

The architecture follows a three-step pattern:

• Index once: Run index_code_folder(path) to parse all source files with tree-sitter and store structured symbol metadata (signature, kind, qualified name, byte offsets, summaries) in a local index at ~/.code-index/.
• Query by intent: The AI calls search_symbols(query) or get_symbol(id). The server performs semantic and keyword search against the index in milliseconds using a bundled ONNX local model (all-MiniLM-L6-v2, 384-dim) for zero-config semantic search.
• Surgical retrieval: Only the matching symbol's source and metadata is returned — not the surrounding file, not unrelated classes. A 6,000-token file read becomes a ~400-token symbol pull.

Symbol summaries fall through a 4-tier chain — Anthropic → Gemini → OpenAI-compatible (Ollama, LM Studio, OpenRouter) → signature-only fallback — so it works with whatever LLM is available, or none. The index format has gone through 9 generations and remains fully backward-compatible.

Benchmark Evidence

The vendor publishes reproducible benchmarks measured with tiktoken cl100k_base across three public repositories pinned to upstream commits (expressjs/express, fastapi/fastapi, gin-gonic/gin), run on v1.108.316. Against a grep-top-3 baseline (the approach a competent agent without the tool actually uses):

• expressjs/express: 15.5x fewer tokens
• fastapi/fastapi: 38.4x fewer tokens
• gin-gonic/gin: 20.3x fewer tokens
• Grand total (15 task-runs): 28.3x fewer tokens (96.5% reduction)

Per-query results range from 7.6x to 81.2x (median 26.1x). An independent 50-iteration A/B test on a Vue 3 + Firebase production codebase showed an 80% vs 72% success rate and a 32% vs 40% timeout rate in favor of jCodeMunch. The vendor also reports live telemetry showing over 1.1 trillion tokens saved across 227,000+ reporting installs since March 2026, though these figures are vendor-published and self-reported.

Structural Capabilities Beyond Search

jCodeMunch exposes over 90 MCP tools. Several answer questions that grep-based approaches cannot:

• get_blast_radius — shows what breaks if a symbol changes
• find_importers — traces who imports a given symbol across the import graph
• get_call_hierarchy — maps call chains through exact function bodies
• find_dead_code — surfaces unreachable symbols
• check_edit_safe / check_delete_safe — preflight risky changes with terminal verdicts
• get_changed_symbols — diffs symbol-level changes between commits
• search_ast — anti-pattern sweeps across the AST

The tool also ships Claude Code hooks: PreToolUse intercepts large file reads and routes to symbol retrieval; PostToolUse auto-reindexes edited files; SubagentStart injects a condensed repo orientation into every spawned subagent.

Companion Tools and Starter Packs

jCodeMunch ships alongside two companion MCP servers: jDocMunch for surgical retrieval of Markdown and documentation sections, and jDataMunch for querying CSV, Excel, Parquet, and JSONL datasets without reading entire files. All three conform to the vendor-published jMRI v1.0 (Munch Retrieval Interface), an Apache-2.0 open retrieval spec.

Pre-built starter packs provide symbol indexes for popular frameworks (Node.js, FastAPI, Django, Flask, React, LangChain, Anthropic SDK, MCP SDK, Laravel, Spring Boot) without requiring a local clone. The Node.js pack (76,700+ symbols) is free. Packs are rebuilt weekly from the latest tagged releases.

Update: v1.108.318 (September 2026)

The latest release is v1.108.318, published 2026-09-11, with the note "the process is code that cannot skip a step, and the field is measured from result files." Recent prior releases include v1.108.317 (2026-09-04, CI harness runs on every change) and v1.108.316 (2026-09-02, benchmark display fix). The project has 2,683 GitHub stars and 366 forks as of the last data snapshot. Language support covers 70+ languages via tree-sitter, including Python, JavaScript/TypeScript, Go, Rust, Java, C/C++, C#, PHP, Ruby, Swift, and Kotlin.