A token-efficient MCP server for AI agents that leverages tree-sitter AST parsing for symbol-level precise code retrieval across 70+ languages, reducing code-reading token consumption by 95%+, with 20+ structural analysis tools covering blast radius, call hierarchy, dead code detection, architectural topology, and more.
jCodeMunch MCP is a code understanding and retrieval server designed for AI agents, compliant with the Model Context Protocol (MCP) standard. Its core architectural principle — "retrieval precision is more efficient than brute-force context expansion" — drives the use of tree-sitter to parse 70+ programming languages into ASTs, extracting functions, classes, methods, constants, types, and other symbols with byte-offset indexing for precise on-demand code extraction, eliminating brute-force full-file reads.
On token efficiency, official benchmarks (express, fastapi, gin repositories, 15 task runs) show total token consumption dropping from 1,865,210 to 92,515 — a 95%+ reduction. The MUNCH compact encoding mechanism achieves median 45.5% byte savings (peak 55.4%) through path prefix compression and dictionary list packing.
The project provides 20+ structural analysis tools covering blast radius analysis with depth-weighted risk scoring (get_blast_radius), N-layer call graph traversal (get_call_hierarchy), inheritance chain traversal (get_class_hierarchy), dead code detection from entry points (find_dead_code), cyclic dependency detection (get_dependency_cycles), logical module topology discovery fusing imports, shared references, and git co-changes (get_tectonic_map), external signal propagation path tracing from HTTP/CLI/cron/events to leaf nodes (get_signal_chains), cross-language AST pattern matching with 10 preset anti-pattern detectors and custom mini-DSL (search_ast), symbol git archaeology with full commit history and semantic evolution narratives (get_symbol_provenance), comprehensive PR risk scoring 0.0–1.0 (get_pr_risk_profile), edit-ready refactoring plans for rename/move/extract/signature changes (plan_refactoring), and more. These capabilities break through traditional LLM tool limitations in structural code analysis.
For security and privacy, the project automatically sanitizes AWS/GCP/Azure/JWT/GitHub keys before they reach LLM context, enforces trusted folder access control, and uses local-first storage (default ~/.code-index/) keeping data on-machine. The context provider framework auto-enriches symbols with dbt model metadata and git author/change metadata, with support for custom providers.
Compatible with Claude Code, Claude Desktop, Cursor, Windsurf, Continue, OpenClaw and other major MCP clients, the project offers an interactive init command for one-shot configuration, watch mode for automatic re-indexing, PreToolUse/PostToolUse/PreCompact execution hooks, and complexity-driven model routing. Installable via pip install jcodemunch-mcp, with Docker and uvx deployment options. Optional dependency groups cover Anthropic, Gemini, OpenAI, semantic search, local embeddings, HTTP/SSE transport, dbt, and Groq integrations.
Note: The official website http://jcodemunch.com/ referenced in the README uses HTTP and its current availability could not be verified. Commercial licensing is dual-mode: free for non-commercial use; paid for commercial use (Builder $79 / Studio $349 / Platform $1,999).