About this project

jCodeMunch MCP is a Model Context Protocol (MCP) server designed to reduce AI token consumption during code exploration. It uses tree-sitter to parse source code into a local index of symbols (functions, classes, methods, constants) with byte-level offsets, allowing agents to retrieve only the exact code they need instead of reading entire files. Key capabilities include: - Symbol-level retrieval: `get_symbol_source` returns exact function bodies, `search_symbols` finds symbols by name, `get_file_outline` provides structural summaries. - Structural queries: `find_importers`, `get_blast_radius`, `get_call_hierarchy`, `find_dead_code`, `get_changed_symbols`, `get_hotspots`, and AST-based anti-pattern searches. - Task context assembly: `assemble_task_context` classifies task intent and runs the appropriate tool sequence under a token budget. - Safety checks: `check_edit_safe`, `check_delete_safe`, `get_pr_risk_profile`, and `plan_refactoring` with edit-ready blocks, plus integration with type checker diagnostics (mypy, pyright, tsc, ruff). - Confidence scores, freshness flags, coverage contracts, secret redaction, and SCIP-based compiler-verified references. - Automatic index freshness via watch modes, agent hooks, and a VS Code extension. Benchmarks claim a 28.3x reduction in tokens versus a grep-and-read baseline (96.5% savings) across three public repositories, with per-repo savings ranging from 15.6x to 39.7x. An independent A/B test on a production Vue 3 codebase reported 15-25% tool-layer savings and improved success rates. Installation is via `uv tool install jcodemunch-mcp` or `uvx jcodemunch-mcp`, with an `init` command that auto-configures MCP clients and installs a CLAUDE.md prompt policy. Supports 70+ languages via tree-sitter, monorepos, and incremental indexing. Local-first design with indexes stored in `~/.code-index/`; optional anonymous savings counter (opt-out). Licensing is dual-use: free for personal use, commercial licenses required for revenue-generating use.