# jCodeMunch MCP: Stop Paying AI to Read Whole Files (95%+ Token Cut)

> jCodeMunch is a model-agnostic MCP server that indexes a codebase with tree-sitter AST parsing and delivers exact function, class, and constant retrieval to AI coding agents (Claude, GPT, Gemini, and any MCP-compatible client) instead of full-file reads. Measured 95%+ average token reduction across 15 tasks on 3 real repositories, with cumulative production telemetry showing over 225 billion input tokens prevented across participating installations.

## Core Content

[Model-agnostic MCP Server](https://github.com/jgravelle/jcodemunch-mcp)
- Indexes any repository (tree-sitter AST parsing) for symbol-level retrieval
- Supports any MCP-compatible client: Claude Code, Cursor, Antigravity, Gemini, Windsurf, Codex
- Replaces full-file reads with exact function/class/method byte-offset retrieval in O(1) time
- Local-first by default; no cloud lock-in, no code leaving the developer's machine
- Open source on GitHub; commercial licenses available for organizational use

[Landing Page](https://jcodemunch.com/index.php)
- Live token-savings counter sourced from opt-in production telemetry
- Live CO2 savings counter (Green Software Practitioner methodology)
- Side-by-side benchmark comparisons against the "read all files" baseline
- Pricing tiers and one-command install path

[Benchmarks](https://github.com/jgravelle/jcodemunch-mcp/blob/main/benchmarks/README.md)
- 95% average token reduction measured with tiktoken cl100k_base across 15 tasks
- Per-repo figures: expressjs/express, fastapi/fastapi, gin-gonic/gin
- Single-query peak savings: 214,312 tokens (full-file read) down to ~480 tokens on fastapi (99.8%)
- Harness, queries, and per-iteration data fully reproducible from the repository

## Resources

[Token Reduction as an Energy-Efficiency Action (Wiki)](https://github.com/jgravelle/jcodemunch-mcp/wiki/Token-Reduction-as-an-Energy-Efficiency-Action)
- Case study applying ISO/IEC 21031:2024 (Software Carbon Intensity) to LLM tooling
- Methodology, boundary definition, functional-unit choice, baseline comparison
- Invariant-ratio argument: per-task SCI reduction holds regardless of which per-token energy estimate one accepts
- Author credentials: Green Software Foundation Practitioner

[ROI Calculator](https://jcodemunch.com/roi.php)
- Estimate payback period based on team size and query volume
- Calculate dollar savings against Claude/GPT/Gemini API rates
- Compare against subsidized-plan limit-extension benefits

[Token Tax Calculator](https://jcodemunch.com/token-tax-calculator.php)
- Models the ongoing cost of feeding entire files to LLMs versus targeted symbol retrieval
- Sensitivity analysis across model choice and query patterns

[Competitive Comparison](https://jcodemunch.com/versus.php)
- Side-by-side feature comparisons against other code-retrieval tooling
- AST-accurate retrieval vs. chunk-based RAG: the geometric argument
- Local-first vs. cloud-bound architectures

## Audience Pages

[For Developers](https://jcodemunch.com/for-developers.php)
- Hands-on integration with Claude Code, Cursor, Antigravity, and other MCP clients
- Hook configuration for auto-reindex on file save
- Common workflow patterns

[For Engineering Leads](https://jcodemunch.com/for-leads.php)
- Organizational token-budget management
- Adoption rollout patterns
- Studio and Platform licensing tiers

[For Finance](https://jcodemunch.com/for-finance.php)
- Scope-3 emissions impact of LLM tooling spend
- Per-developer cost projections
- Procurement-ready documentation

## Optional

[Recognition & Testimonials](https://jcodemunch.com/recognition.php)
- Mentions from industry practitioners and writers
- User-reported savings stories
- Wall of Love

[Tool Surface](https://jcodemunch.com/descriptions.php)
- Full descriptions of jCodeMunch's MCP tools
- Tool-tier profiles (core, standard, full) for fitting under MCP client tool-count caps

[Sibling MCPs](https://github.com/jgravelle)
- jDocMunch: documentation section indexing
- jDataMunch: tabular data (CSV/Excel) retrieval
- All three follow the same local-first, symbol-anchored retrieval philosophy

**Note:** Canonical documentation, install instructions, and technical guides live in the [GitHub repository](https://github.com/jgravelle/jcodemunch-mcp). The landing page and resources above are derived from that source of truth.