Code Graph RAG

Code Graph RAG MCP Server

Sponsor https://accelerator.slider-ai.ru/

Advanced Multi-Language Code Analysis with Semantic Intelligence

A powerful Model Context Protocol server that creates intelligent graph representations of your codebase with comprehensive semantic analysis capabilities.

🌟 11 Languages Supported | ⚡ 5.5x Faster | 🔍 Semantic Search | 📊 26 MCP Methods

---

🚀 Quick Start

Installation

LATEST VERSION ONLY FROM GitHub Realeases

do not use npm repository any more

# Install globally
npm install -g ./er77-code-graph-rag-mcp-2.7.12.tgz
code-graph-rag-mcp --version

Claude Desktop Integration

# Quick setup (recommended)
npx @modelcontextprotocol/inspector add code-graph-rag \
  --command "npx" \
  --args "@er77/code-graph-rag-mcp /path/to/your/codebase"

or

#
  claude mcp add-json  code-graph-rag ' { 
        "command": "npx",
        "args": ["@er77/code-graph-rag-mcp", "/_work_fodler"],
  "env": {
    "MCP_TIMEOUT": "80000"
  }
      }

Manual setup: Add to Claude Desktop config → See detailed instructions

Gemini CLI Integration

# Example
gemini mcp add-json code-graph-rag '{
  "command": "npx",
  "args": ["@er77/code-graph-rag-mcp", "/path/to/your/codebase"]
}'

Codex CLI Integration

# Recommended: add a *global* MCP server entry (works from any project folder)
codex mcp remove code-graph-rag  # optional cleanup
codex mcp add code-graph-rag -- code-graph-rag-mcp

# Or point Codex directly at a local dev build (no npm/npx required)
codex mcp remove code-graph-rag  # optional cleanup
codex mcp add code-graph-rag -- node /absolute/path/to/code-graph-rag-mcp/dist/index.js

Multi-codebase support: Analyze multiple projects simultaneously → Multi-Codebase Setup Guide

Installation Guide (All Clients)

• NPM: npm install -g @er77/code-graph-rag-mcp
• Run server locally: code-graph-rag-mcp [directory]
• Claude: use Inspector (above) or see Quick Start
• Gemini: configure via gemini mcp add-json ... (above)
• Codex: configure via codex mcp add ... (above)

---

🏆 Performance

5.5x faster than Native Claude tools with comprehensive testing results:

Metric	Native Claude	MCP CodeGraph	Improvement
Execution Time	55.84s	<10s	5.5x faster
Memory Usage	Process-heavy	65MB	Optimized
Features	Basic patterns	26 methods	Comprehensive
Accuracy	Pattern-based	Semantic	Superior

---

🔍 Key Features

🔬 Advanced Analysis Tools (26 MCP Methods)

Feature	Description	Use Case
Semantic Search	Natural language code search	"Find authentication functions"
Code Similarity	Duplicate & clone detection	Identify refactoring opportunities
JSCPD Clone Scan	JSCPD-based copy/paste detection without embeddings	Targeted duplicate sweeps
Impact Analysis	Change impact prediction	Assess modification risks
AI Refactoring	Intelligent code suggestions	Improve code quality
Hotspot Analysis	Complexity & coupling metrics	Find problem areas
Cross-Language	Multi-language relationships	Polyglot codebases
Graph Health	Database diagnostics	get_graph_health
Version Info	Server version & runtime details	get_version
Safe Reset	Clean reindexing	reset_graph, clean_index
Batched Indexing	Resumable indexing with progress (Codex-safe for big repos)	batch_index
Agent Telemetry	Runtime metrics across agents	get_agent_metrics
Bus Diagnostics	Inspect/clear knowledge bus topics	get_bus_stats, clear_bus_topic
Lerna Project Graph	Workspace dependency DAG export, optional ingest, cached refresh control	lerna_project_graph (requires Lerna config)
Semantic Warmup	Configurable cache priming for embeddings	mcp.semantic.cacheWarmupLimit

⚡ High-Performance Architecture

Metric	Capability	Details
Parsing Speed	100+ files/second	Tree-sitter based
Query Response	<100ms	Optimized SQLite + vector search
Agent System	Multi-agent coordination	Resource-managed execution
Vector Search	Hardware-accelerated (optional)	Automatic embedding ingestion
AST Analysis	Precise code snippets	Semantic context extraction

🌐 Multi-Language Support (11 Languages)

Language	Features	Support Level
Python	Async/await, decorators, magic methods (40+), dataclasses	✅ Advanced (95%)
TypeScript/JavaScript	Full ES6+, JSX, TSX, React patterns	✅ Complete (100%)
C/C++	Functions, structs/unions/enums, classes, namespaces, templates	✅ Advanced (90%)
C#	Classes, interfaces, enums, properties, LINQ, async/await	✅ Advanced (90%)
Rust	Functions, structs, enums, traits, impls, modules, use	✅ Advanced (90%)
Go	Packages, functions, structs, interfaces, goroutines, channels	✅ Advanced (90%)
Java	Classes, interfaces, enums, records (Java 14+), generics, lambdas	✅ Advanced (90%)
Kotlin	Packages/imports, classes/objects, functions/properties, relationships	✅ Implemented
VBA	Modules, subs, functions, properties, user-defined types	✅ Regex-based (80%)

---

🛠️ Usage Examples

# Single project analysis
code-graph-rag-mcp
code-graph-rag-mcp /path/to/your/project

# CLI helpers
code-graph-rag-mcp --help
code-graph-rag-mcp --version

# Multi-project setup (see Multi-Codebase Setup Guide)
# Configure multiple projects in Claude Desktop config

# Check installation
code-graph-rag-mcp --help

# Health & maintenance
# Health check (totals + sample)
get_graph_health
# Reset graph data safely
reset_graph
# Clean reindex (reset + full index)
clean_index
# Batched index with progress (recommended for strict clients/timeouts)
batch_index
# Lerna workspace graph (ingest into storage)
lerna_project_graph --args '{"ingest": true}'
# Force refresh graph and re-ingest (bypass cache)
lerna_project_graph --args '{"ingest": true, "force": true}'
# Cached runs return `cached: true`; use `force` to break the 30s debounce when configs change.
# Agent telemetry snapshot
get_agent_metrics
# Knowledge bus diagnostics
get_bus_stats
clear_bus_topic --args '{"topic": "semantic:search"}'

# One-shot index from the CLI (debug mode)
node dist/index.js /home/er77/tt '{"jsonrpc":"2.0","id":"index-1","method":"tools/call","params":{"name":"index","arguments":{"directory":"/home/er77/tt","incremental":false,"fullScan":true,"reset":true}}}'

# Relationships for an entity name
list_entity_relationships (entityName: "YourEntity", relationshipTypes: ["imports"]) 

# Adjust semantic warmup (optional)
export MCP_SEMANTIC_WARMUP_LIMIT=25

# Note: when an agent is saturated, `AgentBusyError` responses include `retryAfterMs` hints.

With Claude Desktop:

1. "What entities are in my codebase?"
2. "Find similar code to this function"
3. "Analyze the impact of changing this class"
4. "Suggest refactoring for this file"

Multi-Project Queries:

1. "Analyze the frontend-app codebase structure"
2. "Find authentication functions in backend-api"
3. "Compare user management across all projects"

---

🧰 Troubleshooting

• Codex/VSCode MCP stdio fails to start
  Codex is strict about stdio: stdout must be JSON-RPC only. As of v2.7.12, console stdout logs are redirected to stderr during MCP runs, and heavy initialization is deferred until after handshake / first tool call.
  Recommended Codex config: omit the directory argument and let the server use the workspace root via roots/list:

  [mcp_servers.code-graph-rag]
  command = "code-graph-rag-mcp"
  args = []
  

  If index / clean_index time out on large repos and the transport closes, prefer batch_index with a small maxFilesPerBatch and keep calling it with the returned sessionId until done:true. If you must see logs on stdout for local debugging, set MCP_STDIO_ALLOW_STDOUT_LOGS=1 (not recommended for strict clients). If startup still fails, check the global tmp log mirror: /tmp/code-graph-rag-mcp/mcp-server-YYYY-MM-DD.log (Linux/macOS; uses os.tmpdir()).

• batch_index fails with agent_busy / memory_limit
  Increase the coordinator/conductor limits (these gate task routing in-process): set COORDINATOR_MEMORY_LIMIT / CONDUCTOR_MEMORY_LIMIT and COORDINATOR_MAX_MEMORY_MB / CONDUCTOR_MAX_MEMORY_MB, or edit config/default.yaml.
  If you see a real Node.js OOM, also start the server with a larger heap, e.g. NODE_OPTIONS="--max-old-space-size=4096" code-graph-rag-mcp.

• Database location / multi-repo isolation By default, the server stores its SQLite DB under ./.code-graph-rag/vectors.db (per repo). Add /.code-graph-rag/ to your project’s .gitignore.

• Native module mismatch (better-sqlite3)
  Since v2.6.4 the server automatically rebuilds the native binary when it detects a NODE_MODULE_VERSION mismatch. If the automatic rebuild fails (for example due to file permissions), run:

  npm rebuild better-sqlite3
  

  in the installation directory (globally this is commonly /usr/lib/node_modules/@er77/code-graph-rag-mcp).

• Legacy database missing new columns
  Older installations might lack the latest `embe

---

README truncated. View full README on GitHub.