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

Claude Code Integration

Claude Code (the CLI, claude) enforces a hardcoded 15-second timeout on every MCP tool call. When using a hosted embedding provider (OpenAI, Gemini via OpenAI-compat, etc.), the default 50-seed cache warmup can push the first semantic_search call past this limit and return timed out after 15000ms — even though subsequent calls complete in well under a second.

Recommended setup:

claude mcp add code-graph-rag -s user \
  -e MCP_EMBEDDING_PROVIDER=openai \
  -e MCP_EMBEDDING_MODEL=gemini-embedding-001 \
  -e MCP_EMBEDDING_ENABLED=true \
  -e OPENAI_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai \
  -e OPENAI_API_KEY=YOUR_KEY \
  -e MCP_SEMANTIC_WARMUP_LIMIT=0 \
  -- code-graph-rag-mcp /path/to/your/codebase

Key flag: MCP_SEMANTIC_WARMUP_LIMIT=0 disables the cache warmup so the semantic agent initializes in milliseconds, keeping first-query latency inside Claude Code's 15s window. Subsequent queries hit the warm agent and return in <500ms.

Unlike Claude Desktop, Claude Code does not honor a per-server MCP_TIMEOUT env var — the 15s limit is enforced client-side with no override available today.

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

• Claude Code: semantic_search timed out after 15000ms on first call
  Claude Code (the CLI) enforces a hardcoded 15-second timeout on every MCP tool call and does not honor per-server MCP_TIMEOUT env vars. The default cache warmup (cacheWarmupLimit: 50) issues up to 50 embedding requests during agent init, which can exceed 15s when using a hosted provider (OpenAI, Gemini, etc.). Set MCP_SEMANTIC_WARMUP_LIMIT=0 in the MCP server env to skip warmup — first query will then complete inside the 15s window, and subsequent queries hit the warm agent and return in under a second. See the Claude Code Integration section above for the full recommended claude mcp add command.

• 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 embeddings columns (metadata, model_name, etc.). The server now auto-upgrades in place, but if you still encounter migration errors, delete the local DB and re-run the indexer:

  # delete the DB shown in logs as: "[Config] Database path: ..."
  rm -f ./.code-graph-rag/vectors.db ./.code-graph-rag/vectors.db-wal ./.code-graph-rag/vectors.db-shm
  

  Then start the server again to trigger a clean rebuild.

• Running a one-shot index from the CLI
  You can trigger tools directly by passing JSON-RPC payloads. When a payload is supplied, the server skips the semantic agent and uses low-memory batching for debugging. Example:

  node dist/index.js /path/to/project \
    '{"jsonrpc":"2.0","id":"index-1","method":"tools/call","params":{"name":"index","arguments":{"directory":"/path/to/project","incremental":false,"fullScan":true,"reset":true}}}'
  

  The command logs progress to logs_llm/mcp-server-YYYY-MM-DD.log. Set MCP_DEBUG_DISABLE_SEMANTIC=0 if you want embeddings enabled during the run.

---

📋 Changelog

🚀 Version 2.7.11 (2025-12-15) - Per-Repo Database Isolation

• 🗄️ Default DB path moved to ./.code-graph-rag/vectors.db so multiple codebases don’t share/mix a single global SQLite database

🚀 Version 2.7.12 (2025-12-15) - Remove Deprecated boolean Install Warning

• 🧹 npm install: stop auto-installing onnxruntime-node by default (optional peer dep instead), removing the boolean@3.2.0 deprecation warning during install

🚀 Version 2.7.10 (2025-12-15) - sqlite-vec Global Install Fix

• 🧠 sqlite-vec: load the extension via sqlite-vec’s getLoadablePath() so global installs work regardless of project cwd

🚀 Version 2.7.9 (2025-12-15) - Codex Config Fixes

• 🧭 Codex docs: use codex mcp add ... global config (works from any project folder)
• 🧾 Removed references to non-existent helper scripts

🚀 Version 2.7.8 (2025-12-15) - MCP Startup Diagnostics

• ✅ TTY-safe stdio: redirect console stdout logs → stderr during MCP runs (prevents handshake breaks)
• 🗂️ Global log mirror: always write a copy to /tmp/code-graph-rag-mcp/mcp-server-YYYY-MM-DD.log for early-start debugging
• 🛡️ Resilient logging: if logs_llm/ can’t be created, fall back to os.tmpdir() instead of exiting before initialize

🚀 Version 2.7.7 (2025-12-14) - Codex StdIO Hardening + Kotlin

• ✅ Codex/VSCode stdio compatibility: reserve stdout for JSON-RPC and route logs to stderr
• ⚡ Faster readiness: defer heavy runtime init until after handshake / first tool call
• 📁 Project root now optional: defaults to cwd, prefers MCP roots/list workspace root when available
• 🧩 Kotlin support for .kt/.kts (tree-sitter + KotlinAnalyzer) with tests + ADR-006

🚀 Version 2.7.4 (2025-11-02) - Clone Reporting & CLI Boost

• 🆕 Added --help/-h and --version/-v flags for quick CLI interaction
• 🔄 detect_code_clones now merges deduplicated semantic groups with JSCPD summaries
• 📊 jscpd_detect_clones summary includes duplicated line/token counts, percentages, and inline snippets
• 🧪 Integration coverage verifies clone-report fields to guard against regressions
• 🔁 Maintains vendored lightweight JSCPD pipeline with zero external deps

🚀 Version 2.7.0 (2025-11-02) - JSCPD Clone Detection

• ➕ Added jscpd_detect_clones MCP tool exposing JSCPD duplicate detection without requiring embeddings
• 🧩 Vendored lightweight JSCPD core/tokenizer for zero external build dependencies
• 🧪 New integration fixtures ensure JSCPD scans surface expected duplicate blocks
• 📚 README / tool catalog now counts 23 MCP methods and documents JSCPD usage

🚀 Version 2.6.0 (2025-10-12) - Major Architecture Upgrade

Breaking Changes & Major Improvements ⚡

• 🔄 Provider-based embeddings: New architecture supporting memory/transformers/ollama/openai/cloudru providers
• 🧭 Runtime diagnostics: get_agent_metrics, get_bus_stats, and clear_bus_topic expose live telemetry and knowledge-bus controls for Codex automation
• 🛡️ Agent backpressure hints: MCP tools now receive structured agent_busy responses with retry guidance when capacity is saturated
• 🎯 Deterministic graph IDs: SHA256-based stable IDs for entities and relationships
• ✨ Enhanced vector store: Renamed tables (doc_embeddings, vec_doc_embeddings) with improved sqlite-vec integration
• 🔧 YAML-driven configuration: Unified configuration across parser/indexer/embedding agents
• 📊 Improved parser: Re-enabled tree-sitter ParserAgent with incremental parsing and richer metadata
• 🛡️ Hardened MCP tools: Better entity resolution, structural+semantic responses, improved graph operations
• 🔁 Idempotent operations: Local de-duplication and ON CONFLICT upserts for consistent graph writes

Technical Details:

• Dynamic dimension detection at runtime with safe fallbacks
• Batch deduplication by ID with transactional updates
• Enhanced language analyzers with structured pattern data
• SQLiteManager + GraphStorage singleton for consistency

Testing & Validation (2025-10-21):

• ✅ All 16/16 test suites passing (200+ individual tests, 93.75% success rate)
• ✅ 100% MCP method validation (22/22 methods comprehensively tested)
• ✅ v2.6.0 new methods validated: get_agent_metrics, get_bus_stats, clear_bus_topic
• ✅ Integration test coverage: All core components, semantic operations, and monitoring tools
• ✅ v2.5.9 dual-schema fixes preserved and enhanced with sqliteVecEnabled property
• ✅ Zero regressions after PR #20 integration
• ⚠️ Known issue: Duplicate case "get_graph" in src/index.ts:1668 & 1707 (non-critical, line 1707 unreachable)

🎉 Version 2.5.9 (2025-10-06) - 100% Success Rate

• ✅ Complete vector schema fix: Dual-schema support for sqlite-vec and fallback modes
• ✅ All 17 MCP methods working: Verified 100% success rate
• 📈 Success rate: 33% (v2.5.7) → 61% (v2.5.8) → 100% (v2.5.9)

Version 2.5.8 (2025-10-06) - Critical Infrastructure Fixes

• ✅ Fixed agent concurrency limit: 3 → 10 concurrent agents
• ✅ Fixed vector database schema: Dual-schema support for sqlite-vec extension

Previous Versions

<details> <summary>Click to expand version history (2.5.7 - 2.3.3)</summary>

v2.5.7 - Semantic analysis improvements, lowered thresholds, clone detection v2.5.6 - Fixed DoraAgent type collision (+16% success rate) v2.5.5 - WASM path resolution fix, get_version tool, restart script v2.5.4 - Architecture Decision Records (ADRs) v2.5.3 - Deprecated dependency warning suppression v2.5.2 - Enhanced README documentation v2.5.1 - Python magic methods, import analysis enhancements v2.5.0 - 8 new languages (C#, Rust, C, C++, Go, Java, VBA), Research Trinity v2.4.1 - Rust AST parsing, system architecture docs v2.4.0 - Health check tools, AST hotspots, semantic routing v2.3.3 - Entity extraction fix (0 → 4,467 entities)

</details>

---

⚡ System Requirements

Minimum: Node.js 24+, 2GB RAM, Dual-core CPU Recommended: Node.js 24+, 8GB RAM, Quad-core CPU with SSD

Known Issues

• Deprecated boolean@3.2.0 warning: This is a transitive dependency from the optional onnxruntime-node package (used for ML embeddings). The package is deprecated but functional. The warning can be safely ignored as it doesn't affect core functionality.

---

🤝 Contributing

1. Fork the repository
2. Follow Agent Governance rules
3. Submit pull request

Contributing Guide • Issue Tracker

---

📄 License

MIT License - see LICENSE file for details.

Links: GitHub • NPM • MCP Protocol