Code Intelligence MCP Server

A Model Context Protocol (MCP) server that provides intelligent code analysis and context retrieval for AI coding assistants.

What is this?

Code Intelligence MCP Server enhances AI coding assistants with deep codebase understanding through:

• Semantic Search: Find code by meaning, not just keywords
• Graph-RAG: Context-aware code retrieval using knowledge graphs
• Call Chain Analysis: Trace function dependencies and impacts
• Smart Context: Automatically inject relevant code snippets into AI conversations

Perfect for use with Claude Code, Cline, or any MCP-compatible AI assistant.

Quick Start

Installation

Via npm (recommended):

npm install -g @ozbombor/code-intelligence-mcp

Via git:

git clone https://github.com/Darkbluelr/code-intelligence-mcp.git
cd code-intelligence-mcp
./install.sh

Requirements

• Node.js >= 18.0.0
• Bash shell
• ripgrep (optional, for faster search)
• jq (optional, for JSON processing)

Usage

As MCP Server (recommended):

Add to your MCP client configuration:

{
  "mcpServers": {
    "code-intelligence": {
      "command": "code-intelligence-mcp",
      "args": []
    }
  }
}

Command Line:

ci-search "find authentication code"
ci-search --help

Core Features

Semantic Code Search

# Natural language queries
ci-search "how does user authentication work"

Graph-RAG Context Retrieval

# Get relevant context with smart pruning
ci_graph_rag --query "fix login bug" --budget 8000

Call Chain Tracing

# Trace function dependencies
ci_call_chain --symbol "handleLogin" --depth 3

Impact Analysis

# Analyze change impact
ci_impact --file "src/auth.ts"

Bug Location

# Intelligent bug locator
ci_bug_locate --error "TypeError: Cannot read property 'user'"

Available MCP Tools

Tool	Description
ci_search	Semantic code search with embeddings
ci_graph_rag	Graph-based context retrieval with smart pruning
ci_call_chain	Function call chain tracing
ci_bug_locate	Intelligent bug location with impact analysis
ci_complexity	Code complexity analysis
ci_hotspot	High-churn file detection
ci_impact	Transitive impact analysis with confidence decay
ci_arch_check	Architecture rule validation
ci_vuln	Vulnerability scanning and dependency tracing
ci_index_status	Manage embedding index (status/build/clear)
ci_boundary	Code boundary detection (user/library/generated)
ci_graph_store	Graph store operations (init/query/stats)
ci_federation	Cross-repo API contract tracking
ci_ast_delta	Incremental AST parsing
ci_cod	Architecture visualization (Mermaid/D3.js)
ci_intent	Query history and preference learning

20+ tools available - Each tool supports --help for detailed usage.

Configuration

Embedding Providers

The server supports multiple embedding providers with automatic fallback:

Three-tier Fallback Strategy:

1. Ollama (Local, Free) → Preferred, no network latency, privacy-safe
2. OpenAI API (Cloud, Paid) → Automatic fallback when Ollama unavailable
3. Keyword Search → Final fallback when no embedding service available

Option 1: Ollama (Recommended for Local Use)

Installation:

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull the embedding model
ollama pull nomic-embed-text

Configuration:

# .devbooks/config.yaml
embedding:
  provider: ollama  # or 'auto' for automatic detection
  ollama:
    model: nomic-embed-text
    endpoint: http://localhost:11434
    timeout: 30

Usage:

# Build the embedding index
ci-search build

# Search with Ollama
ci-search "authentication code"

Option 2: OpenAI API (Cloud)

Setup:

# Set your API key
export OPENAI_API_KEY="sk-..."

Configuration:

# .devbooks/config.yaml
embedding:
  provider: openai
  openai:
    model: text-embedding-3-small
    api_key: ${OPENAI_API_KEY}  # or set directly
    base_url: https://api.openai.com/v1
    timeout: 30

Usage:

# Build index with OpenAI
ci-search build

# Or force OpenAI for a single search
ci-search "user authentication" --provider openai

Option 3: Keyword Search (No Setup Required)

Keyword search works out of the box without any configuration:

# Force keyword search
ci-search "error handling" --provider keyword

Advanced Configuration

Automatic Provider Detection:

# .devbooks/config.yaml
embedding:
  provider: auto  # Tries Ollama → OpenAI → Keyword
  fallback_to_keyword: true
  auto_build: true

Command-line Override:

# Override provider for a single search
ci-search "test" --provider ollama --ollama-model mxbai-embed-large

# JSON output with custom settings
ci-search "payment" --format json --top-k 10 --threshold 0.7

Environment Variables:

• OPENAI_API_KEY - OpenAI API key
• EMBEDDING_API_KEY - Generic embedding API key
• PROJECT_ROOT - Project root directory

Optional Features

All features degrade gracefully:

• No embeddings? Falls back to keyword search
• No SCIP index? Uses regex parsing
• No external tools? Core features still work
• No git history? Hotspot analysis disabled

This means you can start using the MCP server immediately without any configuration.

Documentation

For advanced usage, see the local documentation:

• Complete tool reference and examples
• Architecture and system design
• Performance tuning guide
• Troubleshooting tips

Run ./install.sh to access full documentation locally.

Examples

Automatic Context Injection

Enable automatic context injection for Claude Code:

ci-setup-hook

This installs a hook that automatically injects relevant code context when you interact with Claude, making your AI assistant more aware of your codebase without manual queries.

Custom Queries

# Search with specific mode
ci-search "user service" --mode semantic --limit 5

# Trace call chains
ci_call_chain --symbol "processPayment" --direction both

# Check architecture
ci_arch_check --path src/

Demo Suite

Run comprehensive demos with A/B comparison:

# Run quick comparison demo
bash demo/00-quick-compare.sh

# Run all demos
bash demo/run-suite.sh

# Compare versions (if available)
bash demo/compare.sh baseline.json current.json

The demo suite showcases:

• Automatic context injection vs manual queries
• Semantic search and Graph-RAG capabilities
• Performance benchmarks and metrics
• A/B comparison between different configurations

Performance

Metrics are generated by python3 benchmarks/run_benchmarks.py into benchmark_result.json. Use python3 benchmarks/update_readme.py to refresh this section.

Updated at: 2026-01-24T06:27:38+00:00

Metric	Value	Notes
Semantic search P95 latency	222 ms	iterations=3
Graph-RAG cold P95 latency	862.64 ms	iterations=3
Graph-RAG warm P95 latency	69.37 ms	iterations=3
Graph-RAG speedup	91.96 %	cold vs warm
Retrieval MRR@10	0.4264	dataset=self, queries=12
Retrieval Recall@10	1.0	dataset=self, queries=12
Retrieval Precision@10	0.3377	dataset=self, queries=12
Retrieval P95 latency	6.10 ms	dataset=self, queries=12
Cache hit P95 latency	75 ms	iterations=20
Full query P95 latency	89 ms	iterations=20
Precommit staged P95	35 ms	iterations=20
Precommit deps P95	27 ms	iterations=20
Compression latency	8.95 ms	iterations=1

Contributing

Contributions welcome! Please read CONTRIBUTING.md first.

License

MIT License - see LICENSE file for details.

Acknowledgments

Built with: Model Context Protocol, tree-sitter, SCIP

---

Need help? Open an issue on your repo or check docs/TECHNICAL.md.