Local-First Graph-Native Code Intelligence Engine for AI Agents
CodeAtlas is an open-source, local-first graph-native code intelligence engine. It parses source code repositories into a high-performance structural knowledge graph (Neo4j + SQLite metadata), connects to AI Agents (Cursor, Antigravity, Claude Desktop) via the Model Context Protocol (MCP), and delivers graph-grounded evidence with zero cloud dependence.
CodeAtlas eliminates token bloat and hallucination by retrieving exact symbol definitions, execution chains (POST /upload ➔ uploadController ➔ pipelineService), and file line ranges (file, startLine, endLine).
| Feature | Description |
|---|---|
| ⚡ Graph-RAG Retrieval | Combines concept normalization, lexical search, and 6-factor Neo4j graph reranking. |
| 🤖 8 Progressive MCP Tools | Exposes codeatlas_search, codeatlas_get_context, codeatlas_find_symbol, codeatlas_get_callers, codeatlas_get_callees, codeatlas_get_dependencies, codeatlas_analyze_impact, codeatlas_trace_execution. |
| 🔒 100% Local Privacy | Operates completely offline (~/.codeatlas/). Zero code uploads, zero telemetry, zero mandatory cloud APIs. |
| 🚀 Multi-Core Parallel Indexer | Bounded CPU worker pool featuring SHA-256 content hash skipping for sub-second re-indexing. |
| 🔌 Multi-Provider AI Factory | Supports Anthropic Claude, Google Gemini, OpenAI, Ollama, or 100% Local-First Graph-RAG mode. |
| 📊 Control Center Dashboard | Next.js 16 interactive graph workspace with tabbed node inspector, execution tracing, and Docker configurator. |
Clone CodeAtlas and link it globally on your machine:
git clone https://github.com/Hellnight2005/CodeAtlas-.git
cd CodeAtlas-
npm install
npm linkNavigate to any codebase folder on your system:
cd /path/to/your/project
# Initialize configuration
codeatlas init
# Index the repository AST into structural code graph
codeatlas index .
# One-command start (Boots REST API Engine, Dashboard UI, and MCP transport)
codeatlas start- 🌐 Dashboard UI: http://localhost:3001
- 🔌 REST API Engine: http://localhost:5001
Add CodeAtlas to your AI Agent's configuration file (mcp_config.json):
{
"mcpServers": {
"codeatlas": {
"command": "codeatlas",
"args": ["mcp", "start"],
"cwd": "/path/to/your/project"
}
}
}Now ask your AI Agent natural questions:
- 💬 "How does the upload pipeline work in this repository?"
- 💬 "What happens after POST /upload?"
- 💬 "If I modify
directUpload.js, what components could be impacted?"
User Question (Natural Language)
│
▼
AI Agent (Cursor / Antigravity / Claude)
│
│ Model Context Protocol (stdio)
▼
┌──────────────────┐
│ CodeAtlas MCP │
└─────────┬────────┘
│
▼
Concept Normalizer & Search
│
┌────────┴────────┐
▼ ▼
Lexical Search Neo4j Graph Expansion
│ │
└────────┬────────┘
▼
6-Factor Reranker
│
▼
Token-Budgeted Context
│
▼
Grounded Answer to Agent
| Command | Description |
|---|---|
codeatlas start |
One-command environment start (API Engine + Dashboard UI + MCP). |
codeatlas stop |
Gracefully stop background services and release project storage locks. |
codeatlas reset |
Safely reset graph index data for a target project (source code untouched). |
codeatlas init |
Initialize local .codeatlas/config.yaml in active repository. |
codeatlas index . |
Scan repository, parse ASTs, build structural code graph. |
codeatlas watch |
Watch repository for file modifications and apply incremental updates. |
codeatlas serve |
Launch REST API engine (5001) and Dashboard UI (3001). |
codeatlas doctor |
Run environment diagnostics (Node.js, SQLite, Neo4j, Search). |
codeatlas benchmark |
Run automated latency & context token efficiency benchmarks. |
| Metric | Measured Target |
|---|---|
| Initial Indexing Throughput | ~550 files / sec (Bounded CPU Concurrency) |
| Incremental Re-indexing | sub-second (< 180ms SHA-256 hash skip) |
| Search & Concept Normalization | 12ms |
| Neo4j Graph Expansion (Depth 2) | 48ms |
| 6-Factor Reranking & Context Build | 24ms |
| Total MCP Response Latency | ~184ms |
Built for developers and AI pair programmers worldwide. Contributions and issues are welcome!