CodeGraph turns your local codebase into a searchable knowledge graph so any Model Context Protocol (MCP) client can answer questions with project-aware context. The CLI indexes source code, pushes semantic embeddings into FAISS, and exposes a fast MCP server that tools such as LM Studio or Claude Desktop can call.

This README focuses on the streamlined developer workflow we ship today. For a deeper dive into the agent architecture or legacy experiments, see the material moved under docs/legacy/.

1. Install prerequisites

   # macOS / Linux prerequisites curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh brew install faiss # or your distro's libfaiss-dev package

2. Install CodeGraph

   cd codegraph-rust cargo install --path crates/codegraph-mcp --features "embeddings,faiss"

3. Prepare LM Studio (recommended runtime)

   • Download LM Studio from lmstudio.ai
   • Download the models:
     • Embeddings: jinaai/jina-code-embeddings-1.5b
     • LLM (optional): lmstudio-community/DeepSeek-Coder-V2-Lite-Instruct-Q4_K_M
   • Start the LM Studio local server on http://localhost:1234

4. Configure CodeGraph Create .env in your repository (or export the variables):

   CODEGRAPH_EMBEDDING_PROVIDER=lmstudio CODEGRAPH_EMBEDDING_MODEL=jinaai/jina-code-embeddings-1.5b CODEGRAPH_LMSTUDIO_URL=http://localhost:1234 CODEGRAPH_EMBEDDING_DIMENSION=1536 RUST_LOG=warn

5. Index and serve

   codegraph index . # builds .codegraph/ with embeddings + graph codegraph start stdio # exposes MCP interface for LM Studio / Claude

That is the entire end-to-end workflow. LM Studio now has a project-specific vector store it can query through MCP.

Tips:

• Run codegraph index from each repository root you want to serve.
• Keep RUST_LOG=warn in your environment for a clean TUI while indexing.
• To script updates, rerun codegraph index after significant code changes; the cache handles incremental updates.

These numbers come from the same benchmarking harness described in docs/performance/. They are preserved here so you can quickly sanity-check your own runs.

The full LM Studio checklist (model recommendations, port configuration, batch-size tuning, and troubleshooting) lives in LMSTUDIO_SETUP.md. It expands on:

• Switching between embedding and completion models within LM Studio
• Running multiple projects side by side
• Recommended .env / .codegraph.toml settings

For Ollama or ONNX-only setups, see docs/INSTALLATION.md and the legacy guides under docs/legacy/.

• Graph analytics: The indexer produces ownership graphs, dependency edges, and metadata that RAG tools consume automatically.
• MCP tools: vector.search, graph.traverse, code.read, and more are exposed through the MCP server.
• FAISS tuning: Set CODEGRAPH_FAISS_TRAINING_THRESHOLD or provide a pre-built index in .codegraph/faiss.index for massive monorepos.

Explore the docs/ tree for deeper architecture notes, RAG pipeline details, deployment manifests, and operator playbooks.

We welcome pull requests! Running the full test suite:

cargo fmt --all cargo clippy --workspace --all-targets cargo test --workspace

Please open an issue if you want to discuss large changes or integrations.

CodeGraph is dual-licensed under MIT and Apache 2.0. See LICENSE-MIT and LICENSE-APACHE for details.