Code-Graph-RAG parses a codebase into a knowledge graph and lets you query, edit, and optimize the code in plain English. It reads every source file, extracts functions, classes, methods, modules, and the relationships between them, and stores the result in Memgraph under one language-agnostic schema, so a monorepo of mixed languages shares a single graph.
It has two components. A Tree-sitter parser ingests the code structure, and where a toolchain is available, compiler-grade frontends add exact facts on top (libclang for C and C++, the Go type checker, and opt-in Roslyn, javac, and Jedi for C#, Java, and Python). Dynamic tracing of a test run, or production eBPF profiles, merges calls observed at runtime into the graph to expose dispatch that static analysis misses. A RAG layer then turns a natural-language question into a Cypher query, retrieves the matching code, and drives editing and optimization.
Python, TypeScript, TSX, JavaScript, Rust, Go, Java, C, C++, C#, PHP, Lua, and Dart are fully supported. Scala is in development, and Ruby, Kotlin, Swift, Elixir, Haskell, Solidity, Bash, and Nix get structural support (modules, functions, classes, and imports) through an ast-grep tier.
Code-Graph-RAG runs as an interactive CLI, as an MCP server for Claude Code and other MCP clients, and as a Python SDK for graph loading, Cypher generation, and semantic search. Memgraph and Qdrant run as a packaged local stack, several repositories can share one graph, and the project also offers managed cloud and on-premise deployments.
Features
- Knowledge graph: functions, classes, methods, modules, imports, and call relationships stored in Memgraph under one schema
- Natural-language queries: questions translated into Cypher queries and answered from the graph
- Code retrieval: returns the source of any function, class, or method by name or intent
- AST-based editing: surgical patches targeted at functions, with a diff preview before changes apply
- Optimization: suggestions against language best practices or your own coding standards
- Dead code detection: walks call and reference edges from entry points
- Structural search and rewrite: AST pattern matching with ast-grep
- Runtime tracing: merges calls from test runs or production eBPF profiles into the graph
- Compiler frontends: libclang, Go types, Roslyn, javac, and Jedi add exact facts where a toolchain exists
- Semantic search: vector search over code through Qdrant
- MCP server: query and edit the codebase from Claude Code and other MCP clients
- Python SDK: graph loader, Cypher generator, and semantic search modules
