>_faiz.palwala

GraphContext MCP

[IN PROGRESS]Shipping as a paid product, not an open-source release. No public repo link, by design.

The Problem: Chunk-Based RAG Fails at Systems Code

Modern AI coding assistants (Cursor, Claude Desktop) rely heavily on naive text chunking and vector similarity search. In large, interconnected software codebases, this approach systematically loses structural context:

  1. Lost Dependency Topology: Text embeddings cannot trace transitive call graphs, type hierarchies, or cross-file state mutations.
  2. High Token Bloat: Injecting arbitrary file chunks into agent contexts consumes excessive tokens while failing to inform the model of downstream blast radius.
  3. Privacy Vulnerability: Sending proprietary internal codebases to cloud vector stores violates enterprise compliance.

The Systems Architecture

GraphContext MCP operates as a local-first structural “brain” running natively on macOS via the Model Context Protocol (MCP):

1. Deterministic AST Parsing (Tree-sitter)

Constructs Heterogeneous Code Property Graphs (CPGs) locally with sub-millisecond AST extraction across Python, C++, TypeScript, and Rust, mapping precise symbol references and control-flow dependencies.

2. Local Graph Neural Network Inferences (Apple MLX)

Executes Graph Neural Network (GNN) embeddings directly on Apple Silicon unified memory using Apple’s MLX framework, calculating structural symbol representations without cloud roundtrips.

3. Change Blast-Radius Prediction (GraphSAGE)

Predicts the regression impact and dependency blast radius of code edits before generation. The underlying GraphSAGE architecture was pre-trained across historical open-source bug-fix commits processed via PySpark and PyTorch Geometric on an HPC cluster.

4. Zero-Latency Model Context Protocol (MCP) Server

Integrates natively with Claude Desktop and Cursor via a high-throughput local FastAPI/MCP daemon, performing live structural cache invalidation as files change in the editor.