NeuroCode: In-Memory AST Context Slicer & Local SWE Agent
NeuroCode is a zero-daemon, in-memory Code Property Graph (CPG) context slicer and autonomous software engineering agent built specifically for local edge models (Ollama, Qwen2.5-Coder) and MCP clients (Cursor, Claude Desktop) running on 4GB–8GB VRAM consumer laptops (e.g. NVIDIA RTX 3050).
Rather than dumping full repositories into 30,000+ token cloud prompts or requiring heavy graph databases (like Neo4j), NeuroCode parses repository ASTs into an in-memory reachability graph and solves a bidirectional knapsack problem in <1 ms to compress issue context into a strict sub-1,200 token budget.
✨ Features
⚡ Instant AST Reachability Slicing: Slices caller/callee signatures, control dependencies, and target function bodies directly from the active file into $<1,200$ tokens in under 100 ms.
📋 1-Click Copy for AI Chat: Copy the minimal, high-density context prompt directly to your clipboard for pasting into Cursor, ChatGPT, Copilot, or Claude without context overflow.
🛠️ Autonomous Edge Bug Repair: Runs single-shot local bug repair via Ollama (qwen2.5-coder:1.5b or 7b) and displays the unified diff patch side-by-side.
📊 Real-Time VRAM & Hardware Health: Directly inspects your GPU VRAM headroom and /dev/shm tmpfs zero-disk-wear status.
🔌 Model Context Protocol (MCP) Integration: Generates .vscode/mcp.json or Claude Desktop configuration in one click to expose in-memory AST slicing as an MCP tool.
🚀 Quick Start
1. Install Backend
Ensure the lightweight neurocode-ai package is installed on your machine:
pip install neurocode-ai
(Optional for local LLM inference): Install Ollama and pull your model:
ollama pull qwen2.5-coder:1.5b
2. Use in VS Code
Status Bar: Click $(circuit-board) NeuroCode in the status bar for instant actions.
Context Menu: Right-click in any Python/code file $\rightarrow$ NeuroCode: Slice Context for Active File.
Sidebar: Open the NeuroCode sidebar tab to view live hardware telemetry and triggers.
⌨️ Commands
Command
Description
NeuroCode: Slice Context for Active File
Generates sub-1,200 token CPG closure for active cursor position.
NeuroCode: Copy Sliced Context to Clipboard
Copies high-density prompt ready for Cursor/Claude/Copilot.
NeuroCode: Autonomous Repair with Local Ollama
Prompts for a bug description and generates a unified diff patch.
NeuroCode: Check Hardware Health & VRAM
Displays RTX 3050 / GPU memory usage and tmpfs diagnostics.
NeuroCode: Export MCP Config (.vscode/mcp.json)
Configures VS Code and Cursor for MCP tool calling.
⚙️ Extension Settings
neurocode.pythonPath: Path to the Python executable where neurocode-ai is installed (default: python3).
neurocode.tokenBudget: Maximum token budget for context closure (default: 1200).
neurocode.ollamaModel: Default edge model for autonomous repair (default: qwen2.5-coder:1.5b).
neurocode.ollamaHost: URL of the local Ollama API (default: http://localhost:11434).
📄 License
MIT License. Created by Deepak R. (Dhaanish Ahmed College of Engineering).
