AntiGravity: Node-Based LLM Logic Engine

Stack: Python 3.12+, PySide6 (Qt), httpx
Status: Professional Prototype / Proof of Concept

Overview

AntiGravity is an exploration into non-linear LLM interaction. While the industry has settled on the "linear chat" paradigm rooted in 1970s terminal interfaces, AntiGravity looks towards the VFX and Gaming industries for inspiration, using a Directed Acyclic Graph (DAG) to structure AI logic.

Developed as a "vibe coding" experiment using Google AntiGravity, this tool allows you to branch, merge, and visualize complex LLM workflows that would be impossible to manage in a standard chat box.

Why a Node Graph?

• Non-Linear Branching: Explore multiple "what-if" scenarios simultaneously without losing context.
• Smart Context Assembly: Inherit history from specific parent nodes or @reference connected outputs.
• Visual Token Management: Real-time meters show exactly how much context you're sending to the model.
• Multi-Provider Hub: Seamlessly mix nodes from Ollama (local), OpenAI, Gemini, and OpenRouter in a single canvas.

Key Features

• Professional Canvas: Drag-and-drop node editor with smooth pan/zoom and resizable components.
• MVP Architecture: Clean separation between business logic (core/) and UI rendering (ui/).
• Command System: Complete Undo/Redo support for every graph operation.
• Queue-Based execution: Run multiple nodes concurrently with real-time status indicators and timers.
• High-Level API: Includes a developer API (core/api.py) for automated testing and advanced flow control.

Installation

Prerequisites

• Python 3.12 or higher
• uv package manager (recommended)

Setup

# Clone the repository
git clone <repository-url>
cd AntiGravity

# Install dependencies and run
uv sync
uv run python main.py

Quick Start

1. Add Nodes: Use Ctrl+N or right-click the canvas.
2. Connect: Drag from an output port (right) to an input area (left).
3. Configure: Click the model label on a node to set specific provider/model overrides.
4. Reference: Use @NodeName in your prompt to pull in data from connected nodes.
5. Run: Click the "RUN" button or Ctrl+Enter to execute the node.

Architecture & Contribution

This project follows a strict modular design:

• Core: Graph logic, command pattern, and context assembly.
• UI: Customized Qt components using a centralized theme system.
• Services: Async workers for non-blocking LLM requests.

For a deeper dive into the system design, check out ARCHITECTURE.md. Contributions are welcome—please see CONTRIBUTING.md for testing and style guidelines.

---

AntiGravity - Because linear chat is just a bit boring.