Guaardvark: Your Self-Hosted AI Studio for Agents, Media, and Code

Introduction

Guaardvark is a powerful, self-hosted AI studio that brings a comprehensive suite of AI capabilities directly to your machine. Designed for privacy and control, it enables local video, image, music, and voice generation, LoRA training, sophisticated coding swarms, RAG (Retrieval Augmented Generation), and screen agents, all optimized to run efficiently on a single GPU. Whether you're a developer, a creator, or an AI enthusiast, Guaardvark offers a complete AI workstation experience without relying on cloud services.

Why Use Guaardvark and Key Features

Guaardvark stands out by prioritizing a local-first approach, ensuring your data remains on your machine and under your control. This eliminates per-token or per-minute fees, content policy restrictions, and vendor lock-in associated with cloud platforms.

Key features and differentiators include:

Comprehensive Media Studio: Generate high-quality video with 11 local models (Wan 2.2, CogVideoX, LTX, HunyuanVideo, MiniMax H3), create images, compose full-song music, and perform neural voice cloning with consent-gated features. It also includes 4K/8K GPU upscaling for images and video frames.
Advanced Agent Capabilities:
- Coding Agents and Swarms: Deploy up to 20 parallel coding agents in isolated Git worktrees, complete with dependency-aware merging and cost tracking. It integrates seamlessly with popular coding agents like Claude Code, Cursor, and Codex via its built-in MCP server and agent skills.
- Screen Agents: Experience agents that control a real Ubuntu/XFCE desktop, equipped with vision, closed-loop servo clicking, and a live VNC viewer, allowing them to interact with any application as a human would.
- Film Crew and Music Video Directors: Specialized agents collaborate to turn a logline into a finished video, handling scripting, casting, storyboarding, and editing. A beat-synced music video director automatically creates films from your songs.
- Self-Improvement: The system can run its own test suite, identify failures, dispatch an AI agent to fix bugs, verify the fix, and broadcast the learning, all behind guardian review and kill switches.
Robust Knowledge Management (RAG): Implement hybrid RAG on pgvector with cross-encoder reranking, AST code chunking, and layout-aware document parsing. It provides grounded citations, per-project isolation, and an Autoresearch feature that tunes retrieval parameters overnight.
Code Intelligence: Utilize a Monaco editor, AST-aware analysis, dependency graphs, and a System Mapper that visualizes your entire codebase, complete with ranked findings and proposed fixes.
GPU Memory Orchestrator: Efficiently manages GPU resources, ensuring that large models do not conflict and optimizing performance across various tasks.
Offline Operation (Flight Mode): All core functionalities, including agent tasks and swarms, are validated to work entirely offline, providing true independence from internet connectivity.

Installation

Guaardvark requires Python 3.12 for its ML stack. For quick setup on Linux, use the following command:

curl -fsSL https://guaardvark.com/install.sh | bash

This command clones the repository to ~/guaardvark (customizable via GUAARDVARK_HOME) and launches ./start.sh. The first run handles all dependencies, including Python 3.12, venv, Node.js, PostgreSQL, Redis, Ollama, and frontend build.

Alternatively, you can clone the repository manually:

git clone https://github.com/guaardvark/guaardvark.git
cd guaardvark
./start.sh

To stop Guaardvark services, use:

./stop.sh

Examples

Guaardvark offers a rich set of capabilities that can be explored through its comprehensive walkthrough series. Witness the system in action, from generating complex music videos to agents fixing their own code.

Media Production: See how Guaardvark generates images, videos, and music, and even clones voices, all locally.
Agent Automation: Watch screen agents interact with a real desktop or observe coding swarms refactor code in parallel.
Knowledge Retrieval: Explore the RAG system as it processes documents and provides grounded answers with citations.

For a detailed visual demonstration of every feature, watch the official Walkthrough Series on YouTube.