guaardvark

Media and generation features

The media studio runs locally with no cloud APIs. It lists seven local video models, including Wan 2.2, CogVideoX, and LTX, plus image generation, full song music generation, and neural voice with consent gated cloning. Upscaling reaches 4K and 8K using the Real-ESRGAN family and related models, with two pass and frame by frame video modes. A built in video editor provides a three lane timeline for video, text, and audio, real ffmpeg drawtext overlays, drag and drop from a media library, visual trims, undo, and keyboard shortcuts.

Agents and orchestration

Guaardvark routes every message through an AgentBrain three tier engine: Reflex fires pattern matched replies in under 100 milliseconds with no inference, Instinct handles single shot requests in one model call, and Deliberation runs a full ReACT loop for multi step work. Screen agents control a real Ubuntu or XFCE desktop at 1024 by 1024, seen through vision models that move the mouse, click, type, and verify their own actions. The swarm orchestrator launches up to 20 coding agents in isolated git worktrees, merging results with dependency ordered conflict detection and cost tracking.

Local guarantees and MCP

Everything runs locally by default, with no telemetry or cloud contact unless the Interconnector is explicitly enabled. Flight Mode provides fully offline operation with automatic network detection and local model fallback, and swarm and agent tasks were validated end to end without internet. The MCP server uses a default deny policy where desktop control, agent execution, system or shell, browser automation, and test execution tools are hidden unless explicitly allowlisted. Voice cloning requires an explicit consent prompt and keeps reference clips under user control.

Editorial conclusion

Guaardvark is released under the MIT license and installs with a single curl command, keeping all media generation and agent work on the user's own hardware.