Guaardvark

What guaardvark/guaardvark MCP server does

guaardvark/guaardvark MCP server exposes a self-hosted AI studio to compatible coding agents. The studio combines local chat and retrieval with agents that can operate a desktop and browser, parallel coding worktrees, and GPU-backed generation for images, video, music, and voice.

Its MCP interface is intended to let an agent run complete workflows rather than only issue isolated media-generation requests. Supported workflows described by the project include asking questions over uploaded files, coordinating coding swarms, generating or editing media, producing music videos, and operating an Ubuntu desktop. The repository also includes an MCP client, allowing the studio’s chat experience to call tools provided by other MCP servers.

How it works

Local language models, retrieval components, agents, and generators run through Ollama, ComfyUI, and the project’s GPU services. A single orchestrator schedules work on the available GPU, so language-model activity and rendering jobs take turns instead of competing for VRAM.

For retrieval, documents are stored in a Postgres-backed pgvector system. Queries combine keyword and vector search, can use cross-encoder reranking, and return citations with source information. The media side includes local image and video models, LoRA support, music and speech generation, video editing, and queued jobs. The exact models and capabilities are enumerated separately in the project’s capability list.

The guaardvark/guaardvark MCP server can also expose desktop, shell, and browser-related operations. Those tools remain hidden unless they are permitted by configuration. Actions with external consequences, including posting, require explicit approval, and voice cloning is consent-gated.

Setup and configuration

The documented quick start targets Linux with an NVIDIA GPU. The project provides an install script that clones the repository into ~/guaardvark and starts the application; the web interface is then available at http://localhost:5173. The README also points to Docker, Apple Silicon, a PyPI CLI, and a full installation guide, but hardware and feature availability vary by platform.

Claude Code can add the project through its plugin marketplace with:

/plugin marketplace add guaardvark/guaardvark
/plugin install guaardvark@guaardvark

The README describes hardware guidance of 16 GB VRAM for video and 12 GB for the Wan 2.2 5B model, while chat, retrieval, and agents can run with less. Review the installer before execution if you need to inspect its behavior.

Tools and capabilities

The guaardvark/guaardvark MCP server is documented as supporting agent access to:

• Local file chat and RAG with citations, hybrid retrieval, and retrieval traces
• Desktop and browser agents with screen viewing and click targeting
• Parallel coding agents using separate git worktrees
• Image, video, music, voice, and speech-generation workflows
• Video editing, music-video direction, Film Crew production, and job queues
• GPU orchestration, scheduling, backups, and connections to other MCP servers

Limitations and notes

This project is designed around local hardware, and demanding generation workloads depend on available GPU memory. Scanned-document OCR is not installed by default. Outbound access is disabled until enabled or started, and the project records allowed outbound paths in egress.json.

The MCP interface does not automatically grant unrestricted computer control: desktop, shell, and browser tools are subject to an allowlist. Posting and voice cloning require approval, while code changes proposed by the project’s self-improvement features wait for review. The provided material does not enumerate individual MCP tool names or specify a transport beyond describing a built-in MCP server.