Guaardvark

Guaardvark

The self-hosted AI studio. Chat with your own files, hand tasks to agents that use a real desktop and browser, run parallel coding agents, and generate video, images, music and voice. Everything runs locally, on your own GPU.

Quick start · Features · Videos · FAQ · Full capability list · guaardvark.com

https://github.com/user-attachments/assets/6ee72d6a-3a25-4380-92a9-a4b66e8b7e05

Why Guaardvark

• It runs on your machine. Language models, retrieval, agents and every generator run on your own hardware, through Ollama, ComfyUI and Guaardvark's own GPU services. No cloud API keys, no per-token or per-minute fees, nothing reported home. Web access, posting and every other outbound path stay off until you switch them on or start them, and all of them are declared in one file, egress.json.
• One 16 GB card runs the whole studio. Every image model and every video model except the two larger MiniMax H3 builds run on a single 16 GB NVIDIA card, along with chat and RAG, screen agents, coding swarms, music and voice. One orchestrator hands the GPU to each job, so a render and a language model never fight over VRAM (what runs on 16 GB).
• Your coding agent can drive it. A built-in MCP server, fifteen agent skills and plugins for Claude Code, Grok Build and Antigravity let Claude Code, Grok, Antigravity, Cursor, Codex, Gemini CLI and other agents run every flow: "make a music video from this song", "film this script", "train a LoRA of this character", "swarm this refactor".
• You approve what matters. Nothing posts and no voice is cloned without an explicit yes from you, and fixes Guaardvark writes for its own code wait for your review. The MCP server hides desktop, shell and browser tools unless you allowlist them.

Quick start

Linux with an NVIDIA GPU (16 GB for video, 12 GB for the Wan 2.2 5B; chat, RAG and agents run on less, see the hardware tiers). A fresh Ubuntu desktop has no curl or git; run sudo apt install -y curl git first.

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

Then open http://localhost:5173. The script is install.sh from this repository: it clones to ~/guaardvark and runs ./start.sh, which installs everything on the first run. To read it before it runs, download it first (curl -fsSL https://guaardvark.com/install.sh -o install.sh, read it, then bash install.sh), or install by hand.

Add it to Claude Code (two lines, no clone):

/plugin marketplace add guaardvark/guaardvark

/plugin install guaardvark@guaardvark

Other ways in: Docker · macOS on Apple Silicon · the CLI from PyPI · the full install guide. The current release is in the VERSION file.

Features

• Chat with your own files. Local RAG on pgvector: keyword and vector search fused per query, cross-encoder reranking, source citations, and retrieval that shows the chunks and scores behind an answer. Autoresearch tunes retrieval overnight. ▶ Ep 3
• Plan mode. Turn it on with /plan and every request is looked into with read-only tools first, then comes back as a plan card you approve, edit or reject; approved steps run one at a time and pause when one fails. It is offered on a model only once that model passes the plan-mode test.
• Agents on a real desktop. Screen agents get an Ubuntu desktop of their own, see it with a vision model, click with closed-loop targeting, and work in a real browser while you watch from any page. ▶ Ep 4
• Parallel coding agents. A swarm of coding agents, each in its own git worktree, merged back in dependency order. Runs on Claude Code or fully locally through Ollama.
• MCP, both ways. An MCP server that plugs into Claude Code, Cursor, Codex, Gemini CLI, Zed and more with one command, and an MCP client so chat can use tools from other servers. ▶ Ep 21 · ▶ Ep 16
• Video generation. 13 local video models across five families (Wan 2.2, LTX, HunyuanVideo, CogVideoX, and MiniMax H3 with its own soundtrack), video LoRAs, a batch queue, frame interpolation, and one click into ComfyUI. ▶ Ep 6
• LoRA training and consistent characters. Train character, environment and prop LoRAs from a few reference photos on your own GPU, import LoRAs trained elsewhere, and keep the same character across images, video, music videos and Film Crew productions. Details
• Images. Z-Image, FLUX.1, FLUX 3 Klein, Krea 2, MOSS-Image, Qwen-Image 2.1, SDXL and Sana Sprint generation (Sana Sprint runs even on 8 GB cards), an Image Edit page that keeps a face across new scenes and scores the likeness, photo edits by instruction, inpainting and outpainting, and 4K/8K upscaling. ▶ Ep 5
• Music and voice. Full songs with vocals, sound effects, three neural text-to-speech engines, consent-gated voice cloning, and hands-free voice chat. ▶ Ep 7
• Directors. A beat-synced music-video director, a five-role Film Crew that turns a logline into a finished video, and a built-in video editor. ▶ Ep 8 · ▶ Ep 9
• Self-improvement. Ask how part of Guaardvark works and it answers from its own source with checked file:line links. A change it makes to itself is planned, checked, undone when a check fails, and one button reverts it. The System Map draws the whole codebase from its real imports. ▶ Ep 11 · ▶ Ep 14
• Command center. A GPU orchestrator, a plugin manager with a live VRAM budget, jobs and scheduling, schema-aware backups, and a multi-machine Interconnector. ▶ Ep 12
• Terminal and API. The guaardvark CLI with an interactive REPL, plus REST and Socket.IO. ▶ Ep 20

The complete, enumerated list of models, tools, plugins and pages is in [CAPABILITIES.md](CAPABILITIES.md).

---

Chat with your own files (local RAG)

Upload documents, build a knowledge base per project, and ask questions. Answers are grounded in your files and cite where they came from.

• Hybrid retrieval. Postgres full-text keyword search and vector search, fused with a per-query weighting that leans toward keywords for identifier-like queries and toward meaning for prose.
• Cross-encoder reranking. A reranker (one Install in Settings) reads the query and passage together and reorders the candidates. It is admitted against free VRAM and falls back to CPU rather than competing with image or video generation.
• Layout-aware parsing. PDF, DOCX and PPTX are parsed for reading order and section headers, and PDF passages carry the page they came from. (Scanned documents need OCR, which is not installed by default; they say so rather than indexing as empty.)
• Grounded citations. Every chunk carries its source file and section breadcrumb (and, for PDFs, the page), written into the embedded text as well, so a passage lifted from the middle of a document still says where it came from.
• Smart chunking. Code gets AST-informed chunks, so a function stays one chunk; prose gets semantic splitting.
• Postgres-backed vector store. Embeddings live in pgvector next to the rest of your data, with an ANN index and a persisted full-text index, covered by the same backups.
• Retrieval that shows its work. Live retrieval tests show the actual chunks and scores behind an answer (Episode 3), and every query can return a trace of which retrieval legs ran, so a degraded answer can be told apart from a bad one.
• More: index profiles (the same documents projected more than one way), corpus-level summaries for "what are the themes across all of this", a choice of embedding models from 300M to 4B+, entity extraction, and per-project isolation.

Autoresearch tunes retrieval while you sleep. An overnight loop proposes changes to chunking and retrieval settings, scores them with an LLM-as-judge harness, keeps the wins and reverts the regressions, bounded by a wall-clock budget, a run ledger and a circuit breaker. The morning report says what changed and why.

One chat box, three speeds. Every message goes through a three-tier router that picks the fastest path to the right answer (Episode 2):

Desktop and vision tools are only offered to the model while the agent screen is active.

Agents that use a real desktop and browser

Guaardvark's agents control a real Ubuntu desktop (Xvfb and XFCE on display :99): the same Applications menu, desktop icons and taskbar you would see if you connected to the box over VNC. They see the screen through a vision model, move the mouse, click, type, browse, and check their own work. The desktop is isolated from yours: its own XDG_DESKTOP_DIR and XDG_CONFIG_HOME, so the agent's files and settings never collide with your session.

• One vision brain. Gemma4 sees the screen, decides the next action and returns click coordinates in a single inference call.
• Closed-loop targeting. The servo moves, checks where the cursor actually landed, and corrects until it is on the target, rejecting targets the model cannot really see.
• Live reasoning. Every step (act