Last updated: March 15, 2026
MAUDE runs on Linux with CUDA. It was built for the NVIDIA DGX Spark (Grace Blackwell, 128 GB unified memory) but works on any Linux machine with an NVIDIA GPU and enough VRAM to run a local model.
Step 1
Make sure you have the following installed:
nvidia-smi workingStep 2
git clone https://github.com/mboard8070/terminal-llm.git
cd terminal-llm
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
playwright install chromiumStep 3
The setup script clones llama.cpp, builds it with CUDA support, and downloads the Nemotron 30B model (~38 GB).
./setup_local.shThis will take a while depending on your internet speed and whether cmake needs to compile CUDA kernels for your GPU architecture.
Step 4 (Optional)
If you want MAUDE to analyze images and screenshots, install Ollama and pull the LLaVA vision model:
curl -fsSL https://ollama.com/install.sh | sh
ollama pull llava:13bStep 5
Create a variables.env file in the project root. You only need the keys for the providers you plan to use:
# Required for cloud models (at least one)
MISTRAL_API_KEY=your-key-here # console.mistral.ai
CODESTRAL_API_KEY=your-key-here # same Mistral console (separate key)
# Optional
CLAUDE_API_KEY=your-key-here # console.anthropic.com
OPEN_ROUTER_API_KEY=your-key-here # openrouter.ai/keysTip: You can also set keys from inside MAUDE with /keys set mistral YOUR_KEY. Keys are stored in ~/.config/maude/keys.json.
Step 6 (Optional)
If you want to access MAUDE from other machines (Mac, PC, phone), install Tailscale on both the server and your client devices:
curl -fsSL https://tailscale.com/install.sh | sh
sudo tailscale upFor HTTPS access (required by the mobile app), generate SSL certificates and place them at certs/cert.pem and certs/key.pem. The gateway auto-enables HTTPS when these files exist.
Step 7
The single launcher starts all services and drops you into the MAUDE chat:
./maudeThis starts:
# Other commands
./maude stop # Stop all services
./maude status # Show what's running
./maude -n # Start services only (no TUI)Step 8 (Optional)
The lightweight CLI client connects to your MAUDE server over Tailscale. Install it with pip:
pip install maude-client
maude-client --server http://<your-tailscale-ip>:30080The client supports all models, tool execution traces, conversation history, and cross-machine task dispatch.
Step 9 (Optional)
To use Gmail, Drive, Sheets, Calendar, Contacts, Slides, and YouTube tools, you need a Google Cloud project with OAuth 2.0 credentials:
credentials.json to the project rootVerification
Once MAUDE is running, try these commands in the chat:
# Check available models
/model switch mistral
# Test tool execution
what files are in my home directory?
# Test web search
search for the latest NVIDIA news
# Check system status
./maude statusYou should see tool call traces (╭─ tool_name / ╰─ result) as MAUDE executes commands on your behalf.