Install the agent
Run one command on each machine you want to reach later: a desktop with a GPU, a Mac Studio, a box in a closet.
The installer sets the agent up as a service that starts with the machine, offers to install an inference engine, and pairs the machine to your account. Pairing needs an invited account while Qadenza is in beta.
The agent and your engines
The agent is a single program. It runs your engines, watches the hardware, and answers requests for models. It does not replace the engines:
- Ollama for the simplest model library.
- llama.cpp when you want control: GGUF files from Hugging Face, quantization choice, and how a model is split across GPUs.
The agent's own API listens on 127.0.0.1 only. Nothing on your machine is opened to the internet or your network.
Connecting from anywhere
To reach a machine behind a home router or an office firewall, the agent dials out to the Qadenza relay and keeps one encrypted connection open. When you open the dashboard or call the API from elsewhere, your request goes to the relay, down that connection to your machine, and the reply comes back the same way. That is why there is no port forwarding, dynamic DNS or VPN client to set up.
The relay passes requests and replies through and stores none of them. Security lists exactly what it sees, every host the agent connects to, and what telemetry sends.
If you would rather run the relay yourself, point the agent at it:
Reaching your models
Once a machine is paired, its models are available four ways:
- The dashboard, in any browser, and installable on your phone. Chat, load and unload models, and watch CPU, RAM, GPU temperature and power with an hour of history.
- VS Code, through the Qadenza extension: a coding agent that reads your project and proposes edits, running on your own model.
- MCP, both ways. Give your models tools such as Filesystem, Git and GitHub from a gallery, or add your models to Claude Code and other agents with
claude mcp add qadenza -- qadenza mcp. - An OpenAI-compatible API. Anything that takes a custom base URL works: the OpenAI SDKs, Cursor, Continue, Open WebUI, LangChain, curl. Tokens are created in Settings and can be revoked any time.
Choosing and running models
- Fit, before you download. Each model is checked against the memory that's free on that machine right now, so you don't pull 40 GB to find out it won't load.
- Bigger than one GPU. llama.cpp can split a model across every card in the box, and you choose how.
- Any GGUF. Sharded GGUF downloads work, and safetensors models can be converted to GGUF on the machine.
- Heat. If a GPU reaches its thermal limit, the agent pauses generation instead of ending the engine.
Removing it
qadenza unpair removes a machine from your account. qadenza uninstall removes the service and the agent.