What leaves your machine, spelled out.
Your models run on your hardware. When you reach them from elsewhere, traffic crosses our relay. This page lists exactly what that involves and every host the agent connects to.
What passes through our servers, and what doesn't.
The agent dials out to relay.qadenza.ai and keeps one encrypted connection open. A request you make from elsewhere goes to the relay, down that connection to your machine, and the reply comes back the same way.
- IN TRANSITYour requests and the model's replies, between your browser or tools and your machine. Encrypted on the way.
- STOREDNone of it. We never store prompts, replies or files. We keep your account, the list of machines you've paired, and standard server logs.
- TELEMETRYOn by default: daily speed figures per model, like tokens per second and peak GPU temperature. Never prompts, replies, files or model names. One switch in Organization settings turns it off.
- OPEN PORTSNone. Your machine dials out to the relay; nothing on it listens to the internet.
Everything the agent connects to
Outbound only. Check it against your firewall logs.
- relay.qadenza.ai
- The encrypted tunnel that lets you reach this machine.
- api.qadenza.ai
- Sign-in, pairing, agent updates, and telemetry if it's on.
- huggingface.co · modelscope.cn · civitai.com · registry.ollama.ai
- Only when you download a model.
- github.com
- Only when installing or updating an engine or runtime, such as llama.cpp.
How it's put together.
Outbound only
The agent opens one outbound connection. No port forwarding, no inbound firewall rules.
Encrypted in transit
Requests and replies are encrypted between your device, the relay and your machine.
Nothing stored
The relay passes requests and replies through. It doesn't keep or read them.
Loopback-only agent
The agent API listens on 127.0.0.1 only. Nothing on your machine faces the network.
Revocable tokens
Each tool gets its own API token, revocable at any time from Settings.
Credentials protected
Passwords are stored only hashed. Git access tokens are stored encrypted.
Telemetry off switch
Speed figures are on by default and off with one switch, effective immediately.
Your own relay
Point agents at a relay you run with qadenza config set relay-url.
What the agent reports, and what it never does.
Model recommendations depend on knowing how fast models run on real hardware. That is the only reason telemetry exists. It's on by default; one switch under Organization settings turns it off and deletes anything not yet sent.
Sent, summarized per day
- Speed
- Tokens per second, time to first token, power draw, peak GPU temperature. Grouped into ranges, not per request.
- Setup
- Engine and version, quantization, file size, and a one-way hash of the model's public name.
- Hardware
- OS, CPU and GPU models, rounded memory sizes, driver version, agent version.
Never sent: prompts, replies, files, Workspace contents, model names, hostnames or file paths.
No internet? No problem.
For teams that can't let data leave the building, we deploy the whole stack inside your network: agents, dashboard and relay. The full stack runs with zero outbound connections, so the list of hosts above is empty.
- Runs on isolated networks with no outbound connection
- Models and updates delivered offline, on your schedule
- Same chat, model manager and VS Code extension as online
- Your hardware, your network, your data. Nothing leaves.
Questions about your setup?
Email us at qadenzabeta@mlyth.org, or request an invite and try it on your own hardware.