What leaves your machine, and what does not.

With a local model, nothing leaves

When the active model is served by Ollama, LM Studio, or a local GGUF file, inference happens on your machine over localhost. Your prompt, the files the agent reads, and the code it generates never travel anywhere. There is no request for us to log, because there is no request.

You can verify this rather than trust it: run Osmium with a local model and watch outbound traffic. Inference generates none.

With a cloud model, your prompt goes to that provider

Choosing a cloud model is an explicit, per-model decision. When you make it, your prompt and the context you attach are sent to that provider so they can answer. There is no silent fallback that ships your code off-machine — if no cloud plan is active, the app switches to the best local model instead.

Where your data lives

• Sessions, chat history, and workspace state live in a SQLite database on your machine.
• There is no server-side transcript of your conversations.
• Account and billing data live with the Hub, and cover identity and plan only — not your code.

API keys

There is no API key field anywhere in the product. Cloud providers are either proxied on your behalf or configured by an administrator, which means individual users cannot leak a key they never had. If a model has no key behind it, the app says to ask an admin rather than offering a paste box.

Controls

• A kill switch freezes outbound calls and spend.
• Cloud spend has a hard stop at your balance rather than overrunning it.
• Integrations stay in a dry-run mode until they are explicitly switched live.

Honest limitations

• The Windows build is not Authenticode-signed yet, so SmartScreen will warn on first run.
• We have no third-party security certification such as SOC 2 at this time.
• The agent runs commands on your machine when you ask it to. Review what it proposes, particularly anything destructive.