Install it
On macOS, through Homebrew:
On Windows and Linux, take the installer from the download page. All three platforms run the same runtime, and the same employees run on any of them.
Give it a model
Nebo does not ship a model. It talks to whichever provider you point it at, and you choose per employee later if you want different ones for different jobs.
| Provider | You need |
|---|---|
openai | An API key |
anthropic | An API key |
google | An API key |
deepseek | An API key |
ollama | Nothing. It runs on your own machine. |
Point Nebo at a local model and no part of a conversation leaves the machine, not even the model call. Slower and less capable than a frontier model, and the right choice when the material genuinely cannot go anywhere.
Or route through NeboAI instead and skip key management entirely. That is the one part that is not free, and it is optional.
Confirm it works
Open the app and ask it something you can check. You are not testing the model, you are testing that the provider is wired up. If the answer comes back, you are done.
Almost always the key. An invalid key usually fails silently rather than loudly, so re-paste it before looking anywhere else. If you are on Ollama, check the model is actually pulled and the service is running.
What you now have
A runtime, and nothing hired yet. Conversations, files and memory live on your disk, in SQLite, wherever the app stores its data. Nothing has been sent anywhere except the model calls you just made.
Next: hire something, or connect a mailbox for it to work from.