Features
AI assistant
Bring your own key. The assistant can see your canvas, so it can design, extend, review or explain the architecture in front of you.
Setting it up
- 1Open the panelAssistant in the toolbar, then the Keys tab.
- 2Paste a keyAny one of the four providers is enough. Keys save as you type.
- 3Test itThe Test button makes one tiny request and reports the provider's own error if something is wrong.
- 4Pick the providerProvider buttons sit above the chat; a green dot means a key is stored.
Providers
| Provider | Default model | Cost |
|---|---|---|
| OpenAI | gpt-4o-mini | Usage-billed |
| Google Gemini | gemini-2.5-flash | Free tier available |
| Anthropic Claude | claude-opus-5 | Usage-billed |
| Groq | llama-3.3-70b-versatile | Free tier available |
What to ask it
- “Design a URL shortener that handles 10k writes/sec” — builds a diagram from scratch.
- “Review my current diagram and name its weak points” — reads the canvas and answers concretely.
- “Add a caching layer and a read replica to this design” — returns the whole updated diagram.
- “Explain the trade-offs between a queue and direct calls here” — plain conversation, no diagram.
Applying a generated diagram
Diagrams are never applied automatically. When one comes back you get three choices: Replace canvas, Add to canvas (appended below your existing work with fresh ids), or Discard. Replacing is undoable with ⌘/Ctrl+Z.
Privacy and safety
- Requests go straight from your browser to the provider. Nautilus has no backend and never sees your key, your prompts or your diagrams.
- Keys live in this browser's local storage. Anything with access to the origin could read them — use project-scoped keys with a spend limit and clear them when you're done.
- Model output is validated before it can reach the canvas: unknown fields, invalid colours, duplicate ids and edges to missing nodes are dropped.
- The conversation is kept locally so closing the panel doesn't lose it; New chat clears it.
Why bring-your-own-key
It is the only way to offer AI features on a free, backend-less tool without a shared quota or a paywall — and it means your prompts never pass through anyone else's server.