AI assistant
The assistant is the chat interface in the floating mascot panel. It's a language model with crypto-aware tools — ask for a price, an address, a quote, a transfer, and it does the right thing.
Three places to run the model
Pick one in Settings → AI. Switch any time without losing your conversation.
MLX — on your Mac (default)
The bundled model is nani-qwen-3.5-2B, a 2-billion-parameter model fine-tuned for crypto tool calling. It's quantized to 4-bit, runs on the M-series GPU, and uses about 1.5 GB of RAM during inference.
Best for everyday use. Free, fast, private. The first run downloads the model from HuggingFace (~2 GB) — after that everything is local.
Requires Apple Silicon (M1 or newer).
OpenRouter — bigger cloud models
OpenRouter gives you Claude, GPT-4, Llama, Mistral, and others through one endpoint. Use it when you want a stronger model than the local one — long reasoning chains, longer context windows, frontier capabilities.
Pay per token, billed by OpenRouter. Add your API key in Settings; Nani stores it in the macOS Keychain.
Ollama — your own local server
If you already run Ollama and want to swap in a different local model (a 7B Llama, say), point Nani at your Ollama URL in Settings. The default is http://localhost:11434 with model name nani — you can use any pulled model.
Chatting
Conversations save automatically. Each conversation has a title generated from your first message and a list of all the messages in it. You can have several open and switch between them from the chat window.
History persists across launches.
Tool calls
When you ask something the assistant can answer with a tool — "what's the gas price?", "resolve vitalik.eth", "send 0.05 ETH to alice" — it picks the right tool, runs it, and continues its reply with the result.
You'll see a brief "thinking" state while the tool runs, then the answer. For anything that touches your wallet, the assistant pauses and Nani shows you the parsed transaction. You approve with Touch ID — the assistant cannot sign by itself.
The full tool catalog is in Tools.
What to ask
The local model is trained for crypto tool selection. It excels at things like:
- "What's the gas price on Base?"
- "Resolve
vitalik.eth" - "What's ETH at right now?"
- "Show my USDC balance"
- "Send 0.01 ETH to
alice.eth" - "Quote a swap of 100 USDC for ETH"
- "Is
0x…a known scammer?"
For long-form writing, code generation, or general chat, switch to OpenRouter or Ollama with a larger model.
Privacy
With MLX selected, your prompts and the model's replies stay on your Mac. Nothing about the chat goes to the network.
With OpenRouter selected, your prompts go to OpenRouter and from there to your chosen model provider. With Ollama selected, your prompts go to whichever server you've pointed Nani at.
In every case, the assistant only ever sees your wallet address — never the private key.