AI assistant
Ask NANI to look up wallet information, explain a transaction or help prepare your next action. You can also go straight to Wallet for Receive, Send, Swap and Bridge.
Choose where your messages go
Open Settings → AI provider:
| Choice | How it works |
|---|---|
| NANI | The initial provider. A sharing explanation appears before your first hosted message. |
| Your own provider | Add OpenAI, Anthropic or a custom OpenAI-compatible endpoint with a model and credentials. Requests go directly to that service; its policies, limits and fees apply. |
| On-device model | Download Bonsai 2 · 4B or add a supported public MLX text model. Inference runs on your iPhone and needs sufficient memory and storage. |
NANI and OpenRouter receive messages, attached images and context needed to answer, including wallet information used in replies. OpenRouter routes requests to model providers. A provider you configure receives the requests sent to it. Keep sensitive information out of conversations with a remote provider.
Wallet lookups, research tools, model downloads and transactions use network services, including when you select on-device AI. Tool results can become part of the conversation.
NANI’s saved usage and allowance appear in AI provider settings. Use Refresh allowance for a new reading. Last checked identifies an older reading; a failed refresh does not create a new allowance or prove that the remaining count is current.
Switch models and manage downloads
Tap the model pill in the bottom row of the chat composer to open Select model. It lists NANI, saved providers and available local models. Drag up to expand it; More models opens Settings → AI provider. Switching preserves your draft. Model and mode changes are unavailable while a reply is running.
Bonsai 2 · 4B is the current built-in local model and supports text and images. It uses about 1.87 GB of weights, plus working memory and setup space. Downloads and preparation require an explicit action. If a download pauses when you leave the app, return and resume it. A previously selected NANI 2B model can remain usable, but it is no longer the offered choice for a new local-model selection.
To add your own local model, choose Add model in provider settings, paste a public Hugging Face MLX repository or revision URL, and tap Check model. The check shows compatibility and download size; adding selects the model, and Download is separate. Saved revisions do not update automatically. Custom models are text-only; private/gated repositories and GGUF files are not supported. A compatibility check does not guarantee good replies or sufficient memory on your iPhone.
Normal and Unfiltered replies
Bonsai starts with Normal replies. Its Unfiltered switch is in Select model. First use can prepare a profile locally and take several minutes. Tap Done after setup to enable the mode; pausing, closing setup or an error leaves Normal selected.
Unfiltered may produce fewer refusals and less reliable answers. It changes the next reply, not Private Chat, export inclusion or wallet approval. Other local models do not have this switch.
Local send and swap commands
Type /send or /swap as the first word of a draft to show the matching quick action. These commands use local, editable panels inside chat without inference or AI sharing consent. You can also prefill a panel:
/send 0.01 ETH to robert.eth
/swap 0.1 ETH to USDC
Check the account, network and selected asset contract, then complete any missing fields. The panel still ends in the independent wallet review with fees, simulation and approval. Cancel returns to your draft. Quick actions do not appear for an empty draft, partial commands or unknown commands.
Try a clear request
Include the asset, amount and network when they matter:
- “What is my USDC balance on Base?”
- “Explain the fees for this transaction.”
- “Research this token contract on Base.”
- “Prepare a transfer of 0.01 ETH to this address on Ethereum.”
Capabilities depend on the selected network and available tools. Check the source, freshness and network of results. Ask for clarification when information is missing or uncertain.
For an executable swap comparison, use Wallet → Swap or the local /swap panel. Saved chat figures are not live trading data. A quote can expire and require a fresh request.
Instructions, tools and interrupted replies
In Settings → AI provider, use the info button beside a local model to open Instructions, or open a saved provider’s editor. Each local model or direct-provider profile can have its own prompt and Use tools choice. A custom prompt replaces its default. NANI’s hosted instructions and tools are fixed. Turning tools off does not disable the native wallet controls.
If a reply is interrupted, the chat keeps its partial answer and offers the appropriate retry or recovery action. Check any existing proposal or transaction status before repeating a request. Starting a fresh app process opens a new chat rather than automatically resuming the last one; saved conversations remain in history.
Review before approving
A proposal opens a wallet review. Read the recipient, asset, amount, network, permissions, fees and warnings before you approve. Model replies and risk checks can be wrong or incomplete. Authentication confirms your authorization; it does not establish that a contract or recipient is safe.
See Assistant tools for how requests become results and proposals.
Voice and images
Dictation is transcribed on the iPhone. Review the text before sending. You can select up to three images with the photo picker. Bonsai processes supported images locally; compatible remote providers receive them after sharing confirmation. Text-only models cannot use image attachments. Images are saved with the chat, and image chats are marked private by default. Watch voice has separate provider settings; see setup.
Long-press a message for Copy or Select text. Assistant responses also offer Report response, which lets you review the content before opening an email draft or share sheet; it sends nothing automatically.
Private chats
A private chat stays in local history and is excluded from exports unless you include private chats. Messages you send still go to your selected provider. The wallet keeps signing secrets separate from assistant requests, so never paste recovery phrases, private keys or API keys into chat.
Local-model chats and image chats start private automatically. For saved local activity, select an on-device model and choose New on-device activity chat from conversation actions. That chat uses saved activity tools and blocks remote-model sends; ordinary private chats can still use remote providers.
Read the privacy policy and privacy choices for sharing, retention and deletion.