Set up AI¶
Goal: ask the AI assistant about the active context and have it act for you through the platform's operations (workers, feeds, triggers, scenarios, strategies, state, topics).
Prerequisites
- Signed in to https://app.virtufin.com (Getting started), on the context you want to use (providers are kept per context: set them up on each).
- An API key for an OpenAI-compatible Chat Completions endpoint (OpenAI, Anthropic,
OpenRouter, Mistral, or any other with an
https://public base URL) — unless the deployment has an org provider, which works without a key of your own. - No admin role is needed to set up AI or to ask; the assistant writes only for admins.
Steps¶
- Open Settings > AI (the Settings tab, then AI in the sidebar; route
/settings/ai). The page names the active context: Your AI providers for the assistant on<context>. - Add a provider. Under Provider keys, press Add OpenAI (or Add Anthropic, Add OpenRouter, Add Mistral; Add other for any other base URL). A row appears with Name and Base URL filled in.
- Paste the key into API key and press Save keys. A notice says Provider keys saved; the key field now shows set (…abcd); type to replace.
- Pick a model. Under Model, the table lists the models your key reaches, with
Input / 1M, Output / 1M (USD per million tokens, from LiteLLM's price list) and
Context. Type in Filter models (e.g.
gpt mini) and press Use on a row. The line above says Using<provider>·<model>— …; the row shows In use.- A model the provider does not list: pick the provider in the drop-down below the table, type the id in or type a model id, press Use.
- Open the assistant: press AI in the header (any view). The panel opens on the right.
- Ask, e.g. Which workers are stopped?, and press
Enter(Shift+Enterfor a new line). - Confirm writes. As an admin, ask it to do something (stop the worker
binance-orderbook). On a protected context the dialog Write to the protected context<name>? names the operation (e.g. workers.stop (Stop a worker)): Yes, send it or No. On an unprotected context the write goes out at once.
New in the panel starts a fresh conversation; closing the panel also ends it.
What you should see¶
- Each tool call in the panel as it runs: ✓ or ✗, the operation (
workers.list), its input; then the answer. - Writes the assistant made in Infra > Changes (
done,failedwith the reason, ordeclined).
Troubleshooting¶
| Symptom | Cause and fix |
|---|---|
| no AI provider: add one in Settings > AI | No provider with a key on this context, and no org provider. Do steps 1–3 on this context. |
no model for <name>: choose one in Settings > AI |
The default provider has no model: step 4. |
<name>: base URL must be https / is a private or local address / does not resolve |
The server only calls public https:// URLs. Fix the base URL. |
| provider answered 401: … | The key is wrong or revoked: type a new one, Save keys. |
<provider>: answered 404 … under Model |
The provider has no GET /models: type the model id by hand (step 4). |
| Prices unavailable: … | The server could not load LiteLLM's price list (VIRTUFIN_APP_LLM_PRICES_URL); models still work, prices show –. |
| The assistant only reads | You are not an admin (the badge says read only): it gets read tools only. |
A model price shows – |
LiteLLM's list does not name that model id. |
| Saved keys stopped working for everyone | The server's session key was rotated: enter the keys again. |
More: AI.