Cloud providers¶
Three cloud providers are wired in: Anthropic Claude, OpenAI, and Google Gemini. This page covers where to get an API key and which model each app defaults to.
The setup flow is the same for all three: Settings → Trace AI, pick the provider, paste the key, pick a model, Save. For the env-locked variant (one config for everyone), see Setting up Trace.
The model lists were checked against each vendor's model and deprecation pages on 2026-10-11. NutriTrace 1.4.0, NoteTrace 1.1.0 and LiftTrace 1.4.0 and later have the lists below; CookTrace gets them in its next release (until then its Claude default is claude-haiku-4-5-20251001 and its Gemini default is gemini-3.6-flash). A model you already chose keeps working: one that isn't in the picker any more shows as a custom model.
Anthropic Claude¶
Get a key from the Anthropic Console. Free tier plus prepaid credit, both work.
Default model: claude-haiku-5-5 (fast, cheapest, strong tool use).
Picker also includes claude-sonnet-5-5 (balanced), claude-opus-5-5 (smarter) and claude-fable-5-1 (most capable), plus the previous claude-sonnet-5, claude-opus-5, claude-fable-5 and claude-opus-4-8. Sonnet is the sweet spot for most people once cost stops being the top concern.
Env config:
Wire endpoint is https://api.anthropic.com/v1/messages with header anthropic-version: 2023-06-01. Claude has the richest native tool use, which is why it's the default in every app.
OpenAI¶
Get a key from the OpenAI platform dashboard. Prepaid credit required (no free tier).
Default model: gpt-5.6-luna (fast, cheap).
Picker also includes gpt-6-luna (fast, cheapest), gpt-5.6-terra (balanced) and gpt-5.6-sol (most capable), plus the previous gpt-4o-mini and gpt-4o. GPT-6.1 Sol and GPT-6 Astra aren't listed: Trace calls tools through Chat Completions, which those two don't support. A saved gpt-5.6 is sent as gpt-5.6-sol, the model it stands for.
Env config:
Wire endpoint is https://api.openai.com/v1/chat/completions. The same code path powers oai-compat against Ollama and friends; only the base URL differs.
Google Gemini¶
Get a key from Google AI Studio. There's a generous free tier, which makes Gemini a good "just try Trace" option.
Default model: gemini-3.8-flash (fast).
Picker also includes gemini-3.1-flash-lite (cheapest), gemini-3.5-flash-lite (cheap) and gemini-3.1-pro-preview (most capable), plus the previous gemini-3.6-flash. Gemini 2.5 is open only to Google accounts that used it before, so it's no longer listed; a saved 2.5 model keeps working. A saved gemini-3.1-pro (an ID that never existed) is sent as gemini-3.1-pro-preview, and models Google has shut down are sent as the default. See Model list and retirement remap.
Env config:
Wire endpoint is https://generativelanguage.googleapis.com/v1beta/models/<model>:generateContent. Function-calling round-trips through OpenAI-shaped adapters on the server proxy, so Trace's full tool set works on Gemini too.
Photo uploads
Only NutriTrace routes meal photos through the AI channel by default (for propose_food and propose_quick_calories). LiftTrace and CookTrace don't send images by default, but the picker will accept an attachment if you post one manually. Every provider is happy with base64 data: URIs; the server proxy reshapes them per provider.
Rate limits and cost¶
Trace itself is rate-limited to 30 requests / 60s per user (see Setting up Trace). Beyond that, provider-side limits apply. All three providers surface a per-key spend dashboard; check yours before opening Trace up to a household.