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.
Anthropic Claude¶
Get a key from the Anthropic Console. Free tier plus prepaid credit, both work.
Default model: claude-haiku-4-5-20251001 (fast, cheap, strong tool-use).
Picker also includes claude-sonnet-5 (smarter, roughly 3-5× the cost) and claude-opus-4-8 (smartest, roughly 15× the cost). 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 across all three apps.
OpenAI¶
Get a key from the OpenAI platform dashboard. Prepaid credit required (no free tier).
Default model: gpt-4o-mini (fast, cheap).
Picker also includes gpt-4o (smarter, roughly 15× the cost of mini). If you want a newer model like gpt-4.1 or o3, use Custom… in the picker and type the ID, or set AI_MODEL directly.
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-2.5-flash (fast, cheap, strong).
Picker also includes gemini-2.5-flash-lite (cheapest) and gemini-2.5-pro (smarter). Any earlier Gemini model (1.5-flash, 1.5-pro, 2.0-flash, 2.0-flash-lite) has been retired by Google and will 404 if called directly; the server auto-remaps saved configs pointing at those IDs to the current 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.