airaceGP Docs

Choosing your AI

The app computes the analysis; the model you choose turns it into words. You bring the key, and you decide which providers are asked and in what order.

Get an Ollama Cloud key #

  1. Sign in or create an account at ollama.com.
  2. Open Settings → Keys on ollama.com and create an API key.
  3. Add it to airaceGP, in any one of these places:
    • the Ollama Cloud card in the first-run wizard;
    • Settings → Secrets, field OLLAMA_API_KEY, then Save secrets;
    • the .env file in the app folder:
      .env
      OLLAMA_API_KEY=your-ollama-key

A key saved in Settings works from the next AI call, with no restart. The free tier has usage limits. When you reach one, the app reports AIR-202; a second provider in your chain keeps the engineer talking.

The model chain #

The model chain is the list of providers the engineer asks, in order. It lives in Settings → AI → Model chain. There are two lists:

ListUsed forScope
Unprompted callsPeriodic check-ins and race events: everything the engineer says without being asked.Routine calls and events
Your questionsDriver questions in ACC and F1 25 races.Questions
  • The engineer asks the first entry that has its key. An entry without a key is skipped and marked no key.
  • Use fallbacks on: if a provider fails or is rate-limited, the next one is asked. Off: only the first provider with a key is asked. If every AI provider fails, the driver-question fallback uses measured facts or says there is no answer.
  • Up to six entries per list. ↑ and ↓ reorder, × removes, Reset to default goes back to the chain built from your keys.
  • Nothing outside these lists is ever called.
Settings, AI, Model chain editor. Unprompted calls: 1 Ollama Cloud gpt-oss:120b, 2 Groq openai/gpt-oss-120b marked no key and skipped. Your questions: 1 Ollama Cloud gemma4:31b, 2 Ollama Cloud gpt-oss:120b. Each row has up, down and remove buttons, with Add provider and Reset to default below each list, and the Use fallbacks switch on.
An example saved chain from an earlier app screenshot, not the current default. Here Groq has no key, so it is skipped until one is added.

When the questions list applies

It is used in ACC with the default engineer settings. If you move Call length away from 50, or turn off Say what's on the dash or Allow the engineer to stay quiet, the app switches to its full prompt and your questions use the unprompted list instead.

With an Ollama Cloud key and no saved chain, gemma4:31b is the default for unprompted calls as well as driver questions.

All providers #

ProviderKey in .envPreset modelsNotes
Ollama CloudOLLAMA_API_KEYgpt-oss:120b, gemma4:31b, gpt-oss:20bFree tier. Recommended.
GroqGROQ_API_KEYopenai/gpt-oss-120b, openai/gpt-oss-20bProvider limits can change. The same key enables Groq transcription.
Google GeminiGEMINI_API_KEYgemini-2.5-flash, gemini-2.5-flash-liteFree tier with a limited number of requests a day.
OpenRouterOPENROUTER_API_KEYopenai/gpt-oss-120b:freeFree models are shared and often congested.
AnthropicANTHROPIC_API_KEYclaude-haiku-4-5-20251001Paid.
Custom (OpenAI-compatible)CUSTOM_API_KEYyou name itAny /chat/completions endpoint. Set the URL and model in Settings → AI → Custom API endpoint.
Ollama on this PCnoneqwen3:8bNo key and no internet. Shares the GPU with the game.

Any model name can be typed into a chain entry; the presets are suggestions.

Other providers through Custom

The wizard's List more providers offers OpenAI, Together AI, Fireworks, DeepInfra, Cerebras, xAI, Mistral, DeepSeek and others. They all use the one Custom slot: one CUSTOM_API_KEY and one endpoint URL. Where the wizard knows a provider's URL it fills it in; otherwise enter it in Settings → AI → Custom API endpoint. A missing URL, model or key is AIR-209.

Ollama on your PC

Install Ollama from ollama.com, download a model, and add Ollama on this PC to your chain with that model's name. The app talks to it at localhost:11434.

Command prompt
ollama pull qwen3:8b

A local model runs on the same GPU as the game, so frame rate and answer speed both depend on your hardware. Problems here are the AIR-25x codes.

Where keys live #

  • Keys are stored in .env in the app folder, on your PC. They are never written to web_config.json.
  • Settings → Secrets edits that file. It shows only a masked preview of each key, and it works only on the PC running the app, never from a phone on your network.
  • The race brief goes only to the providers in your model chain.
  • Removing a key is reported as AIR-210; that provider is skipped until a key is added again.