airaceGP reads your live telemetry, works out what matters right now and tells you over the radio. Short calls, real numbers, answers when you ask. Built for Assetto Corsa Competizione, with F1 25 support.
Driver“Where's the car ahead?”
McLaren ahead, 1.4, closing 0.3 a lap.
That is the whole call. Car make, the gap, the trend. Example wording, not a guarantee of what an AI model will say. Missing data limits the answer.
airaceGP works out the race state itself, from the telemetry. The model gets a short, finished picture and only has to phrase it.
App
Computes fuel, pressures, gaps, trends and what to say first.
AI
Phrases it like an engineer, in one short call.
What matters right now, so the important call goes first.
Enough to the end, or how much short.
Each corner labelled. Which one is out, and which way.
By car make, with the gap and whether it is closing.
Slower cars you are about to meet.
Called when they change.
Driver memory keeps the last 8 sessions per game, track and car in driver_memory.json on your PC. The file is never uploaded.
ACC only for now. Remember me between sessions is on by default; turn it off to stop recording and recall, or use Forget everything. Remembered summaries can be sent to your chosen AI provider.
Ask about stops or fuel in an ACC race: the strategy planner works out when to stop, how much to add and whether the fuel reaches the flag, in laps, never litres, with a 1-lap reserve.
Saving suggestions are bounded. Tyre wear, traffic, undercuts and safety cars are not modelled; ACC pit-window units are still unverified.
Setup advice compares measured hot tyre pressures with the dry or wet pressure window and gives the change in psi per corner, using recent sessions or the live frame when memory is off.
This is not a setup builder. Brakes, camber, toe, aero, ride height, dampers, bars and gearing are not measured for setup advice, so it gives no values for them and no target cold pressure.
Pick which providers the engineer asks, in what order, and whether it moves to the next one when a call fails. No single provider is required. Onboarding builds the chain from the keys you enter.
Ollama Cloud
gemma4:31b · free tier · recommended
Asked firstGroq
your key
If 01 failsOllama · local
on your PC · no key
If 02 failsWhen a provider fails, the app says which problem it was and what to do about it. Every failure has a code and a fix, and the next provider in your chain takes over. See the error codes.
If every AI provider fails, a driver-question fallback uses the app's facts or says plainly that there is no answer.
A Windows desktop app with a dark timing-screen dashboard. Open it on a phone over your LAN and keep it next to the wheel.
The headline, the cars around you, fuel, pressures and the radio log, updating as you drive.
Track map, lap times, telemetry traces and sector trends against your best.
A page for the corners of the lap, to see where the time goes.
Talkative or quiet, short calls or longer ones, spaced out or back to back. Settings let you control the radio. See every setting.
Where the engineer was built. Fuel to the flag, pressures, damage, and the cars around you called by make.
Same engineer, same radio, fed by the game's UDP telemetry.
The engineer's voice and the app speak your language. Translation runs on your PC.
Bind a button on your wheel (a detected controller input), a D-pad direction or a key. Hold it to talk, or tap it once. Or let voice detection pick up your question.
faster-whisper on your PC, or Groq Whisper in the cloud.
Text to speech runs locally with Piper. A clear British voice by default, a male American voice one click away.
A desktop app for Windows. Public download coming soon.
Onboarding asks which providers you use and builds your model chain. Keys stay on your PC in a local .env file.
ACC shared memory, plus the Broadcasting API for live timing. For F1 25, turn on UDP telemetry.
The engineer calls what matters. Talk to it when you need something.
Stuck? The docs cover setup, providers and settings. Clear error codes, with a fix for each.
These tools make the remaining hardware checks visible. Developer-page test guide.
The Spotter is ACC only, beta and off by default in Settings, AI, Spotter.
A dedicated fast rivals feed drives car left, car right, three wide and clear calls. A call cuts the engineer off; it can resume afterwards, although a routine line that waits too long is dropped.
Tested in simulation, not verified on a real PC: opponent updates, left/right in traffic and audio timing. Run the ACC update-rate test first, and do not enable the Spotter in a race until it shows the feed is good enough.
The Wheel and buttons card shows one light per detected button, D-pad direction and held key, lit while it is held, to check whether a button can be bound for push to talk. Hardware compatibility is unverified on a real PC; try your device in this view.
The Developer page also has the ACC update-rate test, a Spotter live view and a cut-and-resume demo. The game tools and wheel lights still need a real-PC test.
Le Mans Ultimate is not supported yet. Its game integration is fixture-only and unverified on a real PC; the recorder exists for testers.
The LMU recorder lets testers add notes and stop to write a report zip. It is a data-gathering tool, not a working race engineer for LMU.
Straight answers. More in the docs.
Nothing. airaceGP is free, forever: no subscription, no paid tier, no ads. The AI is your choice: Ollama Cloud has a free tier, local Ollama needs no key, and paid providers bill you directly for what you use. That money goes to them, not to us.
Any one of Ollama Cloud, Groq, Google Gemini, OpenRouter, Anthropic, an OpenAI-compatible endpoint, or Ollama on your PC. Add more to build a chain with fallback.
Into a local .env file on your PC. Race context is sent only to the providers in your chain.
The telemetry analysis, translation, Piper speech and, if you pick it, faster-whisper speech recognition. The phrasing is done by the model you choose, which can also be Ollama on your PC.
Assetto Corsa Competizione, through shared memory and the optional Broadcasting API. F1 25, through its UDP telemetry.
Yes. The dashboard works in a phone browser over your LAN.
The app computes measured facts before asking the AI to phrase them. Answers can still be wrong. New features have specific limits: driver-memory wording is not live-model verified, and the Spotter is beta and unverified on a real PC.
Yes. In Settings, Audio, Push to talk, press Bind button and then the button you want: a detected wheel button, a D-pad direction, or a key. Hold it while you talk, or tap it once and speak. Pressing it while the engineer talks cuts him off, like a real radio. Check your input in the Developer page; hardware compatibility is unverified on a real PC.
Driver memory saves measured session history on your PC when enabled. The file stays local; a remembered summary can enter your chosen AI requests. The app has no account, no analytics and no tracking. Race context goes only to the AI providers you put in your chain, and your spoken questions to Groq only if you pick Groq for speech recognition. This website sets no cookies either.
No. airaceGP is an unofficial fan-made tool, not affiliated with the makers of either game.
airaceGP is a small project built on a lot of open work. These are some of the people whose code, voices, translations and track maps make it possible. Full credits and notices are kept with the app.
Piper by Michael Hansen and the Rhasspy project, and piper-plus by ayutaz. The default English voice, Cori, by Bryce Beattie from public-domain LibriVox recordings; the male English voice trained on the LibriTTS corpus (Zen et al., CC BY 4.0). Voice datasets by Thorsten Mueller (German), Carlos Fonseca (Spanish), czyzi0 (Polish), the Nabu Casa voice-datasets team (Portuguese, Czech), the SIWIS project (French).
Barlow and Barlow Condensed by Jeremy Tribby and the Barlow Project Authors, JetBrains Mono by JetBrains, all SIL Open Font License 1.1, served from this site.
Vosk by Alpha Cephei for the wake word. faster-whisper by SYSTRAN, riding on CTranslate2 by OpenNMT and SYSTRAN, with weights from OpenAI's Whisper. All MIT or Apache 2.0.
Argos Translate by Argos Open Technologies, with models trained on the OPUS parallel corpus collection curated by Joerg Tiedemann and the Helsinki-NLP group.
Traced from Wikimedia Commons maps by Will Pittenger, GabrielStella, HumanBodyPiloter5, Rumbin, Firkin, AlexJ, Ocdp, Sas1998, Dh16dh, Nasho3498, cBuckley, and ごひょううぺこ. CC0, CC BY 3.0, CC BY-SA 3.0 / 4.0.
Instrument Sans by the Instrument Sans Project Authors, Titillium Web by the Accademia di Belle Arti di Urbino, both SIL Open Font License 1.1. pywebview by Roman Sirokov, python-sounddevice by Matthias Geier, PortAudio by Ross Bencina and Phil Burk, NumPy contributors, ONNX Runtime by Microsoft.
Gemma by Google DeepMind is the default model, served through Ollama, OpenRouter or Groq, each under its own terms. None of them is affiliated with airaceGP.
ACC shared-memory documentation and Broadcasting SDK by Kunos Simulazioni. The F1 integration uses the EA/Codemasters UDP specification and packet layouts referenced from hotlaps.io (MIT).