The race
engineer.
airaceGP listens to your live F1 telemetry over UDP, runs it through an LLM pipeline, and speaks engineer-style radio calls back to you in real time.
Calm, terse, factual, modelled on real F1 race engineers. One snapshot in, one line out. Silent until it has something worth saying. And you can talk back.
Completely free.
Zero subscription.
There's no hosted service and nothing to pay us. airaceGP is open source and runs entirely on your machine.
The catch
Bring your own key.
Plug in your own API auth. It runs on whatever provider you already pay for, or fully local for zero cost. You pay your provider directly. Nothing flows through us.
Your keys, your models, your data.
Telemetry in. Radio out.
Four moving parts, one calm voice in your ear.
Listen
Your game streams telemetry over UDP on port 20777. airaceGP catches every snapshot: speed, tyres, fuel, gaps, damage.
Think
Your heavyweight model runs periodically, thinks through the whole race, and writes a live memory: one section per topic. Gap, tyres, fuel, ERS, pit plan, strategy, damage, weather.
Answer
When you ask a question, a fast small model reads that memory and replies in sub-second. It never waits on the heavyweight, so your driver never waits either.
Speak
A Piper neural voice reads the call out loud, engineer-style. Silent until there's something worth saying.
Any model you want. Every stage.
Nothing is hard-coded and nothing is gated. The brain that thinks, the parrot that answers, the voice that speaks: each one is yours to pick, swap, or run fully local.
The heavyweight that thinks
Runs on ticks and race events, digests the whole telemetry state, and writes a memory the driver never waits on. Give it whatever you trust for the tricky calls: cloud, local, whatever.
The fast one that replies
A small model that reads only that memory. Sub-second answers to “gap?”, “fuel?”, “P?”. Ships with a fast local default so it's free out of the box, then point it at any model you like.
How it sounds
A Piper neural voice, fully local and free by default. Swap the voice, the accent, or the whole speech engine if you'd rather.
Ask a question, get an answer in seconds. Emergencies are instant.
Your questions never wait on the slow model. Three lanes, each tuned for its job.
“Gap?” · “Fuel?” · “P?”
Answered by a fast small model, often sub-second. Ask anything and it comes straight back, no waiting on the heavy lifting.
Safety car · Red flag · Box · Blue flag
Pre-recorded and played immediately. Critical calls bypass the AI entirely, zero delay when it counts.
Full periodic telemetry calls
Designed to say things still true a lap later, so the round-trip latency simply doesn't matter.
Built like a real pit wall.
A full race-engineer stack: voice, memory, parrot Q&A and dashboard, all running on your own hardware.
Voice race engineer
Speaks radio calls out loud via a Piper neural voice. Calm, terse, factual. No coaching fluff.
Talk back to it
Voice input via VAD + Whisper speech-to-text. Ask “gap?”, “fuel?”, “should I box?” and it answers.
Instant event calls
Safety car, VSC, red flag, blue flag, “box this lap”, SC ending. All played immediately, no AI delay.
Engineer memory
The brain writes one live section per topic: gap, tyres, fuel, ERS, weather, damage, pit plan, strategy, lap times. It's what everything downstream reads.
Parrot answers
Driver questions go to a fast small model that reads only the live memory. Sub-second, every time. The heavyweight is off in the background; you never wait on it.
Live dashboard
Speed, gear, RPM, throttle/brake/steer, lap & sector times, tyre temps, and a rolling 60-second trace chart.
LAN hosting
Flip a toggle to view the dashboard from your phone or another device on your WiFi.
9 languages
English, Español, Italiano, Français, Deutsch, Português, Čeština, Polski, 日本語.
Bring-your-own model
OpenRouter, Groq, Anthropic, local Ollama, or any custom OpenAI-compatible endpoint.
Fully tunable
Chattiness slider, mute under yellow flag, suppress fuel chatter, tick interval, custom instructions, or override the whole prompt.
Your mic & speaker
Pick any input/output device. Local or cloud (Groq) transcription. First-run wizard; desktop app or headless.
Built and verified for F1 25 / 26.
Older games (F1 21–24) are selectable but unverified. Here's exactly where things stand.
F1 25
F1 24 · 23 · 22 · 2021
It speaks your language.
Put an engineer
in your ear.
The app is on its way. The repo's still private for now. Follow my GitHub profile and subscribe on YouTube to catch it the moment it drops.
Yep, this is vibecoded, and honestly I'm fine with that. If it works, it works. I'm also not careless with it, your API key never leaves your machine. If it's not your thing, no worries at all. As Max once put it: it's always if, if, if...
Standing on other people's work.
airaceGP is a small project built on a lot of open work. This page names some of the people whose code, voices, translations and track maps make it possible. The full list, including every permissive license we depend on, lives in the app repo at CREDITS.md and NOTICES.md.
Piper by Michael Hansen and the Rhasspy project, and piper-plus by ayutaz. Voice datasets by Thorsten Mueller (German), Carlos Fonseca (Spanish), czyzi0 (Polish), the Nabu Casa voice-datasets team (Portuguese, Czech), the SIWIS project (French).
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.
The F1 25 UDP telemetry specification, published by EA and Codemasters for third-party tools. Packet layouts referenced from the hotlaps.io specs (MIT).