Free, Private AI Racing Coach with LM Studio

You don't need a cloud AI subscription to get telemetry coaching. With LM Studio you can run an open model on your own PC and connect it to Flowix. It's free, your conversation never leaves your computer, and only the telemetry the model asks for is fetched from Flowix.

Requirements

• A free Flowix account with laps recorded by the Flowix agent.
• A PC with a decent graphics card, or an Apple Silicon Mac. 16 GB of RAM is a practical minimum; more VRAM lets you run bigger, smarter models.
• LM Studio, a recent version with MCP support.

Step 1: Pick a model that can use tools

The model has to support tool calling, because that's how it asks Flowix for your laps. In LM Studio's model search, tool-capable models have a small hammer icon.

Good starting points:

Your hardware	Try
8–12 GB VRAM	Qwen3 8B, Qwen3 14B (Q4)
16–24 GB VRAM	gpt-oss 20B, Qwen3 30B A3B
32 GB+ VRAM or a high-memory Mac	Qwen3 32B or larger models

Bigger models coach better. Small ones work too, they just like simpler questions.

Step 2: Give the model enough memory

When you load the model, raise the context length to at least 16k tokens, or 32k if your memory allows. Context is how much of the conversation the model can keep in mind. Telemetry answers are long, and with the small default the model forgets the start of the chat.

Step 3: Add Flowix to mcp.json

1. Copy your connector URL from the Flowix AI page with the Copy URL button.
2. In LM Studio open the Program tab in the right-hand panel, then Install → Edit mcp.json.
3. Add Flowix, replacing the placeholder with your URL:

{
  "mcpServers": {
    "flowix": {
      "url": "PASTE_YOUR_FLOWIX_CONNECTOR_URL_HERE"
    }
  }
}

4. Save the file. LM Studio loads the server and shows mcp/flowix in the integrations list.

The URL is your personal key. Don't show your mcp.json in screenshots or forum posts. If it leaks, press Regenerate on the Flowix AI page and update the file.

Step 4: Enable it in a chat

Open a new chat, load your model and switch on the mcp/flowix integration. LM Studio asks before each tool call. You can allow Flowix tools permanently because they only read data.

First test: "Use the Flowix tool to list my sessions." Small models sometimes need the word tool to remember they have one.

How to talk to a local model

Local models are smaller than the big cloud assistants. The trick is a short chat, one small step at a time. Each answer gives the model what it needs for the next question. Keep each question about one lap, or two at most. Asking a local model to go through the telemetry of many laps is slow and often fails.

A first conversation

Ask these one after another, in the same chat:

List my last sessions.

What was my fastest lap in the newest one?

Where is the slowest corner on that lap?

Why is it slow?

Ask it to explain

Even a small model knows a lot about racing. These questions don't need your data at all, so they're fast:

What is a braking zone?

Why do drivers brake in a straight line?

What does a too-early downshift do to the car?

Two laps, one question at a time

Compare that lap with my second fastest.

Which one was faster in the last sector?

Progress over sessions

Session lap times are small and quick to read, so this works well even on modest hardware:

Show my best lap time for each session at this track.

Am I getting faster?

What should I work on next time?

Keeping a local model on track

• New topic, new chat. It keeps the context small and the model sharp.
• Write a short system prompt in the chat settings, for example: "You are a friendly sim racing coach. Use the Flowix tools. Keep answers short."
• Try a bigger model before blaming the data. If answers look confused, a larger model or a higher quantisation (a less compressed version of the same model) usually fixes it.

Prefer a cloud assistant? See the guides for Claude, ChatGPT and Gemini.