# Remixing Setups

> Connect an AI client to your Race Engineer and evolve your setups conversationally — every change explained, every version saved.

Remixing happens **in conversation, not in the app**. You connect an MCP-compatible AI client to your Race Engineer, describe what the car is doing, and ask what you should test. The engineer recommends first; it creates a new version only after you explicitly ask it to apply the change.

> Setups is in **Public Beta** and requires a personal **Plus or Ultra** plan. A Team seat alone does not grant access.

## Connect Your AI Client

Open [Connect agent](https://app.brakinglab.com/ai-setups/connect) in the app. It gives you the MCP server URL and a ready-to-paste config for **Claude Desktop**, and the same URL works with **ChatGPT**, **Cursor**, **Windsurf**, or any MCP-compatible client.

Authentication is **account-based — there is no API key**. The first time your client connects, it opens the browser and you sign in with your Braking Lab account. Full walkthrough: [MCP Setup](/en/docs/race-engineer/mcp-setup).

Your Race Engineer runs inside that client. What it reads from your vault — setup
values, notes, evidence — passes through the client's provider. See
[What stays private](/en/docs/ai-setups/introduction#what-stays-private).

## Talk to Your Engineer

Describe symptoms and goals, not parameter names:

> "Take my Spa GT3 setup and make it more stable on entry — soften the rear and add a touch of rear wing."

> "I'm killing my front tires after 10 laps at Barcelona. Find me some life without giving up all the turn-in."

The engineer reads the setup from your vault, decides which parameters are worth testing, and explains why. If you later ask it to apply the recommendation, it edits **only adjustable values** — settings your garage actually exposes — and never touches anything else.

It can use your [type and personal tags](/en/docs/ai-setups/types-and-tags) to
find the setup you mean and repeat your own description back to you. Those
labels are user-declared context, not a diagnosis or evidence for deciding what
to change.

## How a Setup Decision Uses Context

You can ask about a symptom or a specific setting:

> "Would one click toward less negative rear camber make sense on this setup?"

For that decision, the engineer brings together the context that actually exists:

- Current setup values and which settings are adjustable
- Compatible session telemetry and the feedback you gave it
- Your confirmed experiment history and saved comparisons for this setup
- Reviewed setup tendencies and car-specific notes when coverage exists

Personal evidence takes priority over compatible telemetry and driver feedback,
then reviewed knowledge, then clearly labeled general knowledge. Missing context
does not get silently filled in. The answer is graded as **supported**,
**plausible**, **not supported**, or **insufficient evidence**.

The response follows a deliberately narrow shape:

1. It answers the setting you asked about first.
2. It may offer **one primary alternative**, but only after verifying that the setting exists and is adjustable on your car.
3. Each hypothesis changes at most **one or two settings**, so the next stint can teach you something.

## Clicks, Values and Physical Direction

The engineer can speak in **clicks**, but always with the physical direction:

- **"1 click toward less negative rear camber"**, never "index +1"
- For a paired setting: **"1 click on each rear side"**, not "2 clicks"
- **Current → proposed** values only when the garage mapping is reliable; it never invents a unit or display value

Some questions need signals that the current telemetry does not provide. Camber,
for example, can be proposed as a hypothesis, but it is never called **optimal**
without inner/middle/outer tire temperatures. The engineer states that limitation
instead of turning a plausible direction into a certainty.

## Recommendation Is Not Mutation

Asking "does this make sense?" is read-only. The engineer does not create a
version, recipe, evaluation or association merely because it recommended a test.
If you want the change, ask explicitly in a later message — for example:

> "Apply that one-click rear anti-roll-bar test as a new version."

## What a Remix Produces

| Sim | Result |
|-----|--------|
| **Le Mans Ultimate** | A new version with a downloadable `.svm` — load it directly in the game |
| **iRacing** | A new **recipe version** — the exact changes to apply in the garage ([how](/en/docs/ai-setups/iracing-setups)) |

The rules every remix follows:

- **At most 8 changes.** That is a hard limit per version. One or two changes per
  hypothesis is the recommendation, so you can tell what helped.
- **Only adjustable parameters.** A change to an unknown or non-adjustable
  parameter is rejected, and no version is written.
- **Observed values advise; they do not block.** Braking Lab checks each value
  against what it has seen in uploaded files for that car. A value outside that
  observed range is still saved, with an advisory for you to confirm in the
  garage.
- **The LMU protected guard.** For LMU car variants with reviewed coverage, every
  change must use a certified option stop. If one does not, Braking Lab rejects
  the whole request before it writes a version, a conversation milestone or a
  file, and the parent stays current. Without reviewed coverage, the structural
  checks still run, but the result is never called certified.
- **A concurrent version wins.** When the engineer builds on the latest version
  and another version lands first, the remix fails with `SETUP_BASE_MOVED` and
  saves nothing. The engineer re-reads the setup and tries again.

## What Each Version Records

Every version records its **lineage**: which version it came from and what
changed. Braking Lab also saves two milestones of the conversation with it: your
request (up to 500 characters) and the engineer's rationale (up to 1,200
characters). They appear in the setup's **Conversation** panel. The rest of the
conversation stays in your AI client and is not stored by Braking Lab.

If a session's telemetry motivated the remix, the version shows **Motivated by**
with a link to that session. When the remix follows a saved comparison, see
[Measured iterations](/en/docs/ai-setups/measuring-changes#measured-iterations).

## Setup Interview

For **Le Mans Ultimate**, the Setup Interview can derive a child version from an
existing owned `.svm` for a different exact session, condition, risk or balance
target. Protected references teach bounded relative movements; no reference
file or absolute value is copied into your vault.

Full walkthrough: [Setup Interview](/en/docs/ai-setups/setup-interview).