> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cadenzalabs.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Configuration

> Optional extras, API keys, and environment variables.

The core CLI runs with no configuration. A few features need an extra package or
an API key.

## Optional extras

Install these into a **source install** (`pip install -e .`). The standalone
binary can't load them.

| Extra  | Install                    | Unlocks                                                                                                             |
| ------ | -------------------------- | ------------------------------------------------------------------------------------------------------------------- |
| `gym`  | `pip install -e ".[gym]"`  | MuJoCo simulator. Required for `env run`, `env show` rendering, and the SDK.                                        |
| `lora` | `pip install -e ".[lora]"` | LoRA fine-tuning of the action head. Required for `env lora`, `env run --policy lora`, and the `env vla` GRD loop.  |
| `rl`   | `pip install -e ".[rl]"`   | Residual RL + distillation (`env residual`, `env distill`). `torch` only — keeps the RL stack separate from `lora`. |
| `vla`  | `pip install -e ".[vla]"`  | Real VLA loaders for `env vla` with `--model smolvla\|hf\|local` (adds `transformers` + `lerobot`).                 |

`gym`, `lora`, and `vla` pull in `cadenza-lab` (the [SDK](/sdk/introduction)).
`lora`, `rl`, and `vla` install `torch`. Extras are independent — install only
what a given workflow needs.

## Environment variables

| Variable             | Used by     | Purpose                                                                              |
| -------------------- | ----------- | ------------------------------------------------------------------------------------ |
| `GROQ_API_KEY`       | `env train` | API key for the Groq LLM-as-Judge that rewrites a project's system prompt.           |
| `CADENZA_ACCEL`      | torch paths | Set to `0` to disable the [acceleration layer](/cli/acceleration) (pure eager fp32). |
| `CADENZA_NO_COMPILE` | torch paths | Set to `1` to keep dtype/thread tuning but skip `torch.compile`.                     |
| `CADENZA_NO_NOTIFY`  | shell       | Set to `1` to silence the terminal bell/notification after long runs.                |
| `XDG_CACHE_HOME`     | shell       | Overrides the base cache directory (default `~/.cache`).                             |

```bash theme={null}
export GROQ_API_KEY="gsk_..."     # then: cadenza env train my-project
```

## The Cadenza API

Account, action sync, and all **governed verdicts** (`lora eval`, `residual
train/eval/bench`, `distill eval`, `vla grd/eval`) go through the hosted Cadenza
API at `https://www.api.cadenzalabs.xyz`. This is hard-wired into the client — you
don't configure it, and the client never talks to a database directly. Governed
commands authenticate with your saved session token, which is why they require
[sign-in](/cli-setup/authentication). The scorecard/verdict logic lives
server-side; see [Governance](/cli/governance).

## Files & directories

| Path                      | Contents                                                            |
| ------------------------- | ------------------------------------------------------------------- |
| `~/.cadenza/config.json`  | Your saved sign-in session (token).                                 |
| `<project>/env.json`      | The mission specification for a mission project.                    |
| `<project>/vla.json`      | The model + action library for a [VLA-only project](/cli/vla-mode). |
| `<project>/.cadenza-env/` | Per-run logs and fine-tune records.                                 |
| `<project>/residual/`     | Residual config + trained policy + rollback snapshot.               |
| `<project>/student/`      | Distilled base-free student + baseline.                             |