# Crush

> Connect the Crush terminal AI coding tool to the LMU AI API — add a custom provider in crushrc to use Claude, GPT and Chinese models.

URL: https://docs.lmuai.com/en/docs/tools/crush



Crush is a terminal AI coding tool from [Charm](https://charm.land), offering both a CLI and a TUI for code generation, debugging, conversation, file operations, and multitasking from the command line. It supports custom providers — this page covers connecting it to LMU AI.

***

## Install Crush [#install-crush]

Pick whichever method suits your system:

<Tabs items="['Homebrew (recommended on macOS)', 'NPM (cross-platform)', 'Windows', 'Arch Linux', 'Nix']">
  <Tab value="Homebrew (recommended on macOS)">
    ```bash
    brew install charmbracelet/tap/crush
    ```
  </Tab>

  <Tab value="NPM (cross-platform)">
    ```bash
    npm install -g @charmland/crush
    ```
  </Tab>

  <Tab value="Windows">
    ```powershell
    winget install charmbracelet.crush
    ```

    Or with Scoop:

    ```powershell
    scoop bucket add charm https://github.com/charmbracelet/scoop-bucket.git
    scoop install crush
    ```
  </Tab>

  <Tab value="Arch Linux">
    ```bash
    yay -S crush-bin
    ```
  </Tab>

  <Tab value="Nix">
    ```bash
    nix run github:numtide/nix-ai-tools#crush
    ```
  </Tab>
</Tabs>

***

## Configure LMU AI [#configure-lmu-ai]

Crush is configured with a **Bash script**. Edit the global config file:

* **macOS / Linux**: `~/.config/crush/crushrc`
* **Windows**: `%USERPROFILE%\.config\crush\crushrc`

There are two steps: add LMU AI as a provider with `provider add`, then register the models you want with `model add`. Choose the protocol that matches the models you plan to use:

<Tabs items="['GPT / general (openai-compat)', 'Claude / Chinese models (anthropic)']">
  <Tab value="GPT / general (openai-compat)">
    Works for OpenAI models such as `gpt-5.6-sol` and `gpt-5.5`. Claude and Chinese models can also go through this protocol (the LMU AI backend handles the protocol conversion).

    ```bash
    # Add LMU AI as a provider (OpenAI-compatible protocol)
    provider add lmuai \
      --type openai-compat \
      --base-url "https://api.lmuai.com/v1" \
      --api-key "sk-your-lmu-ai-api-key"

    # Register the models you want
    model add lmuai/gpt-5.6-sol --name "GPT-5.6 Sol" --context-window 400000
    model add lmuai/claude-opus-5 --name "Claude Opus 5" --context-window 200000

    # Set it as the default large-model slot
    model large lmuai/claude-opus-5
    ```

    <Callout type="warn" title="The type is openai-compat, not openai">
      Crush has two OpenAI types. Per the official docs, `openai` is for proxying or routing requests **through OpenAI itself**, while `openai-compat` is for **non-OpenAI providers that offer OpenAI-compatible APIs**. LMU AI is the latter, so use `openai-compat`.
    </Callout>
  </Tab>

  <Tab value="Claude / Chinese models (anthropic)">
    Works for models served over the Anthropic protocol, such as `claude-opus-5`, `claude-sonnet-5`, `qwen3.8-max-preview`, `deepseek-v4-pro`, `glm-5.2`, and `kimi-k3`.

    ```bash
    # Add LMU AI as a provider (Anthropic protocol)
    provider add lmuai-anthropic \
      --type anthropic \
      --base-url "https://api.lmuai.com" \
      --api-key "sk-your-lmu-ai-api-key" \
      --extra-header anthropic-version 2023-06-01

    # Register the models you want
    model add lmuai-anthropic/claude-opus-5 \
      --name "Claude Opus 5" \
      --context-window 200000 \
      --can-reason true \
      --supports-images true

    model add lmuai-anthropic/glm-5.2 --name "GLM-5.2" --context-window 200000

    # Set it as the default large-model slot
    model large lmuai-anthropic/claude-opus-5
    ```

    <Callout type="warn" title="Don't add /v1 to the Anthropic base URL">
      Set `--base-url` to `https://api.lmuai.com` — **not** `https://api.lmuai.com/v1`. Crush appends `/v1/messages` itself, so the extra segment makes it request `/v1/v1/messages`, which fails with:

      ```
      not found: POST "https://api.lmuai.com/v1/v1/messages": 404 Not Found
      ```

      This is the opposite of `openai-compat` (which does need `/v1`) — don't mix them up.
    </Callout>
  </Tab>
</Tabs>

<Callout type="warn" title="Config files are trusted code">
  An official warning: `crushrc` runs with your shell privileges before the UI appears, and any `$(...)` in `crush.json` runs at load time. **Don't launch Crush in a directory whose config you haven't reviewed**, and don't `source` config files of unknown origin.
</Callout>

***

## Start using it [#start-using-it]

Restart Crush once configured:

```bash
crush
```

Switch models from within a session with:

```
/models
```

<Callout type="info" title="LMU AI models missing from /models?">
  Crush's built-in model list comes from the [Catwalk](https://github.com/charmbracelet/catwalk) database, which doesn't include LMU AI model IDs. You must register them with `model add <provider>/<model-id>` first — they'll appear in the `/models` list after a restart.
</Callout>

***

## Common model IDs [#common-model-ids]

| Model ID              | Notes                                 |
| --------------------- | ------------------------------------- |
| `claude-opus-5`       | Claude Opus 5 (flagship, recommended) |
| `claude-sonnet-5`     | Claude Sonnet 5 (balanced)            |
| `claude-haiku-4-5`    | Claude Haiku 4.5 (fast)               |
| `gpt-5.6-sol`         | GPT-5.6 Sol                           |
| `glm-5.2`             | GLM-5.2                               |
| `qwen3.8-max-preview` | Qwen 3.8 Max Preview                  |
| `deepseek-v4-pro`     | DeepSeek V4 Pro                       |
| `kimi-k3`             | Kimi K3                               |

Not sure which model to enter? Go to the [Model Gallery](/en/docs/guide/models) to see every available model ID and copy it with one click.

***

## Legacy JSON config [#legacy-json-config]

<Callout type="warn" title="crush.json is officially deprecated">
  Crush originally used `crush.json`, which is now **deprecated** — the maintainers have said they'll keep supporting it, but new configuration options will only be added to the Bash config (`crushrc`). Plenty of tutorials still teach the JSON format, but **new setups should use `crushrc` as shown above**.

  Already have a JSON config and want to migrate? Start Crush and just ask it in plain language to convert it for you.
</Callout>

If you do want to use JSON, the format looks like this (in a `crush.json` in your project directory or under `~/.config/crush/`):

```json
{
  "$schema": "https://charm.land/crush.json",
  "providers": {
    "lmuai": {
      "id": "lmuai",
      "name": "LMU AI",
      "type": "openai-compat",
      "base_url": "https://api.lmuai.com/v1",
      "api_key": "sk-your-lmu-ai-api-key",
      "models": [
        {
          "id": "claude-opus-5",
          "name": "Claude Opus 5",
          "context_window": 200000,
          "default_max_tokens": 32000
        }
      ]
    }
  }
}
```

<Callout type="warn" title="Don't omit type and models in JSON">
  Just `id` / `name` / `base_url` / `api_key` won't work: `type` decides which protocol is used (omitting it falls back to a default guess), and `models` decides what you can pick under `/models`. This is the most common reason a copy-pasted minimal example fails.
</Callout>

***

## Notes [#notes]

* The config file is `~/.config/crush/crushrc` (`%USERPROFILE%\.config\crush\crushrc` on Windows); `~/.local/share/crush/crush.json` is machine-owned state — **don't edit it**
* Use `--type openai-compat` for LMU AI's OpenAI-compatible endpoint (**not** `openai`), and `--type anthropic` for the Anthropic protocol
* **The two types take opposite `--base-url` forms**: `anthropic` takes `https://api.lmuai.com` (no `/v1`), while `openai-compat` takes `https://api.lmuai.com/v1` (with `/v1`). Adding `/v1` on the Anthropic side produces `/v1/v1/messages` and a 404
* For the API key, just enter the `sk-` key generated in the LMU AI console
* LMU AI models aren't in Crush's built-in list, so they must be registered with `model add` before they appear under `/models`