Use Crush with llmrelay

Setup

1. 1 Create or open crush.json in your project root, or $HOME/.config/crush/crush.json for a global config.
2. 2 Add a providers entry with type: 'anthropic' (use 'openai-compat' for the OpenAI-compatible surface).
3. 3 Set base_url to the relay endpoint and api_key to $LLMRELAY_API_KEY so the value is expanded from the environment.
4. 4 List the models you want with id, name, context_window and default_max_tokens.
5. 5 Export LLMRELAY_API_KEY in your shell, then run `crush` and pick the model.

Config

// crush.json  (project root, or $HOME/.config/crush/crush.json)
{
  "$schema": "https://charm.land/crush.json",
  "providers": {
    "llmrelay": {
      "type": "anthropic",
      "base_url": "https://api.llmrelay.dev",
      "api_key": "${LLMRELAY_API_KEY:?set LLMRELAY_API_KEY}",
      "extra_headers": {
        "anthropic-version": "2023-06-01"
      },
      "models": [
        {
          "id": "claude-opus-5",
          "name": "Claude Opus 4.8 (relay)",
          "context_window": 200000,
          "default_max_tokens": 50000,
          "can_reason": true,
          "supports_attachments": true
        }
      ]
    }
  }
}

# export LLMRELAY_API_KEY=***

Smoke test

Copy, paste, run. If you get JSON back with a message, you're good.

Run `crush` in the project, confirm the llmrelay provider and model appear in the model list, and send one prompt.

Gotchas

• !Ignore the /v1 in the README's Anthropic example (https://api.anthropic.com/v1). We tested both against the relay on 2026-07-28: https://api.llmrelay.dev returns 200, and adding /v1 yields /v1/v1/messages which 404s. Use the bare host.
• !Use type 'openai-compat' (not 'openai') for non-OpenAI providers with OpenAI-compatible APIs; 'openai' is only for proxying to OpenAI itself.
• !crush.json is trusted code: any $(...) in it runs at load time with your shell privileges before the UI appears. Use ${VAR:?message} for required keys so an unset variable fails loudly instead of resolving to empty.

Related integrations