Crush

Crush, from Charm, reads provider setup from a crushrc file: a short shell script that runs Crush's own commands. The openai-compat type speaks Chat Completions, so the base URL is https://deference.si/v1.

Set up

1. Create a key in API keys and export it as DEFERENCE_API_KEY.
2. Create crushrc in ~/.config/crush/ (on Windows, %USERPROFILE%\.config\crush\crushrc). A .crushrc file in a project folder works too, and is safe to commit because it holds no key.

provider add deference \
  --name Deference \
  --type openai-compat \
  --base-url "https://deference.si/v1" \
  --api-key "${DEFERENCE_API_KEY:?set DEFERENCE_API_KEY}"

model add deference/anthropic/claude-sonnet-5.5 \
  --name "Claude Sonnet 5.5" \
  --context-window 200000 \
  --default-max-tokens 32000

model large deference/anthropic/claude-sonnet-5.5
model small deference/anthropic/claude-sonnet-5.5

3. Start Crush.

The key is read from your environment when the file runs, so it is never written down. model large is the model Crush uses for the main agent, and model small is the lighter one it uses for light tasks. Point small at a smaller model if you like.

Declare the models you want with model add so their context and output limits are explicit. Crush can also discover models for an OpenAI-compatible provider with no declared models. Add --discover-models true to merge discovered models with your declarations; your explicit settings take priority. See Crush's configuration guide.

Verify

1. Run crush models. It lists deference/anthropic/claude-sonnet-5.5.
2. Run a one-off prompt.

crush run -m deference/anthropic/claude-sonnet-5.5 \
  "Reply with exactly: gateway-ok"

3. Open Activity. The request is the top row, with the model you chose.

If something fails, crush logs --tail 50 shows the provider errors.

Pick a model

• Crush splits provider/model on the first slash. In deference/anthropic/claude-sonnet-5.5 the provider is deference and the model id is anthropic/claude-sonnet-5.5. Without the deference/ prefix, Crush reads anthropic as the provider and fails with "model not found".
• Change the model in a session with Ctrl L.
• Use --type openai-compat for Deference. The openai type is for OpenAI itself.

Troubleshooting