Providers overview
crab’d is model-agnostic. A model is chosen with a <provider>/<model> specifier; the provider
decides how the request is authenticated and routed.
model: anthropic/claude-sonnet-4-6
modes: review: model: openai/gpt-5.5 # per-mode overrideProvider guides
Section titled “Provider guides”| Provider | ID(s) | Guide |
|---|---|---|
| Anthropic (Claude) | anthropic |
Anthropic |
| OpenAI | openai |
OpenAI |
| Google (Gemini, Vertex AI) | google, google-vertex |
|
| OpenAI-compatible / local | your id, ollama |
OpenAI-compatible & local |
OpenRouter (openrouter, key OPENROUTER_API_KEY) and other Pi-catalog providers work the same way.
Pick a <provider>/<model> specifier and supply the provider’s credential.
The provider allowlist
Section titled “The provider allowlist”By default the allowlist is empty, which allows any provider — crab’d works with zero config. Set it to restrict which providers your repo code may reach; then any model whose provider isn’t listed fails the run before a byte leaves your CI:
providers: allowlist: [anthropic, openai]Routing & self-hosting
Section titled “Routing & self-hosting”- Route built-in providers through a proxy with the egress gateway.
- Point at any OpenAI-compatible URL with a custom provider.
