Roder
Providers

Fireworks provider

Roder ships a first-party fireworks inference provider. It preserves Fireworks account-scoped model ids, streams turns through the OpenAI-compatible Responses transport, and appears in provider discovery even before credentials are configured so the setup path is visible in clients.

Setup. Set FIREWORKS_API_KEY orRODER_FIREWORKS_API_KEY, then select providerfireworks. Missing credentials fail locally with provider-specific guidance instead of an opaque remote authentication error.

Provider shape

  • Canonical provider id: fireworks.
  • Default endpoint: https://api.fireworks.ai/inference/v1.
  • Transport: OpenAI-compatible Responses requests.
  • Model ids keep the full Fireworks path, such as accounts/fireworks/models/qwen3-235b-a22b.
  • Model discovery is non-blocking; offline fallback metadata remains available when credentials are missing.

Configuration

The provider uses Fireworks-specific credentials and does not fall back to OpenAI environment variables. The default model is available from offline catalog data, while authenticated discovery can expose the models available to the account.

provider = "fireworks"
model = "accounts/fireworks/models/qwen3-235b-a22b"

[providers.fireworks]
api_key_env = "FIREWORKS_API_KEY"
base_url = "https://api.fireworks.ai/inference/v1"

Runtime contract

Fireworks turns stream through the same canonical Roder inference events as OpenAI, OpenRouter, Vertex, Cursor, and other providers. Roder-owned tools remain client-executed through the normal tool executor, so policy, approvals, transcripts, and cancellation work the same way as other provider routes.