Go Feather Route

Appearance

Providers and routing ​

The gateway supports any OpenAI-compatible endpoint, including OpenAI, DeepSeek, and local Ollama. Public model names and upstream mappings are YAML configuration data; the Go implementation does not contain a provider model catalog.

yaml
providers:
  ollama:
    base_url: "http://127.0.0.1:11434/v1"
    api_key_env: "OLLAMA_API_KEY"

model_list:
  - model_name: ollama-qwen3
    provider: ollama
    upstream_model: qwen3:4b

route_rules:
  - match: ollama/*
    provider: ollama

model_list entries are exact aliases advertised by /v1/models. route_rules can accept provider-qualified model names that are not listed, such as ollama/qwen3:8b; the gateway forwards only qwen3:8b upstream.

Ollama is expected to run outside Docker during local development. From a Docker container, use OLLAMA_API_BASE=http://host.docker.internal:11434/v1. The Ollama API key is a local placeholder; Ollama ignores it. For requests with reasoning_effort: "none", the gateway sends Ollama think: false so thinking traces do not enter Cloud’s final response.

Provider keys are referenced by environment-variable name. The gateway keeps keys server-side and forwards only the provider authorization header.

Provider kind selects provider-specific request behavior. The default openai-compatible kind forwards the shared OpenAI-compatible contract; ollama additionally maps configured aliases to local model names and maps reasoning_effort: "none" to Ollama's think: false. These behaviors belong to the provider adapter, not to a particular model.

The gateway does not retry a request after an ambiguous transport failure. A provider may have accepted the POST before the connection failed, so replaying it could duplicate work or charges. Bounded retries are limited to explicit retryable upstream responses such as rate limits and selected server errors.

Fallbacks are configured per exact model_list alias and are evaluated in order. A successful fallback response is marked in the usage event with fallback_used; a provider that reaches the failure threshold is temporarily cooled down. Cost rates are configured on the model route as USD per million input/output tokens. Provider-reported cost takes precedence when a provider adapter exposes it.