Configure Model Providers

WeaveFlow uses an OpenAI-compatible client for model-driven nodes. It supports OpenAI, Azure, DeepSeek, Gemini, vLLM, Mistral, xAI, and OpenRouter, with either the Chat Completions or Responses request format.

Environment variables for examples

The simplest local setup is:

export OPENAI_API_KEY="your-key"
export OPENAI_BASE_URL="https://api.openai.com/v1"
export OPENAI_MODEL="your-model"

OPENAI_BASE_URL is optional for the default OpenAI endpoint. Keep credentials in a local .env file or a secret manager. Never commit them, put them in a Graph Definition, or print them in logs.

Configure a provider explicitly

When a provider needs request-specific fields, configure the client explicitly in Go:

model, err := openai.New(
    openai.WithToken(os.Getenv("OPENAI_API_KEY")),
    openai.WithModel(os.Getenv("OPENAI_MODEL")),
    openai.WithBaseURL(os.Getenv("OPENAI_BASE_URL")),
    openai.WithProvider(openai.ProviderDeepSeek),
    openai.WithAPIFormat(openai.APIFormatChatCompletions),
)
if err != nil {
    log.Fatal(err)
}

The default client uses the Chat Completions format. Select openai.APIFormatResponses when the endpoint implements the Responses API. The provider and format must match the endpoint you are calling.

Server-managed assistant

cmd/server can expose the optional Assistant API when all required variables are present:

export WEAVEFLOW_ASSISTANT_API_KEY="your-key"
export WEAVEFLOW_ASSISTANT_MODEL="your-model"
export WEAVEFLOW_ASSISTANT_BASE_URL="https://api.openai.com/v1"
export WEAVEFLOW_ASSISTANT_PROVIDER="openai"
export WEAVEFLOW_ASSISTANT_API_FORMAT="responses"
go run ./cmd/server -data .local/server

Assistant configuration is independent of the model IDs referenced by a Graph Definition. A graph node can select a model from the runtime model context with its model_id; use the Workbench settings or your own server integration to provide multiple models.

Compatibility checklist

1. Verify the base URL points to the API root, commonly ending in /v1.
2. Verify the model ID is accepted by the provider.
3. Choose the request format (chat_completions or responses) supported by that endpoint.
4. Run a small text_generation or llm_turn graph before enabling tools or long plans.
5. Inspect the Run Events and provider error details if the request fails.

For a credential-free first run, use the model-free examples.