model-providers

OpenAI Agents: Custom Model Providers

Shows how to pass a custom (non-default) ModelProvider to OpenAIAgentsPlugin. Any OpenAI Agents SDK ModelProvider can drive the model Activity, so you can point an agent at any OpenAI-compatible endpoint (a local server such as Ollama, OpenRouter, or another gateway) by constructing OpenAIProvider with a custom baseURL.

The provider runs inside the model Activity, never in the Workflow sandbox.

Run

Start a Temporal dev server:

temporal server start-dev

Point the Worker at your OpenAI-compatible endpoint. For example, a local Ollama server:

export OPENAI_BASE_URL=http://localhost:11434/v1
export OPENAI_API_KEY=ollama        # any non-empty value; most local servers ignore it
export OPENAI_MODEL=llama3.2        # a model your endpoint serves

Pull the model and start Ollama:

ollama pull llama3.2
ollama serve

Or an OpenRouter endpoint:

export OPENAI_BASE_URL=https://openrouter.ai/api/v1
export OPENAI_API_KEY=sk-or-...
export OPENAI_MODEL=meta-llama/llama-3.1-8b-instruct

Then start the Worker and run the Workflow (run from the openai-agents/ root, after npm install there):

npx ts-node src/model-providers/worker.ts

npx ts-node src/model-providers/client.ts "Say hello in one sentence."

OPENAI_MODEL is forwarded to the run via runConfig.model and resolved by your custom provider.

Test

The tests run fully offline. They inject a FakeModelProvider — itself a custom ModelProvider — and assert the agent run is handled by the injected provider, including resolving the requested model name. This is the same injection point the live custom-provider setup uses.

npx mocha --exit --require ts-node/register --require source-map-support/register "src/model-providers/mocha/*.test.ts"