ai-providers

Mule supports configuring multiple AI providers. These are used to provide the models that will be used by Agents.

Mule uses pi as its AI runtime, which supports a wide range of providers. Mule’s internal provider system supports the following core providers:

Core Supported Providers

Mule directly supports these providers through its provider configuration:

ProviderDescriptionAuthentication
OpenAIGPT-4, GPT-4o, and other OpenAI modelsAPI Key
AnthropicClaude models (Claude 3.5 Sonnet, Opus, etc.)API Key
Google GeminiGemini 1.5 and Gemini 2.0 modelsAPI Key
OllamaLocal LLM inference with OpenAI-compatible APILocal (no key)
OpenRouterUnified access to 100+ modelsAPI Key

Note: Mule creates a default OpenRouter provider on first startup. OpenRouter provides access to multiple AI models through a single API and offers free credits for new users. Sign up at openrouter.ai to get started.

OpenRouter is recommended because it:

  • Provides access to 100+ models from various providers
  • Offers free credits for new users
  • Uses a unified API key for all supported models
  • Supports models from OpenAI, Anthropic, Google, Meta, Mistral, and more

Configuration via API

Create a Provider 

curl -X POST http://localhost:8140/api/v1/providers \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-provider",
    "api_base_url": "https://api.openai.com/v1",
    "api_key_encrypted": "sk-your-api-key"
  }'

Provider Configuration Fields

FieldDescription
nameDisplay name for the provider
api_base_urlBase URL for the API endpoint
api_key_encryptedYour API key (stored securely)

Provider-Specific Configuration

OpenAI / OpenAI-Compatible 

# Standard OpenAI
api_base_url: "https://api.openai.com/v1"
api_key_encrypted: "sk-your-openai-key"

# OpenRouter
api_base_url: "https://openrouter.ai/api/v1"
api_key_encrypted: "sk-or-v1-your-openrouter-key"

# Azure OpenAI
api_base_url: "https://your-resource.openai.azure.com"
api_key_encrypted: "your-azure-key"

Anthropic 

api_base_url: "https://api.anthropic.com"
api_key_encrypted: "sk-ant-your-anthropic-key"

Google Gemini 

api_base_url: "https://generativelanguage.googleapis.com"
api_key_encrypted: "your-gemini-key"

Ollama (Local) 

api_base_url: "http://localhost:11434/v1"
api_key_encrypted: "ollama"  # or any placeholder

Local Models

For local models, configure Ollama or any OpenAI-compatible server:

Ollama Setup

  1. Install Ollama from ollama.ai
  2. Pull a model: ollama pull llama3.2
  3. Configure Mule:
curl -X POST http://localhost:8140/api/v1/providers \
  -H "Content-Type: application/json" \
  -d '{
    "name": "ollama-local",
    "api_base_url": "http://localhost:11434/v1",
    "api_key_encrypted": "ollama"
  }'

LM Studio / vLLM

Any OpenAI-compatible inference server works:

{
  "providers": {
    "local": {
      "baseUrl": "http://localhost:1234/v1",  # LM Studio default
      "api": "openai-completions",
      "apiKey": "local",
      "models": [
        { "id": "model-name" }
      ]
    }
  }
}

Note: LM Studio typically runs on port 1234, while vLLM uses port 8000 by default.


## Subscription Providers (OAuth via pi)

Mule uses [pi](https://pi.dev) as its agent runtime, which supports OAuth-based authentication for subscription providers:

- **Claude Pro/Max** - Use `/login` in pi interactive mode
- **ChatGPT Plus/Pro (Codex)** - Use `/login` in pi interactive mode
- **GitHub Copilot** - Use `/login` in pi interactive mode
- **Google Gemini CLI** - Use `/login` in pi interactive mode

For these providers, use the pi CLI's OAuth login flow rather than API keys.

## Cloud Provider Configuration

### Amazon Bedrock

```bash
export AWS_PROFILE=your-profile
# Or IAM keys
export AWS_ACCESS_KEY_ID=AKIA...
export AWS_SECRET_ACCESS_KEY=...
# Optional region
export AWS_REGION=us-west-2

Google Vertex AI 

gcloud auth application-default login
export GOOGLE_CLOUD_PROJECT=your-project
export GOOGLE_CLOUD_LOCATION=us-central1

Custom Providers

For custom providers with specific API requirements, you can create a pi extension to register custom provider implementations. See the pi custom provider documentation for details.

All providers support model switching, tool assignment, and dynamic configuration.

Not all fields are required—only the information necessary to interact with your chosen provider must be provided.

Was this page helpful?