crewAI/docs/oauth2_llm_providers.md

OAuth2 LLM Providers

CrewAI supports OAuth2 authentication for custom LiteLLM providers through configuration files.

Features

• Automatic Token Management: Handles OAuth2 token acquisition and refresh automatically
• Multiple Provider Support: Configure multiple OAuth2 providers in a single configuration file
• Secure Credential Storage: Keep OAuth2 credentials separate from your code
• Seamless Integration: Works with existing LiteLLM provider configurations
• Error Handling: Comprehensive error handling for authentication failures
• Retry Mechanism: Automatic retry with exponential backoff for token acquisition
• Thread Safety: Concurrent access protection for token caching
• Configuration Validation: Automatic validation of URLs and scope formats

Error Handling

The OAuth2 implementation provides specific error classes for different failure scenarios:

OAuth2ConfigurationError

Raised when there are issues with the OAuth2 configuration:

• Invalid configuration file format
• Missing required fields
• Unknown OAuth2 providers

OAuth2AuthenticationError

Raised when OAuth2 authentication fails:

• Network errors during token acquisition
• Invalid credentials
• Token endpoint errors

OAuth2ValidationError

Raised when configuration validation fails:

• Invalid URL formats
• Malformed scope values

Example Error Handling

from crewai import LLM
from crewai.llms.oauth2_errors import OAuth2ConfigurationError, OAuth2AuthenticationError

try:
    llm = LLM(
        model="my_custom_provider/my-model",
        oauth2_config_path="./litellm_config.json"
    )
    result = llm.call("test message")
except OAuth2ConfigurationError as e:
    print(f"Configuration error: {e}")
except OAuth2AuthenticationError as e:
    print(f"Authentication failed: {e}")
    if e.original_error:
        print(f"Original error: {e.original_error}")

Configuration Validation

The OAuth2 configuration is automatically validated:

• URL Validation: token_url must be a valid HTTP/HTTPS URL
• Scope Validation: scope cannot contain empty values when split by spaces
• Required Fields: client_id, client_secret, token_url, and provider_name are required

Retry Mechanism

Token acquisition includes automatic retry with exponential backoff:

• 3 retry attempts by default
• Exponential backoff: 1s, 2s, 4s delays
• Detailed logging of retry attempts
• Thread-safe token caching

Configuration

Create a litellm_config.json file in your project directory:

{
  "oauth2_providers": {
    "my_custom_provider": {
      "client_id": "your_client_id",
      "client_secret": "your_client_secret", 
      "token_url": "https://your-provider.com/oauth/token",
      "scope": "llm.read llm.write"
    },
    "another_provider": {
      "client_id": "another_client_id",
      "client_secret": "another_client_secret",
      "token_url": "https://another-provider.com/token"
    }
  }
}

Usage

from crewai import LLM

# Initialize LLM with OAuth2 support
llm = LLM(
    model="my_custom_provider/my-model",
    oauth2_config_path="./litellm_config.json"  # Optional, defaults to ./litellm_config.json
)

# Use in CrewAI
from crewai import Agent, Task, Crew

agent = Agent(
    role="Data Analyst",
    goal="Analyze data trends",
    backstory="Expert in data analysis",
    llm=llm
)

task = Task(
    description="Analyze the latest sales data",
    agent=agent
)

crew = Crew(agents=[agent], tasks=[task])
result = crew.kickoff()

Environment Variables

You can also use environment variables in your configuration:

{
  "oauth2_providers": {
    "my_provider": {
      "client_id": "os.environ/MY_CLIENT_ID",
      "client_secret": "os.environ/MY_CLIENT_SECRET",
      "token_url": "https://my-provider.com/token"
    }
  }
}

Supported OAuth2 Flow

Currently supports the Client Credentials OAuth2 flow, which is suitable for server-to-server authentication.

Token Management

• Tokens are automatically cached and refreshed when they expire
• A 60-second buffer is used before token expiration to ensure reliability
• Failed token acquisition will raise a RuntimeError with details

Configuration Schema

The litellm_config.json file should follow this schema:

{
  "oauth2_providers": {
    "<provider_name>": {
      "client_id": "string (required)",
      "client_secret": "string (required)",
      "token_url": "string (required)",
      "scope": "string (optional)",
      "refresh_token": "string (optional)"
    }
  }
}

Legacy Error Handling

For backward compatibility, the following error types are still supported:

• OAuth2 authentication failures raise OAuth2AuthenticationError (previously RuntimeError)
• Invalid configuration files raise OAuth2ConfigurationError (previously ValueError)
• Network errors during token acquisition are wrapped in OAuth2AuthenticationError

Examples

Basic OAuth2 Provider

from crewai import LLM

llm = LLM(
    model="my_provider/gpt-4",
    oauth2_config_path="./config.json"
)

response = llm.call("Hello, world!")
print(response)

Multiple Providers

{
  "oauth2_providers": {
    "provider_a": {
      "client_id": "client_a",
      "client_secret": "secret_a",
      "token_url": "https://provider-a.com/token"
    },
    "provider_b": {
      "client_id": "client_b", 
      "client_secret": "secret_b",
      "token_url": "https://provider-b.com/oauth/token",
      "scope": "read write"
    }
  }
}

# Use different providers
llm_a = LLM(model="provider_a/model-1", oauth2_config_path="./config.json")
llm_b = LLM(model="provider_b/model-2", oauth2_config_path="./config.json")