Connections¶

Every model call and every MCP server in this provider goes through an Airflow connection. The connection holds the credential and, for model connections, the model name in provider:model form, so a Dag names a connection id and never a key or a vendor. Create connections in the UI under Admin > Connections, with an AIRFLOW_CONN_<ID> environment variable, or through a secrets backend.

Connection type	Use it for	Page
`pydanticai`	Any vendor pydantic-ai supports through an API key and an optional base URL: OpenAI, Anthropic, Google Gemini, Groq, Mistral, DeepSeek, Ollama, vLLM, TypeSafe and others. The default for every operator and decorator.	Pydantic AI connection
`pydanticai_azure`	Azure OpenAI, with the endpoint and API version Azure needs.	Pydantic AI (Azure OpenAI) connection
`pydanticai_bedrock`	AWS Bedrock, with IAM keys, a bearer token or the default credential chain.	Pydantic AI (AWS Bedrock) connection
`pydanticai_vertex`	Google Vertex AI, with a project, a location and a service account.	Pydantic AI (Google Vertex AI) connection
`mcp`	An MCP server the agent calls as a toolset, over HTTP, SSE or a stdio subprocess.	MCP server connection
`langchain`	Chat and embedding models built through LangChain, for the LangChain hook and toolset bridge.	LangChain connection
`llamaindex`	OpenAI models for the LlamaIndex embedding and retrieval operators.	LlamaIndex connection

Which vendors work, and the extra and model prefix each one needs, is on Supported model providers. Vendor outages can fail over to a second connection with Provider fallback. Error messages from a misconfigured connection are listed with their fixes in Troubleshooting.