Use Pydantic AI with open-source models hosted in the EU

Pydantic AI is an open-source Python framework for building agents and model-backed features. You declare tools as typed Python functions and describe the output you expect as a Pydantic model; the framework handles the conversation with the model, tool calls, retries on validation errors and streaming.

It supports many providers. For services without a dedicated integration, the documented route is OpenAICompatibleProvider combined with OpenAIChatModel, which speaks the Chat Completions format. That is how you connect Pydantic AI to inference.eu while keeping the same agent code.

OpenAICompatibleProvider takes a name, base_url (the inference.eu base URL ending in /v1) and api_key. Keep the name stable, for example inference-eu, because Pydantic AI records it on every response and uses it to decide which data from earlier turns may be sent back.

OpenAIChatModel takes the inference.eu model id and the provider. Use it rather than OpenAIResponsesModel, since Chat Completions is the most widely supported format. Avoid OpenAIProvider with a custom URL: it assumes OpenAI's own models and features. If a feature misbehaves, adjust the OpenAIModelProfile passed as profile.

Pydantic AI relies on tool calls for structured output and for your own tools, so models with dependable tool calling work best. glm-5.3 and kimi-k3 are strong defaults for agents with several tools or complex output types, and they follow nested schemas well.

For extraction, classification and other quick typed calls, gpt-oss-120b keeps latency low. For agents that generate or review code, use qwen3-coder-480b-a35b-instruct. You can create several models on the same provider and assign them to different agents, for example a fast triage agent that hands off to a stronger specialist.

A 401 error means the key is missing or wrong. A 404 usually means base_url has an extra path or the model id does not exist at inference.eu.

If the agent raises an error after retrying output validation, simplify the output model or raise output_retries, and choose a model with strong tool calling. If tool definitions are rejected, set openai_supports_strict_tool_definition=False in the profile. If a request fails because of multiple system messages, set openai_chat_supports_multiple_system_messages=False. Both are fields of OpenAIModelProfile.

With the provider pointed at inference.eu, every prompt, tool result and output your agents exchange with the model is processed in EU data centres and never used for training. We sign a data processing agreement under Article 28 GDPR, publish our subprocessor list and bill in euros.

Pydantic AI itself only calls the provider you configure. It integrates with OpenTelemetry, and Logfire can record prompts and outputs when you enable instrumentation; check where your tracing backend stores that data. Tools your agents call, such as search APIs, also receive data under their own terms.