Messages API: Anthropic-compatible requests, processed in the EU

POST /v1/messages takes a list of messages that alternate between user and assistant, plus an optional top-level system prompt, and returns the next assistant turn. Unlike the OpenAI format, the system prompt is not a message but its own field, and max_tokens is required on every request.

Authentication uses the x-api-key header with your inference.eu key, and the anthropic-version header is accepted as Anthropic clients send it. Because the endpoint follows the Anthropic shape, the official Python and TypeScript SDKs work once you set their base URL to the API host. Note that those SDKs add /v1/messages themselves, so give them the host without the /v1 suffix.

Message content can be a string or a list of content blocks: text, image (base64 or URL) for vision models, tool_use and tool_result. The response is a message object with id, role: "assistant", the model, and a content array of blocks. Plain answers contain one text block; reasoning models can add a thinking block first.

stop_reason is end_turn when the model finished, max_tokens when it ran out of room, stop_sequence when it hit one of your stop_sequences, and tool_use when it wants you to run a tool. usage reports input_tokens and output_tokens.

Define tools in a top-level tools array, each with name, description and an input_schema written as JSON Schema. When the model wants to use one, the response ends with stop_reason: "tool_use" and contains a tool_use block with an id and input. Run the tool, then send a user message containing a tool_result block that references the same tool_use_id.

This is the loop coding agents run all day: read a file, edit, run tests, repeat. Pointing such an agent at inference.eu keeps source code inside European data centres. Pick a model that handles long tool loops well, and set max_tokens generously so edits are not cut off mid-file.

Errors follow the Anthropic shape: {"type": "error", "error": {"type": "...", "message": "..."}} with an HTTP status. Expect invalid_request_error (400) for a missing max_tokens or messages that do not alternate correctly, authentication_error (401) for a bad key, not_found_error (404) for an unknown model, and rate_limit_error (429) when you reach a limit for your plan.

Anthropic model names such as Claude are not available here; send an open model id instead. Most clients let you override the model with an environment variable or setting. Features tied to Anthropic's own models may be accepted but have no effect.