Introduction

maKI is an OpenAI-compatible API service running entirely on Uni Mannheim infrastructure. No requests or data leave the university.

Overview

How it works

1. You send a request to https://maki.uni-mannheim.de/v1/chat/completions — exactly like the OpenAI API.
2. maKI validates your API key and routes the request to the chosen model.
3. The model runs locally on GPUs at Uni Mannheim and generates the response.
4. The response is streamed back — token usage is logged per key.

Prompts and responses are not stored (see Privacy Policy).

Available models

For a current list with context windows and usage recommendations, see Available Models.

API keys

Every request to maKI must carry an API key — a randomly generated string that identifies your application. Keys are issued by the administrator and can be revoked at any time. For key types, requesting, and handling, see API Keys.

Features

• OpenAI-compatible: Any SDK or tool that supports the OpenAI API works without modification.
• Fully local: All models run on dedicated GPUs at Uni Mannheim. No third-party providers involved.
• Streaming: Responses are streamed token by token (stream: true).
• Structured output: Models can produce JSON-schema-compliant output (see Structured Output).
• Embeddings: In addition to chat models, embedding models are available for vector search and RAG.