It is a thin wrapper over DeepInfra with one tool per task, so the choice of model is configuration rather than code, and the tools you never use can be switched off before the client ever sees them.
An MCP server over DeepInfra's OpenAI-compatible API. Each capability is a separate tool with its own default model, and you decide which of them the server exposes at all.
- Images generated from a prompt
- Speech transcribed to text
- Text generated against a chat model
- Text classified as a whole, and the tokens inside it labelled
- Images classified against fixed labels, or against labels you supply at call time
- Objects located in an image
A DeepInfra API key in DEEPINFRA_API_KEY, and uv to install and run. Each tool's default model is set by its own variable in the .env file, so image generation and transcription can be pointed at different models. Another variable decides which tools appear: give it a comma-separated list to expose only the ones you want.
One command plus a key — uv sync, then supply credentials