HuggingChat tutorial: OmniRouter and MCP tools
Learn model selection and MCP setup in HuggingChat, including dynamic Spaces for image generation and documentation search for 4-bit quantization.
Player not loading? Watch on YouTube
This tutorial introduces HuggingChat as Hugging Face's open source chat interface for models served through inference providers. The walkthrough uses hosted inference; it does not demonstrate how to run models locally.
The speaker starts with OmniRouter, which automatically chooses a model for a prompt. A Python sorting function provides the example. Users can also select a model manually, with icons indicating tool calling and image understanding. The speaker recommends choosing a model with tool calling support when using MCP servers.
The Hugging Face MCP server supplies tools for searching the Hub and documentation. Its experimental dynamic Spaces feature finds a compatible Space and uses it as an MCP tool. The demonstration generates an image of a Bavarian town with the Eiffel Tower without manually adding a separate image generation tool.
Other examples search for text-to-image models and ask how to load a model with 4-bit quantization. A documentation chat bubble opens HuggingChat with the current page as context. Custom MCP setup requires a name, endpoint URL and authentication when needed. The speaker also explains enabling Exa for web search. The account-credit and pay-as-you-go remarks describe the service at recording time; this tutorial does not establish permanent free usage.
Design Blog
