Hugging Face agent app store

Hugging Face has turned its Spaces marketplace into a distribution layer for artificial intelligence agents, giving users a directory where agent-like apps can be found and run from one place. (huggingface.co) On April 15, 2026, the Spaces homepage described itself as “The AI App Directory” and showed more than 1,208,985 Spaces, with filters for running apps and labels including “MCP,” short for Model Context Protocol. (huggingface.co) An artificial intelligence agent is a program that does more than answer a prompt once: it can choose tools, take multiple steps, and keep going until it finishes a task. Hugging Face’s own smolagents library says agents are useful when a language model needs to determine the workflow of an app. (huggingface.co) Hugging Face started laying the plumbing for this in late 2024, when it launched smolagents on December 31, 2024 as an open-source library for building agents that “write actions in code.” Its documentation now says agents and tools can be shared to and loaded from the Hub as Gradio Spaces. (huggingface.co) The store-like piece is the combination of that agent-building stack with a discovery layer users already know how to browse. Hugging Face’s welcome page now pitches Spaces as “The community AI apps directory,” not just a place to host demos. (huggingface.co) The other change is distribution into outside assistants. Hugging Face’s “Agents on the Hub” documentation says its Model Context Protocol server works with clients including ChatGPT, Claude Desktop, OpenAI Codex, Cursor, Visual Studio Code, Gemini Command Line Interface, and Zed. (huggingface.co) That means a Space is no longer only a webpage someone visits in a browser. Hugging Face says community Gradio Spaces can expose functions as tools through Model Context Protocol, so a chat assistant can call them directly with arguments and descriptions. (huggingface.co) The result is closer to an app store model than a model repository model. Instead of downloading a base model and wiring tools by hand, a user can discover a finished workflow, run it as a Space, or connect it as a tool inside another assistant. (huggingface.co) Hugging Face is also betting that agent parts will be interchangeable. Its documentation says developers can mix models from Hugging Face, OpenAI, Anthropic, Ollama, or LiteLLM-supported providers, and can use tools from any Model Context Protocol server, LangChain, or a Hub Space. (huggingface.co) That makes the storefront less about one model winning and more about packaging, discovery, and reuse. Hugging Face already became the default shelf for open models; with Spaces, Model Context Protocol support, and agent tooling, it is trying to do the same for agents. (huggingface.co)