huggingface/smolagents

smolagents is a small Python library from Hugging Face for building "agents", programs that use a large language model to plan and act step by step, calling tools and reading their results. Its central idea, stated in the project's tagline, is "agents that think in code": instead of producing structured tool-call commands, the agent writes short snippets of code that perform the actions, and that code is what gets executed. Because letting a language model run arbitrary code is risky, smolagents supports running the generated code inside sandboxed environments such as Blaxel, E2B, Modal, Docker, or a Pyodide and Deno WebAssembly sandbox. The library keeps its abstractions deliberately thin, the core agent logic is described as fitting in roughly a thousand lines of code, which makes it easier to read and adapt. It is model-agnostic: you can plug in a local model through transformers or ollama, use providers from the Hugging Face Hub, or reach OpenAI, Anthropic and many others through a LiteLLM integration. It is also modality-agnostic, supporting text, vision, video and audio inputs, and tool-agnostic, so tools can come from an MCP server, from LangChain, or from a Hugging Face Hub Space. You would reach for smolagents when you want to prototype or run an LLM-driven assistant that has to use external tools, searching the web, calling APIs, browsing pages, without committing to a heavy framework. The package ships with a default toolkit, a quick example built around a web search tool, and two command-line entry points, "smolagent" for a general code-writing agent and "webagent" for browsing the web.