smolagents

smolagents is a lightweight open-source Python library from Hugging Face that lets you build and run powerful AI agents in just a few lines of code. The core logic fits in under 1,000 lines, keeping abstractions minimal while supporting advanced agentic workflows. It features first-class support for Code Agents that write their actions as Python code rather than JSON blobs, which research shows uses 30% fewer LLM steps and achieves higher performance on difficult benchmarks.

• CodeAgent — the flagship agent type that writes actions as Python code snippets, enabling complex multi-step reasoning and tool use via function calls
• ToolCallingAgent — a standard ReAct-style agent that writes actions as JSON/text blobs for use cases where code execution is not desired
• Model-agnostic support — works with local transformers or Ollama models, Hugging Face inference providers, OpenAI, Anthropic, Azure, Amazon Bedrock, and 100+ LLMs via LiteLLM integration
• Modality-agnostic inputs — agents support text, vision, video, and audio inputs out of the box
• Tool-agnostic ecosystem — use tools from any MCP server, LangChain, Hugging Face Hub Spaces, or define your own custom tools
• Sandboxed code execution — supports E2B, Blaxel, Modal (managed cloud sandboxes), Docker (self-hosted), and Pyodide+Deno WebAssembly for secure agent code execution
• Hub integrations — share and pull tools or entire agents to/from the Hugging Face Hub as Space repositories with a single line of code
• Multi-agent hierarchies — compose multiple agents together for complex, hierarchical agentic workflows
• CLI support — run agents directly from the command line using smolagent (generalist) or webagent (web-browsing) commands
• MCP server support — load tool collections directly from any MCP server for instant extensibility

To get started, install the package with pip install "smolagents[toolkit]", then instantiate a model, create a CodeAgent with your chosen tools, and call agent.run("your task"). The library is designed to be hackable — you can use only the parts you need and extend or override any component.