SmolAgents

Hugging Face's minimalist agent library where agents act by writing Python code

Agents & AutomationFor developersOpen sourceActively maintained

About

SmolAgents is Hugging Face's lightweight Python library for building agents in a few lines of code. Its CodeAgent writes its actions as Python code instead of JSON tool calls, which makes multi-step reasoning and tool composition compact and capable. It is model-agnostic — Hugging Face Inference, LiteLLM, OpenAI-compatible APIs, or local models via Transformers/Ollama — and is best for developers who want a small, hackable agent framework rather than a large platform.

pip install "smolagents[toolkit]"

Requires Python and, depending on your setup, an API key (e.g. a Hugging Face token or OPENAI_API_KEY) or a local model runtime. For untrusted code execution, use one of the supported sandboxes (E2B, Modal, Docker) — the built-in local executor is not a security boundary.