smolagents

smolagents is Hugging Face's minimal Python library for building agents in a few lines of code. Its headline idea is the CodeAgent: instead of emitting JSON tool calls, the agent writes its actions as Python code, which composes naturally with loops, conditionals, and function nesting; a classic ToolCallingAgent covers JSON/text tool-calling when that paradigm fits better.

It is model-agnostic (Hugging Face Inference providers, OpenAI, Anthropic, 100+ providers via LiteLLM, or local models via Transformers/Ollama), tool-agnostic (tools from any MCP server, from LangChain, or even a Hub Space as a tool), and modality-agnostic (text, vision, video, audio). Code execution can run locally in a restricted interpreter or in a remote sandbox (Modal, Blaxel, E2B, Docker), and it ships CLI utilities (smolagent, webagent) for running agents without boilerplate.

1. Install with the default toolkit: pip install 'smolagents[toolkit]' (includes default tools like web search)
2. Authenticate to Hugging Face (set HF_TOKEN / log in with huggingface-cli) if you use the hosted Inference API; for other providers, have your API keys ready for LiteLLM
3. Create your first agent per the docs: from smolagents import CodeAgent, InferenceClientModel → agent = CodeAgent(tools=[], model=InferenceClientModel()) → agent.run("your task")
4. Add tools, e.g. DuckDuckGoSearchTool(), or point at other models: LiteLLMModel for OpenAI/Anthropic (needs smolagents[litellm]), TransformersModel for local models (needs smolagents[transformers])
5. If the agent will run untrusted or production workloads, switch code execution to a sandboxed backend (Modal, Blaxel, E2B, or Docker) instead of the default local interpreter