smolagents
Minimal library for building secure code-based AI agents.
View on GitHub Updated 2026-06-15
What is smolagents?
smolagents is an open-source Python framework designed to run lightweight yet powerful AI agents. It centers on CodeAgent, which generates and executes Python code to solve tasks instead of relying on text-only action formats.
The library connects to any LLM through local runtimes, Hugging Face inference, or third-party providers via LiteLLM. It accepts tools from MCP servers, LangChain, or Hub Spaces and supports text, image, video, and audio inputs.
It targets Python developers who want minimal abstractions, quick prototyping, and the ability to share finished agents directly on the Hugging Face Hub.
Capabilities
run code agents in minimal code
execute actions in sandboxed environments
integrate any LLM provider
support text/vision/video/audio inputs
use tools from MCP/LangChain/Hub
What you can build with smolagents
Web research queries
Combine a CodeAgent with WebSearchTool to answer factual questions that require live data lookup and calculation.
Multimodal analysis
Process images or video by feeding them directly to supported models and letting the agent write code to interpret results.
Agent sharing
Push a finished agent to the Hub as a Space so teammates can load and reuse it instantly with a single command.
Install smolagents
Install
pip install "smolagents[toolkit]"Quick start
pip install "smolagents[toolkit]"- 1Install the package with pip install "smolagents[toolkit]".
- 2Import CodeAgent, a tool such as WebSearchTool, and your chosen model class.
- 3Create an InferenceClientModel or any other supported LLM instance.
- 4Instantiate CodeAgent with the selected tools and model, then call agent.run with your task.
- 5Optionally push the agent to the Hub using agent.push_to_hub for sharing.
Works with
LangChainMCPLiteLLMOpenAI APIPythonHugging Face Hub
smolagents: pros & cons
Pros
- +Extremely small core codebase for easy inspection and customization
- +Native code execution with sandbox options for safer runs
- +Works with virtually any LLM and many tool sources
- +Simple Hub integration for publishing and loading agents
Cons
- –Code-based action format may not suit every use case
- –Sandbox setup adds extra steps for production security
- –Still early-stage so documentation and examples are growing
Did you find this helpful?
Frequently asked questions
No, it supports local models, Hugging Face endpoints, and many other providers through LiteLLM.
User reviews
Verified reviews from the community shape this listing's rating.
Loading reviews…
Similar agents
Other agent frameworks options worth comparing.
Explore & compare smolagents
Data-driven comparisons, alternatives and rankings — kept current by our agents.
Featured in