Smolagents

Smolagents is Hugging Face’s minimalist open-source library for building agents that write and execute code, in a few lines of Python.

Overview

Where most frameworks add abstraction layers, Smolagents stays deliberately tiny, a few thousand lines you can actually read. Its signature idea is the code agent: instead of emitting JSON tool calls, the agent writes executable Python that runs in a sandboxed interpreter, which turns out to be a more efficient way to express complex actions.

The library supports any model from Hugging Face’s Inference Providers to OpenAI-compatible endpoints and local models, and includes both code agents and standard tool-calling agents. It integrates with MCP servers, supports multi-agent hierarchies, and has become a favorite for teaching, prototyping and embedding agents inside larger ML pipelines. The project maintains a set of examples that double as a practical education in agent design, from simple tool loops to multi-agent hierarchies. Its small surface area also means upgrades rarely break your code, a real advantage over larger frameworks.

Key Features

• Code agents that write and run Python instead of emitting JSON tool calls
• Sandboxed local execution plus secure remote sandboxes
• Model-agnostic: Hub inference, OpenAI-compatible APIs and local models
• MCP client support for connecting external tool servers
• Multi-agent hierarchies with managed agents and tool agents
• Extremely small, readable codebase that is easy to audit and extend

Pricing

Smolagents is open source under the Apache license and entirely free. You only pay for model inference, and pairing it with local models or free inference credits makes it possible to build and run agents at no cost. There are no fees, tiers or usage dashboards to manage, and pairing Smolagents with cheap local models keeps experiments essentially free.

Who Is It For

Smolagents suits Python developers and ML engineers who want agent capabilities without framework lock-in, and educators who need a codebase small enough to explain in one sitting. It is ideal for embedding lightweight agents inside existing data and ML workflows. It is equally at home inside data pipelines and notebooks, where a lightweight agent can automate steps of an analysis without heavyweight dependencies.

FAQ

Is Smolagents free to use?

Yes, it is free and open source. There is no paid tier; your only cost is the model API or hardware you choose to run inference on.

Does Smolagents offer an API?

Smolagents is a library rather than a hosted service, so you expose agents through your own API layer. It can consume external tools over MCP and HTTP, making integration with existing services straightforward.

How does Smolagents compare to alternatives?

Smolagents trades features for clarity: it lacks the enterprise platform around LangGraph or CrewAI, but its code-action approach often solves in three lines what tool-calling agents need thirty for. For quick, hackable agents it is hard to beat. If your agent work is research-flavored or embedded in existing Python code, Smolagents provides most of the value of larger frameworks with a fraction of the complexity.