Type: agent framework, requiring developer integration. Hugging Face smolagents offers CodeAgent and ToolCallingAgent, model backends through hosted inference, LiteLLM, Transformers and Ollama, and tool integrations including MCP. Use cases documented include agentic retrieval, SQL workflows and browser agents. Library license: Apache-2.0; model endpoints and remote execution services have separate costs. Sources: https://huggingface.co/docs/smolagents/en/index; https://github.com/huggingface/smolagents/blob/main/LICENSE.
Getting started
Official quickstart: pip install 'smolagents[toolkit]'. Configure a model backend and permitted tools in Python. Optional toolkit includes default tools such as web search. Commands were documented, not executed.
Security & permissions
CodeAgent executes model-generated code locally by default using the project executor. The documentation warns about this risk and describes sandbox options. Configure permitted tools, imports and credentials deliberately. Source: https://huggingface.co/docs/smolagents/en/tutorials/secure_code_execution