xiry.

Development · Web app

smolagents

Lightweight open-source Python library for building code-writing and vision AI agents

From the Xiry library · Updated 10 October 2026

Visit huggingface.co ↗

smolagents is Hugging Face's lightweight agent framework that lets AI agents write and execute code actions rather than relying solely on JSON tool-calling. With added support for Vision-Language Models (VLMs), developers can create autonomous web-browsing agents that view screenshots, dismiss popups, and interact with live web pages dynamically.

What it's for

  • Build autonomous browser agents to navigate web apps and extract data visually
  • Inspect web pages and extract data hidden behind dynamic JavaScript or visual elements
  • Dismiss modal popups, cookie banners, and interact with complex page elements via automated actions
  • Process multi-page visual documents like PDFs using vision models in an agent loop

Features

  • Native support for Vision-Language Models (VLMs) via step callbacks and screenshot logging
  • CodeAgent architecture where models output executable Python code instead of fixed JSON schemas
  • Easy integration with browser automation tools like Helium and Selenium
  • Dynamic memory management that prunes old screenshots to keep model contexts lean
  • Built-in support for open-source VLMs (SmolVLM, Qwen2-VL) as well as proprietary vision models

Pricing

Open source. Free and open-source (Apache-2.0 license); underlying model API calls may incur standard provider fees

Details

Type
Web app
Category
Development
Topic
Agent Frameworks
Open source
Yes
Platforms
Code library, Linux, macOS, Windows
For
Python developers and AI engineers building autonomous web scraping and multi-modal agents
Alternative to
OpenAI Operator, Browser Use, LangChain, CrewAI
Tags
smolagents, huggingface, ai agents, browser automation, vision models, vlm, python

Why it was saved

  • @mervenoyann, 26 January 2025: Lets you build autonomous browser agents in Python: CodeAgent + browser tools (get_driver, scroll_down, click, close_popups, final_answer), with a prompt pattern giving stepwise instructions and helium_instructions baked in. Demo in the video: agent navigates GitHub Trending, opens a repo, inspects the author profile and returns '308' commits in a year.
  • @mervenoyann, 26 January 2025: Companion guide linked by the author explaining the VLM browser-agent capability shown in the demo.

Compare

Related

  • ai-hedge-fund

    Proof-of-concept AI multi-agent trading system with paper trading and backtesting.

  • Eliza

    Open-source multi-agent framework for building autonomous social and on-chain AI agents

  • memary

    Open-source persistent memory layer for autonomous AI agents using knowledge graphs

  • Self-Improving Agent Loop via skill.md

    A workflow pattern where agents document completed tasks into reusable shortcuts

  • awesome-llm-apps

    Open-source collection of runnable AI agent, skill, and RAG templates

  • Browser Use Cloud

    Cloud-hosted stealth browser infrastructure and execution platform for AI agents.

All 5 in Agent Frameworks →

Updated