Skip to content

Latest commit

 

History

History
248 lines (159 loc) · 12.1 KB

File metadata and controls

248 lines (159 loc) · 12.1 KB

Browser Agents

Browser agents turn the web into an API for LLMs. They wrap headless browsers, computer-vision models, and accessibility trees so an agent can navigate pages, fill forms, extract data, and execute multi-step workflows without a human in the loop.

This page covers the dominant open-source stacks: Browser Use, Stagehand, Skyvern, Playwright MCP, Steel, LaVague (deprecated), Midscene, and the broader split between DOM-vision and accessibility-tree approaches.

Why this matters

Most real-world data and actions still live behind web UIs. Browser agents solve the "last mile" problem: they let agents operate sites that have no API, dynamic front ends, or anti-bot protection. They sit between the planning layer and the evaluation layer: the planner decides what to do, the browser agent performs it, and the evaluation layer checks whether the resulting page state, download, or database change is correct. If you care about reliable agentic automation, browser agents are where theory crashes into reality.


Must Learn

browser-use/browser-use

Make websites accessible for AI agents. Automate tasks online with ease.

Why it matters: The de facto open-source reference for LLM-driven browser automation; huge ecosystem, active core, and the community web-ui project makes it easy to adopt.

Stats: ⭐ 104843 / 🍴 11550 / last commit 2026-07-15 / Active

Priority & difficulty: 🎯 5 / Intermediate

Tags: ai-agents, ai-tools, browser-automation, browser-use, llm, playwright, python

  • Learn how it maps high-level goals to Playwright actions and DOM observations.
  • Study the agent loop and how it handles failures, retries, and page state changes.
  • Compare its accessibility-tree approach with vision-only alternatives.
  • Use the project as a baseline before evaluating newer frameworks.

microsoft/playwright-mcp

Playwright MCP server

Why it matters: Microsoft's official MCP server exposes the Playwright API to LLMs via the Model Context Protocol, making it a key bridge between browser automation and agent tool stacks.

Stats: ⭐ 35112 / 🍴 2922 / last commit 2026-07-15 / Active

Priority & difficulty: 🎯 5 / Intermediate

Tags: mcp, playwright

  • Understand how MCP exposes browser tools as a standardized protocol.
  • Compare direct Playwright API use with MCP-mediated access.
  • Learn how to compose MCP servers for agents that need multiple tools.
  • Study the security and sandboxing implications of browser-as-a-tool.

browserbase/stagehand

The SDK For Browser Agents

Why it matters: A production-grade SDK built on Playwright/Puppeteer with deterministic extraction primitives; widely used by companies that need reliable agentic web data.

Stats: ⭐ 23516 / 🍴 1616 / last commit 2026-07-14 / Active

Priority & difficulty: 🎯 5 / Intermediate

Tags: agents, ai, llms, playwright, puppeteer, selenium

  • Study its extraction primitives and how they reduce non-determinism.
  • Compare Playwright vs. Puppeteer driver trade-offs in the same SDK.
  • Learn how it balances agentic flexibility with deterministic reliability.
  • Use it as a reference for building SDKs around browser automation.

Strongly Recommended

Skyvern-AI/skyvern

Automate browser based workflows with AI

Why it matters: Enterprise-focused browser-agent framework with strong visual understanding and workflow primitives; good for structured RPA-style tasks.

Stats: ⭐ 22243 / 🍴 2084 / last commit 2026-07-15 / Active

Priority & difficulty: 🎯 4 / Intermediate

Tags: ai, api, automation, browser, browser-automation, computer, gpt, llm

  • Learn how it structures workflows for repetitive enterprise tasks.
  • Study its visual understanding integration and when it beats DOM-only agents.
  • Compare its task DSL with Browser Use's goal-driven loop.
  • Evaluate its AGPL license against your deployment model.

trycua/cua

Open-source infrastructure for Computer-Use Agents. Sandboxes, SDKs, and benchmarks to train and evaluate AI agents that can control full desktops (macOS, Linux, Windows).

Why it matters: Desktop-grade Computer-Use Agent infrastructure with sandboxes and benchmarks; not limited to the browser, but a key adjacent category for agents that control full OS environments.

Stats: ⭐ 19706 / 🍴 1296 / last commit 2026-07-15 / Active

Priority & difficulty: 🎯 4 / Advanced

Tags: agent, ai-agent, apple, computer-use, computer-use-agent, containerization, cua, desktop-automation

  • Study how it sandboxes desktop sessions for safe agent execution.
  • Learn the difference between browser-only and full-desktop agent control.
  • Use its benchmarks to compare agent performance across environments.
  • Consider its macOS/Windows/Linux support for cross-platform agents.

browser-use/web-ui

🖥️ Run AI Agent in your browser.

Why it matters: Official web interface for browser-use, lowering the barrier to running AI agents in the browser without writing code.

Stats: ⭐ 16197 / 🍴 2715 / last commit 2026-05-15 / Active

Priority & difficulty: 🎯 4 / Beginner

Tags: ai-agent, browser-automation, browser-use-box, cloud-browser

  • Use it to quickly prototype browser automation without writing scripts.
  • Study how the UI serializes agent state and actions for non-technical users.
  • Compare its no-code approach with the core library's code-first model.
  • Learn how to deploy it as a demo or internal tool.

web-infra-dev/midscene

AI-powered, vision-driven UI automation for every platform.

Why it matters: Vision-driven UI automation from the Web Infra team; works across web, desktop and mobile, making it a versatile foundation for agentic UI control.

Stats: ⭐ 14078 / 🍴 1072 / last commit 2026-07-15 / Active

Priority & difficulty: 🎯 4 / Intermediate

Tags: ai, ai-test, browser-use, computer-use, gpt-operator, javascript, phone-use, testing

  • Learn how vision-driven assertions work across different platforms.
  • Compare its image-based element detection with accessibility-tree selectors.
  • Study its integration with testing workflows.
  • Explore how it handles dynamic UIs where DOM selectors are brittle.

nanobrowser/nanobrowser

Open-Source Chrome extension for AI-powered web automation. Run multi-agent workflows using your own LLM API key. Alternative to OpenAI Operator.

Why it matters: Open-source Chrome extension that brings multi-agent browser automation to end users with their own API keys; strong traction as an OpenAI Operator alternative.

Stats: ⭐ 13467 / 🍴 1413 / last commit 2025-11-24 / Moderate

Priority & difficulty: 🎯 4 / Beginner

Tags: agent, ai, ai-agents, ai-tools, automation, browser, browser-automation, browser-use

  • Study how it packages multi-agent workflows into a browser extension.
  • Learn the UX trade-offs of running agents inside the user's own Chrome profile.
  • Compare its client-side architecture with server-side browser agents.
  • Evaluate its API-key model and privacy implications.

steel-dev/steel-browser

🔥 Open Source Browser API for AI Agents & Apps. Steel Browser is a batteries-included browser sandbox that lets you automate the web without worrying about infrastructure.

Why it matters: Batteries-included browser sandbox API; focuses on infrastructure for running headless browsers at scale for agents and apps.

Stats: ⭐ 7339 / 🍴 948 / last commit 2026-07-12 / Active

Priority & difficulty: 🎯 4 / Intermediate

Tags: ai, ai-agents, ai-tools, browser-automation, llm

  • Learn how it abstracts browser infrastructure into an API.
  • Study its sandboxing and session management for scalable automation.
  • Compare hosted vs. self-hosted browser execution models.
  • Use it when you need reliable headless browsers rather than a full agent framework.

microsoft/Webwright

A simple SWE style browser agent framework that achieves SOTA results on long horizon web tasks.

Why it matters: Microsoft's SWE-style browser agent framework that reports SOTA results on long-horizon web tasks; good for research and complex automation.

Stats: ⭐ 5808 / 🍴 367 / last commit 2026-06-03 / Active

Priority & difficulty: 🎯 4 / Advanced

Tags: (none listed)

  • Study its SWE-style architecture and how it decomposes long tasks.
  • Compare its SOTA results with other frameworks on the same benchmarks.
  • Learn how it handles planning, memory, and verification over long horizons.
  • Use it as a research baseline for complex browser agents.

Comparison Table

Repo ⭐ Stars Language License Last Commit Activity Difficulty Ranking
browser-use/browser-use 104843 Python MIT 2026-07-15 Active Intermediate Must Learn
microsoft/playwright-mcp 35112 TypeScript Apache-2.0 2026-07-15 Active Intermediate Must Learn
browserbase/stagehand 23516 TypeScript MIT 2026-07-14 Active Intermediate Must Learn
Skyvern-AI/skyvern 22243 Python AGPL-3.0 2026-07-15 Active Intermediate Strongly Recommended
trycua/cua 19706 HTML MIT 2026-07-15 Active Advanced Strongly Recommended
browser-use/web-ui 16197 Python MIT 2026-05-15 Active Beginner Strongly Recommended
web-infra-dev/midscene 14078 TypeScript MIT 2026-07-15 Active Intermediate Strongly Recommended
nanobrowser/nanobrowser 13467 TypeScript Apache-2.0 2025-11-24 Moderate Beginner Strongly Recommended
steel-dev/steel-browser 7339 TypeScript Apache-2.0 2026-07-12 Active Intermediate Strongly Recommended
microsoft/Webwright 5808 Python MIT 2026-06-03 Active Advanced Strongly Recommended

Browser Agent Architecture

graph TD
    P[Planner / LLM] -->|task + context| BA[Browser Agent]
    BA -->|actions| HB[Headless Browser / Playwright]
    BA -->|DOM / vision| OBS[Observation Layer]
    OBS -->|accessibility tree| AT[Accessibility Tree]
    OBS -->|screenshots| CV[Computer Vision Model]
    AT -->|parsed state| BA
    CV -->|element coordinates| BA
    HB -->|page state| OBS
    BA -->|result + artifacts| EVAL[Evaluation Layer]
    EVAL -->|pass/fail| P
Loading

Recommended Learning Order

  1. browser-use/browser-use — start with the most adopted open-source browser agent.
  2. browser-use/web-ui or nanobrowser/nanobrowser — get a no-code or extension feel for what end users see.
  3. microsoft/playwright-mcp — understand how browser tools plug into the MCP protocol.
  4. browserbase/stagehand — study deterministic extraction and SDK design.
  5. Skyvern-AI/skyvern — learn enterprise workflow primitives and visual understanding.
  6. steel-dev/steel-browser — explore browser-as-infrastructure and sandboxing.
  7. web-infra-dev/midscene — compare vision-driven UI automation across platforms.
  8. trycua/cua — expand from browser to full desktop computer-use agents.
  9. microsoft/Webwright — study SOTA long-horizon browser agents for research.

When to use which

  • Quick demos and prototyping: browser-use/web-ui or nanobrowser.
  • Production agent workflows: browser-use, stagehand, or Skyvern.
  • MCP-native tool integration: microsoft/playwright-mcp.
  • Desktop / OS-level control: trycua/cua.
  • Vision-first or cross-platform UI: midscene.
  • Scalable headless browser infrastructure: steel-browser.
  • Long-horizon research tasks: microsoft/Webwright.

Cross-links