Browser Automation
SaaS, open source and skills for browser automation work — with the limits of each written down.Before you pick
What decides it
Whether it runs in your logged-in browser or a fresh remote one. Login walls, expiring cookies and bot checks decide more than model quality.
What to avoid
Page text and your instructions reach the model through one channel, and it acts with your logged-in privileges. Restrict which sites it may open.
What is changing
Agents are learning to record a successful run and replay it as deterministic steps, calling the model only when the page changes.
SaaS
Hosted products you sign up for.Leading
Established picks, at the top of today's aggregate.
- 1ChatGPT Agent (Operator)Consumer-facing agent that operates a real browser inside ChatGPT to complete multi-step web tasks end to end.
- 2Claude for ChromeAnthropic's browser agent that drives Chrome to read pages and take actions on the user's behalf.
- 3Perplexity CometAI-native consumer browser that combines answering and agentic actions across tabs for research-heavy workflows.
- 4Gemini in ChromeGoogle's built-in Chrome agent, generally available to Workspace users for in-browser assistance and actions.
- 5OpenAI AtlasOpenAI's standalone AI browser for agentic browsing, competing directly with Comet and Dia.
Emerging
Newer challengers, surfaced by two or more sources.
- 1BrowserCloudHosted pool of headless Chrome instances you point Puppeteer, Playwright or Selenium at, with proxies, stealth fingerprints and parallel sessions.
- 2Open Claude in ChromeClean-room reimplementation of Anthropic's Chrome extension exposing the same 18 MCP browser tools, without the 58-domain blocklist, on any Chromium browser.
- 3AutomatioNo-code point-and-click builder for scraping and form-filling bots, now fronted by chat agents that run them on schedule in the cloud.
Open source
Repositories you run yourself.Leading
- 1Browser UseThe recognized leader for AI browser agents (~107k★, ~89% WebVoyager): gives the LLM full control of the browser through an agent loop — the default for building custom browser agents.
- 2PlaywrightMicrosoft's foundational browser-automation framework (~93k★) — the deterministic substrate many AI agents (Stagehand, Skyvern) build on; the de-facto standard for scripted control and E2E testing.
- 3StagehandPlaywright-based framework (~24k★) exposing structured act/extract/observe actions — deterministic where possible, AI only when needed.
- 4SkyvernVision + LLM open-source agent (~23k★) resilient to layout changes; leads on form-filling (~85.85% WebVoyager).
- 5NanobrowserOpen-source Chrome extension (~13.5k★) for multi-agent web automation, running on your own LLM API keys as a free OpenAI-Operator alternative.
Emerging
- 1vercel-labs/agent-browserRust CLI that drives Chrome over the DevTools Protocol and returns compact accessibility trees instead of raw DOM, keeping agent token use low.
- 2zhu1090093659/deepseek-ppBrowser extension that turns DeepSeek's web chat into an agent workspace with MCP tools, memory, skills, and scheduled tasks acting on the open tab.
- 3AIPexStudio/AIPexChrome extension that lets MCP clients drive the browser you already use, sending DOM snapshots rather than screenshots over a local WebSocket daemon.
- 4feder-cr/Jobs_Applier_AI_Agent_AIHawk
- 5hyperbrowserai/HyperAgentTypeScript library that wraps Playwright with natural-language page methods and cached action replay, running locally or on Hyperbrowser's hosted browsers.
Skills
Skills you drop into a coding agent.Emerging
- 1pencil20388-eng/browser-automation-skillsSix drop-in SKILL.md files giving Claude Code, Cursor or Codex reference patterns for Playwright, Selenium, Puppeteer, scraping and antidetect profiles.
- 2wrsmith108/dev-browser-claude-skillClaude Code skill keeping named Playwright pages alive on a background server, so cookies and DOM state survive between short scripts; fork of dev-browser.
- 3Salmonbird/evoui-browser
- 4scrapeless-ai/scrapeless-agent-browserCLI that drives Scrapeless-hosted cloud browsers using short element references instead of selectors; needs an API key, and residential proxies come built in.
- 5appautomaton/webmatonFive symlinkable Python skills for Claude, Codex and OpenCode covering nodriver and Playwright control, DevTools debugging, cited web research and HTML-to-Markdown capture.