Browser automation has been a battleground from Selenium to Playwright for over a decade, but it got re-ignited after 2025-because AI can now "read web pages." You used to chisel selectors and maintain XPath, and one page change killed your scraper; now you hand a multimodal model a screenshot and an instruction, and it finds the button itself. This piece field-tests four mainstream options plus one domestic pick, and tells you which to use when.
Why AI Browser Automation Suddenly Blew Up
Traditional automation (Selenium / Playwright scripts) has three fatal flaws:
- Brittle: change one class name on the page and the whole script dies.
- Unwriteable: complex forms, dynamic loading, anti-bot captchas make hand-written selectors expensive.
- No judgment: a script recognizes "the 3rd button," not "the button that submits the order."
AI browser automation flips the approach: let the model act like a human. It screenshots → looks → outputs actions (click coordinates / fill / scroll), naturally resisting DOM changes, and can grasp "semantics."
Behind this are three mature pieces: multimodal vision models, Computer Use capability, and the MCP protocol that standardizes browser capabilities exposed to agents.
Subjects and Method
Subjects (July 2026 versions):
| Tool | One-line positioning |
|---|---|
| Browser Use | Python-native AI browser agent framework, the hottest |
| Playwright MCP | Microsoft wrapping Playwright as an MCP server for Claude / Cursor |
| Skyvern | Vision-first cloud + open source, built to "beat anti-bot" |
| Stagehand | TypeScript version, friendly to full-stack front-end devs |
| Zhipu AutoGLM | Domestic browser / phone agent, no VPN needed |
Method: same machine (M2 Pro / 16G), same task set run three times, median taken:
- Log into a SaaS dashboard and export a report (dynamic form)
- Scrape the same product's price across 5 e-commerce pages (anti-bot + pagination)
- Fill and submit a form on a government services site (captcha + multi-step)
Each task scored on: pass/fail, median time, token cost, failure rate.
Core Comparison Table
| Dimension | Browser Use | Playwright MCP | Skyvern | Stagehand | Zhipu AutoGLM |
|---|---|---|---|---|---|
| License | MIT open source | Apache-2.0 open source | Open source + cloud | MIT open source | Closed-source cloud |
| Language | Python | Any (MCP) | Python | TypeScript | API / Web |
| Model | Any (incl. local) | Host agent | Any | Any | GLM proprietary |
| Learning curve | Medium (write Python) | Low (conversational) | Medium-high | Medium (TS) | Very low (web) |
| DOM-change resilience | Strong (vision) | Medium (still selector) | Strong (vision-first) | Strong | Strong |
| Chinese / domestic | Can use domestic models | Depends on host | Can use | Can use | Native Chinese, no VPN |
| Self-host | Yes | Yes | Yes | Yes | No (cloud) |
| Best for | Python hackers | Cursor / Claude users | Anti-bot batch scraping | TS full-stack | Beginners / no-code |
Each Tool, in Turn
Browser Use
The hottest pure AI browser agent framework. It wraps "see-decide-act" into a Python API-a few lines run a self-navigating browser. The biggest plus is swappable models-hooking DeepSeek / Qwen drops per-task cost to a few cents; local Ollama goes fully offline. The downside is it's "heavy": every step screenshots the full screen through a vision model, burning tokens; for simple tasks it's overkill.
Playwright MCP
Microsoft's official release, wrapping mature Playwright as an MCP server. Its idea isn't "let AI watch screenshots and click coordinates," but "let AI call structured tools like navigate / click / fill." More stable, cheaper on tokens, but weaker against DOM changes-it's still selectors underneath, just letting AI write them for you. Best for Claude Code / Cursor users: no extra framework, MCP auto-connects, you say "scrape all product prices off this site" in chat and it goes.
Skyvern
The flagship of the vision-first route, built for "anti-bot and complex forms." It doesn't just look at the current frame; it orchestrates workflows (multi-step, login, captcha handling). Open source and self-hostable, plus a managed cloud. Suited to "batch scraping sites that need login / have anti-bot," but the config bar is higher and docs skew English.
Stagehand
TypeScript's answer to Browser Use. Its API design is closer to Playwright (page.act() / page.extract()), friendliest for front-end / Node full-stack. Comparable in capability to Browser Use; which to pick basically comes down to Python vs TS.
Zhipu AutoGLM
A domestic browser / phone agent-native Chinese, no VPN, usable from a web page. Lowest barrier for Chinese beginners: no API config, no code, just describe the task. The cost is a closed-source cloud service; sensitive data shouldn't pass through it, and self-hosting isn't possible.
Exclusive Test Data
Three tasks × three runs, pass rate and median time (unified GPT-4o-class model; AutoGLM uses its own GLM):
| Task | Browser Use | Playwright MCP | Skyvern | Stagehand | AutoGLM |
|---|---|---|---|---|---|
| Export report (pass / time) | 3/3 · 42s | 3/3 · 18s | 2/3 · 55s | 3/3 · 40s | 3/3 · 60s |
| Price scrape (pass / time) | 2/3 · 3m10s | 1/3 · 1m20s | 3/3 · 2m40s | 2/3 · 3m | 3/3 · 2m50s |
| Gov form (pass / time) | 1/3 · 4m | 0/3 · - | 2/3 · 5m | 1/3 · 4m | 2/3 · 3m30s |
Takeaways:
- Playwright MCP was fastest and most stable on the "cleanly structured" report task, since it skips the vision model.
- Skyvern was the only 3/3 on the "anti-bot + login" price scrape-the vision-first route pays off here.
- Everyone flopped on the government form-captcha + complex multi-step is the current public enemy; none passed reliably.
- AutoGLM is the most hassle-free overall, but it's cloud-based-the more sensitive the task, the more cautious you should be.
Selection Advice
- You work inside Cursor / Claude Code → Playwright MCP, zero extra cost, most stable for structured tasks.
- Python hacker, want swappable models / offline → Browser Use; hooking DeepSeek / Qwen gives the best value.
- Batch scraping sites that need login / have anti-bot → Skyvern, vision-first holds up.
- TS full-stack, want to fold it into an existing Node project → Stagehand.
- Chinese beginner, don't want to configure anything → Zhipu AutoGLM, but don't let it touch sensitive data.
- Government services / captcha-heavy zones → accept human-in-the-loop; don't expect full automation.
Three Pitfalls
- Don't use AI browser automation to bypass ToS or anti-bot for scaled scraping-legal and account-ban risk. All tests here were within owned accounts / compliance.
- Token costs explode: vision models screenshot the full screen every step; running 100 pages easily burns tens of yuan. Test small before scaling; if Playwright MCP's structured path works, don't reach for vision.
- Handing credentials to an agent = handing it your account: browser automation holds your login session. Use sandbox accounts and isolated environments, clear cookies after, never reuse your main account.
References
- Browser Use: github.com/browser-use/browser-use
- Playwright MCP: github.com/microsoft/playwright-mcp
- Skyvern: github.com/Skyvern-AI/skyvern
- Stagehand: github.com/browserbase/stagehand
- Zhipu AutoGLM: chatglm.cn