What Is a Computer Use Agent (CUA)?

Paste into ChatGPT, Claude, or any LLM

On this page

A Computer Use Agent (CUA) is an AI agent that logs into a portal as the user, navigates the UI, handles MFA and CAPTCHAs, and returns structured data. Different from web scraping in three ways: the user grants permission (no ToS conflict), it operates on data the user already owns, and it works on portals that have no public API. Anthropic's Computer Use, OpenAI's Operator, Skyvern (85.8% WebVoyager), and Browser Use (89% WebVoyager, the leading open-source option at 78k+ GitHub stars) are the current production-grade implementations.

WebVoyager — Browser Use	89% (open-source, leading)
WebVoyager — OpenAI Operator/CUA	87% (controlled VM environments)
WebVoyager — Skyvern	85.8%
WebVoyager — Anthropic Computer Use	56% (real desktop environments)
Browser Use stars	78k+ (May 2026)

CUA vs web scraping — different categories

Classic web scraping: anonymous HTTP/browser requests against public pages. You see what a logged-out visitor sees. Per-request billing, high concurrency, low latency.

Computer Use Agent: the user grants the agent permission, the agent logs in as the user. Per-task billing, low concurrency (one VM per task), higher latency (30 seconds to several minutes per task), clean legal posture because the data belongs to the user.

Mental model: CUAs are "Plaid for any website" — bringing the open-banking pattern (user permission + structured data extraction) to portals that have no public API. Utility bills, bank statements, payroll exports, insurance claims, tax filings, e-commerce backend orders.

When each one wins

Use a CUA when: the data lives behind a login the user owns; the portal has no API; the workflow requires MFA, step-up authentication, or human-grade UI navigation; you need one-off retrievals at human scale (5 documents per user × 200 users).

Use traditional scraping when: the data is public-facing (e-commerce listings, SERPs, social media, news, real estate); you need low latency (sub-second); you need high concurrency (100+ parallel requests); the per-request cost matters (scraping is 10–100× cheaper for the same data when both work).

At 100k items, scraping might cost €20–€100 on Scrappey. Running 100k CUA tasks could cost $5,000–$100,000 depending on platform. The cost gap is what makes the categories distinct rather than competing.

The market in May 2026

Anthropic Computer Use — direct API, operates the actual host machine via mouse and keyboard primitives. Best for custom agent pipelines. Scores 56% on WebVoyager because it operates real desktops with all their messy complexity, not controlled browser-only VMs.

OpenAI Operator (CUA) — hosted product with built-in browser capabilities, scores 87% on WebVoyager in controlled environments.

Skyvern — open-source (YC-backed), Vision-LLM driven, 85.8% WebVoyager. Strong on invoice retrieval, job applications, government forms, insurance quotes. Both cloud-hosted and self-hostable.

Browser Use — the leading open-source browser-only agent at 89% WebVoyager, 78k+ GitHub stars. Connect any LLM, run locally or self-hosted. Supports OpenAI, Anthropic, Gemini, Ollama for local models.

Deck — managed VMs with credential vault, SOC 2, positioned as "Plaid for any website" with 100k+ utility provider integrations.

Code example

python

# Browser Use (open source, 89% WebVoyager) — the standard open-source CUA
# pip install browser-use

from browser_use import Agent, ChatOpenAI

agent = Agent(
    task="Log into example-utility.com using the credentials in the env, "
         "navigate to billing history, download the last 12 months of "
         "statements as PDFs, and return the file paths.",
    llm=ChatOpenAI(model="gpt-4o"),    # or ChatAnthropic, Gemini, local Ollama
)

result = agent.run()
# Returns structured output of the task — file paths, total billed, dates.

Related terms

What Is AI Web Scraping?

AI web scraping is an approach that replaces CSS selectors with natural-language prompts, LLM-based extraction, and Markdown-first output. I…

What Is a Web Scraping API?

A web scraping API is a managed HTTP service that fetches a target URL on your behalf and returns the rendered HTML, JSON, or parsed data. I…

What Is a Headless Browser?

A headless browser is a real web browser — Chrome, Firefox, or WebKit — that runs without a visible graphical interface, controlled entirely…

What Is Firecrawl?

Firecrawl is an AI-native scraping API that takes a URL and returns clean Markdown or JSON — no CSS selectors, no XPath, no page parsing. It…

How to Scrape Infinite-Scroll Pages

Scraping infinite-scroll pages means programmatically triggering the scroll events that load new content, waiting for that content to render…

Concept map

How Computer Use Agent connects

The terms most directly tied to this one. Hover a node to see its neighbours, click to preview, drag to rearrange.

0 terms · 0 connections

You are here · Web Scraping APIs

Tools & solutions for this topic

Frequently asked questions

Is a CUA the same as a headless browser?

A headless browser is the substrate — Chrome or Firefox running without a visible UI. A CUA is an AI agent layered on top of a headless browser (or sometimes a real desktop) that interprets the page visually or via DOM, decides what to do next, and acts. The headless browser is the body; the CUA is the brain.

Why is Anthropic's Computer Use score lower than OpenAI's?

WebVoyager measures browser-only tasks in controlled environments. OpenAI Operator runs in optimised browser-only VMs and scores 87%. Anthropic's Computer Use is more general — it can operate any desktop application, not just browsers — and was benchmarked on the harder real-desktop environment, scoring 56%. They are solving overlapping but not identical problems.

Should I use Browser Use or Skyvern?

Browser Use if you want the highest open-source WebVoyager score and the most active community (89%, 78k+ stars). Skyvern if you specifically want a Vision-LLM driven agent that uses screenshots rather than DOM (85.8%, useful when the DOM is dynamically obfuscated). For invoice retrieval and form-filling specifically, Skyvern has more documented production deployments.

When is a CUA cheaper than a managed scraping API?

Almost never for public data. CUAs are billed per-task ($0.05–$1 each); managed scraping APIs are billed per-request ($0.0002–$0.003 each). The CUA premium pays for user-consent flow, MFA handling, and login-required data — none of which apply to public data. Pick CUAs for portals behind a user login; scraping APIs for everything else.

Last updated: 2026-05-26