Hardcore Reviews
Hardcore Reviews

AI Computer Use Agents Compared: Claude vs OpenAI vs Mariner vs 3 Open-Source Picks

A comparison of 6 AI computer-use agents: closed-source Claude Computer Use / OpenAI Operator / Google Project Mariner plus open-source trycua/cua (21,113 stars) / simular-ai/Agent-S (12,147 stars) / microsoft/fara (6,140 stars). Includes a capability matrix and a pricing/availability table, selection guidance, and 5 FAQs. Closed-source tools are restricted in China; open-source ones run locally. Representative comparison, not a personal benchmark.

Published August 11, 20269 min read
<!-- ai-computer-use-agents-comparison-review | review | AI Computer Use Agents Compared: Claude vs OpenAI vs Mariner vs 3 Open-Source Picks -->

In October 2025, Anthropic released Computer Use, letting Claude "see the screen, click the mouse, type on the keyboard" to operate a computer directly. Less than a year later, OpenAI's Operator/CUA and Google's Project Mariner followed suit, and the open-source side produced three 10k-star contenders: trycua/cua at 21k stars, simular-ai/Agent-S at 12k, and microsoft/fara at 6k. AI has officially crossed from "can chat" to "can work"-but if you actually want to pick one to operate your computer, which of the three closed-source giants and three open-source challengers is more reliable? This review lays them all side by side.

Scope first: this is based on each tool's GitHub README, official site, and public benchmarks, as of 2026-08-11. Open-source star counts verified via GitHub API (cua 21,113 stars, Agent-S 12,147 stars, fara 6,140 stars); closed-source tools have no public repos so no star counts, and pricing and availability are per each official site. This is a representative comparison, not a personal full-scale benchmark. This site also published the Gemini Robotics ER-2 hotspot (another thread in embodied AI), the cua open-source resource (dedicated open-source piece), and the computer-use agent build SOP (how to build after you pick)-this piece covers the "selection" step.

One: Know the Difference-Computer Use Is Not RPA, Nor a Personal Assistant

Three easily confused concepts, demarcated first. Computer Use / GUI Agent: the AI uses visual recognition of screen screenshots plus simulated mouse and keyboard input to operate a graphical interface like a human-any button you can click, any form you can fill, it can operate, without depending on the software exposing an API. RPA (Robotic Process Automation): relies on recording or scripts to click along a fixed path, with no "understanding" capability-once the flow changes, it breaks. Personal assistants (e.g., OpenClaw): run locally, operating email/calendar/files/terminal via API calls and scripts rather than "seeing the screen"-this site covered it in the OpenClaw resource; it does not belong to GUI computer-use and is excluded from this comparison. This piece only compares the "watch screen, move mouse" category.

Two: Capability Matrix-Six-Tool Specs Table

Six representative tools, compared across positioning, open-source status, supported OS or browser, and core capability.

ToolPositioningOpen sourceSupported OS/BrowserCore capability
Claude Computer UseDesktop-first GUI agentNoDesktop (cross-app) + WebScreenshot recognition + mouse/keyboard simulation, cross-app workflows
OpenAI Operator/CUABrowser-first cloud agentNoCloud virtual browserWeb task automation, cloud-isolated execution
Google Project MarinerIn-browser agentNoChrome browserBrowser-anchored automation, Google ecosystem integration
trycua/cuaCross-OS open-source computer-useYes (MIT)macOS/Windowscomputer-use 2.0, cross-OS local control
simular-ai/Agent-SOpen-source agentic frameworkYes (Apache-2.0)macOS/Windows/LinuxFramework that uses computers like a human, customizable
microsoft/faraOpen-source computer-use modelYes (MIT)Cross-platform (model-level)Fara1.5 frontier computer-use agent model

A few clarifications. First, the three closed-source players take different architecture routes: Anthropic is desktop-first (broadest cross-app reach), OpenAI is browser-first (most secure via cloud isolation but limited to web), and Google is browser-anchored (in between, winning on ecosystem). Second, the three open-source tools occupy different layers: cua is a ready-to-use cross-OS tool, Agent-S is a customizable framework, and fara is model-level (you build the screenshot and operation pipeline yourself). Third, "open source" does not mean "zero cost"-the software is free, but you pay for LLM API calls or local compute.

Three: One by One-Each Tool's Best Range

Claude Computer Use: strongest cross-app, desktop-first. From Anthropic, provided as a computer-use tool via the Claude API. The model takes a screenshot to "see" the screen, outputs mouse coordinates and keyboard instructions, and directly operates the local desktop-without depending on software APIs; theoretically any software you can use, it can use. Its strength is cross-app versatility: scraping data from a web page into Excel, moving information across windows, operating legacy software with no API-this is its home turf. Shortcomings: slow (each step requires screenshot + reasoning + action), non-trivial error rate on complex multi-step tasks, and Chinese-language software compatibility remains unverified. Best for: cross-app information transfer, complex desktop workflows. Pricing is per Claude API usage, per the official site. China availability: API requires VPN, restricted.

OpenAI Operator/CUA: most stable for browser tasks, cloud-isolated. OpenAI's Computer-Using Agent runs all operations in a server-side virtual browser-it never touches your local computer, which is secure but limits it to web apps. Its strength is stable web task execution: booking flights, filling forms, price comparison, information retrieval, all run in the cloud with results returned. Shortcoming: the browser-first architecture limits desktop app operation-it can't reach your local Excel or WeChat. Best for: web task automation, online forms, information retrieval. Pricing via ChatGPT Pro subscription, per the official site. China availability: ChatGPT is unavailable in China.

Google Project Mariner: in-browser agent, wins on ecosystem. Google's solution, anchored in the browser, deeply integrated with the Google ecosystem (Gmail, Docs, Search). Its strength is smooth task execution within Google services. Shortcomings: weak desktop operation capability, and abilities are locked inside the Google ecosystem. Best for: browser automation for heavy Google ecosystem users. Pricing per the official site. China availability: Google services are restricted in China.

trycua/cua: open-source cross-OS, ready to use. 21,113 stars, MIT license. computer-use 2.0, supporting macOS and Windows cross-OS control-the highest-starred and most-watched open-source contender, with a dedicated open-source piece on this site. Its strengths are cross-OS + open source + self-hostable: you can run it entirely locally, with data never leaving your machine. Shortcomings: requires self-installation and configuration, model costs are on you, and stability varies. Best for: developers who want local, cross-OS automation. China availability: open source, runs locally, unrestricted.

simular-ai/Agent-S: open-source framework, customizable. 12,147 stars, Apache-2.0. Positioned as "an agentic framework that uses computers like a human"-not just a tool, but a framework for building custom computer-use agents, supporting macOS/Windows/Linux. Its strength is customizability: modifying interaction logic, plugging in your own model, adding safety boundaries-framework-level solutions are the most flexible. Shortcoming: higher barrier than cua, requires writing code. Best for: developers building custom computer-use agents. China availability: open source, runs locally, unrestricted.

microsoft/fara: Microsoft's official open-source model. 6,140 stars, MIT. Fara1.5 is a frontier computer-use agent model-note that it is a "model" rather than a "tool"; you build the surrounding screenshot and operation pipeline yourself. Its strengths are official Microsoft backing plus model-level capability: suited for researching frontier computer-use models or swapping models in your own pipeline. Shortcoming: model-level means more integration work, not ready to use like cua. Best for: research and engineering teams working with frontier computer-use models. China availability: open source, model can run locally.

Four: Pricing and Availability

The second table looks at go-to-market: free tier, paid, and China availability. Closed-source pricing is uniformly "per official site"-no invented numbers.

ToolFree tierPaidChina availability
Claude Computer UseNoneClaude API pay-per-use (per official site)Restricted, API needs VPN
OpenAI Operator/CUANoneChatGPT Pro subscription (per official site)Unavailable
Google Project MarinerNonePer official siteRestricted
trycua/cuaOpen-source freeModel costs on you (API or local compute)Available, runs locally
simular-ai/Agent-SOpen-source freeModel costs on youAvailable, runs locally
microsoft/faraOpen-source freeModel costs on youAvailable, runs locally

The conclusion is straightforward: all three closed-source tools are restricted or unavailable in China, while all three open-source tools run locally without restriction-this is the biggest dividing line for developers in China. Closed-source wins on out-of-the-box readiness and model capability; open-source wins on autonomy and zero software cost.

Five: Selection Advice-Pick by Need

Direct conclusions by your need.

Your needTop pickReason
Cross-app desktop workflow automationClaude Computer UseDesktop-first, strongest cross-app capability
Web task automation (booking/forms/comparison)OpenAI Operator/CUABrowser-first, cloud isolation most stable
In-browser automation within Google ecosystemGoogle Project MarinerDeep integration with Google services
Local cross-OS automation (macOS/Windows)trycua/cuaHighest open-source stars, ready to use
Build a custom computer-use agentsimular-ai/Agent-SFramework-level, most customizable
Research frontier computer-use modelsmicrosoft/faraMicrosoft-backed, model-level
China-available + data stays localAny of the three open-sourceRuns locally, unrestricted

Most developers' actual combo: one closed-source tool as primary (Claude or Operator) + one open-source tool as local backup (cua is easiest to start with). No need to try all six-pick by your scenario and region.

Six: Three Pitfalls

One: Don't take benchmarks as the whole story-real success rates are still low. Public benchmark reports show that the best-performing open-source agent (Agent-S's S2 version) achieved only about 34.5% success rate within 50 steps-nearly two-thirds of tasks could not be completed. All agents perform poorly when limited to 15 steps; multi-step reasoning and planning remain the bottleneck. Allowing more steps improves success rates, but at the cost of higher latency and cost. Bottom line: today's computer-use agents can work, but don't expect them to get it right in one shot-plan for human review.

Two: Closed-source isn't "better," open-source isn't "worse." Closed-source wins on model capability and out-of-the-box readiness, but is unavailable in China, sends data off your machine, and has opaque pricing; open-source wins on autonomy and local execution, but requires self-setup, model costs on you, and stability varies. Choose by region, privacy requirements, and technical ability-don't just take sides along "closed vs. open."

Three: Always require human confirmation for sensitive operations. The ability to operate a computer equals the ability to cause harm. Banking, payments, and internal company systems should always be operated manually or at least confirmed by a human before execution. The safety red line in the Computer Use era is: sandbox first + operation allowlist + human confirmation for sensitive steps. This site's computer-use agent build SOP has the complete secure setup process.

Seven: FAQ

Q: What's the difference between Computer Use and RPA? A: RPA relies on recording or scripts to click along a fixed path, with no "understanding" capability-once the flow changes, it breaks. Computer Use uses a visual model to "see" the screen and reason about actions, handling interface changes and non-standard scenarios. For simple, repetitive, fixed-flow tasks, choose RPA; for tasks that need to "understand the screen before acting," choose Computer Use.

Q: Are the three closed-source tools usable in China? A: All are restricted or unavailable. Claude Computer Use's API requires a VPN and is restricted, OpenAI Operator/CUA is unavailable in China via ChatGPT Pro, and Google Project Mariner is restricted with Google services. Developers in China should look at the three open-source options first-they run locally without restriction.

Q: How do I choose among the three open-source tools? A: Depends on what you need. cua (21,113 stars) is ready to use, cross-OS, and easiest to start with; Agent-S (12,147 stars) is a framework, suited for developers building custom agents; fara (6,140 stars) is a model, suited for teams researching frontier computer-use models. Beginners start with cua, advanced users pick Agent-S, researchers pick fara.

Q: Can I use it for banking or payments? A: Not recommended. The ability to operate a computer equals the ability to cause harm-banking, payments, and internal company systems should always be operated manually or at least confirmed by a human before execution. Computer Use is suited for information transfer, form filling, and data cleanup-scenarios where "nothing major can go wrong." For anything involving real money, keep a human in the loop.

Q: Is it mature enough to do my job for me yet? A: Not quite. The best open-source agent has about a 34.5% success rate within 50 steps, with a high error rate on complex multi-step tasks. Its current positioning is "assistant" not "replacement"-it can handle repetitive, well-defined work, but you need to decompose tasks, allow room for trial and error, and do human review. Use it as a "digital intern," not a "digital substitute."


References

This article is AI-assisted and human-edited. Last updated: 2026-08-11

FAQ

What's the difference between Computer Use and RPA?
RPA relies on recording or scripts to click along a fixed path, with no "understanding" capability-once the flow changes, it breaks. Computer Use uses a visual model to "see" the screen and reason about actions, handling interface changes and non-standard scenarios. For simple, repetitive, fixed-flow tasks, choose RPA; for tasks that need to "understand the screen before acting," choose Computer Use.
Are the three closed-source tools usable in China?
All are restricted or unavailable. Claude Computer Use's API requires a VPN and is restricted, OpenAI Operator/CUA is unavailable in China via ChatGPT Pro, and Google Project Mariner is restricted with Google services. Developers in China should look at the three open-source options first-they run locally without restriction.
How do I choose among the three open-source tools?
Depends on what you need. cua (21,113 stars) is ready to use, cross-OS, and easiest to start with; Agent-S (12,147 stars) is a framework, suited for developers building custom agents; fara (6,140 stars) is a model, suited for teams researching frontier computer-use models. Beginners start with cua, advanced users pick Agent-S, researchers pick fara.
Can I use it for banking or payments?
Not recommended. The ability to operate a computer equals the ability to cause harm-banking, payments, and internal company systems should always be operated manually or at least confirmed by a human before execution. Computer Use is suited for information transfer, form filling, and data cleanup-scenarios where "nothing major can go wrong." For anything involving real money, keep a human in the loop.
Is it mature enough to do my job for me yet?
Not quite. The best open-source agent has about a 34.5% success rate within 50 steps, with a high error rate on complex multi-step tasks. Its current positioning is "assistant" not "replacement"-it can handle repetitive, well-defined work, but you need to decompose tasks, allow room for trial and error, and do human review. Use it as a "digital intern," not a "digital substitute."

Related

Hardcore Reviews

One compromised agent loses everything: a comparison of four credential and permission governance approaches

Credentials went from a config item to an attack surface, yet most teams' defenses are still stuck at "put the agent in a sandbox." This review splits cleanly from our sandbox-isolation comparison: the sandbox governs where code runs; credential governance governs how secrets are used, who approves actions, and whether they can leave. It contrasts four approaches — OpenClaw 2.0, OpenWorker, OpenHuman and traditional secret storage — across six lifecycle stages (store / use / approve / exfiltrate / audit / multi-agent): OpenClaw with masked requests plus an opt-in proxy allowlist; OpenWorker with hard floors, an autonomy ladder, a reviewer model and a circuit breaker, and never self-approving unattended; OpenHuman with Privacy Mode enforced in the Rust core and E2E-encrypted inter-agent comms. Secondhand data (SaaS Sentinel transcription, no primary source located) shows compromise probability 0.24 with one agent rising to 0.86 with seven — risk grows superlinearly with count, under the premise "any agent proposes, execute."

Sep 1, 202611 min read
Hardcore Reviews

Comparing 11 Models by Real Token Cost After the August 31 Repricing: Peak Hours, Cache Hits, and Tokenizer Effects

A model's list price wears at least three more layers. Time of day: DeepSeek moved to peak and off-peak pricing on August 17, charging peak rates on weekdays from 09:00-12:00 and 14:00-18:00, halving them off-peak, and applying off-peak rates all weekend, so the same model costs twice as much at 3pm as at 10pm. Caching: prefix cache hits are billed far below standard input, and the variable sits with your prompt structure rather than the vendor. Tokenization: Sonnet 5 changed tokenizers, so the same input now maps to 1.0x to 1.35x more tokens, and the multiplier floats with content type. This comparison fixes one unit throughout, blended rate equals input plus output divided by two, assuming equal token volumes, as a neutral starting point, then recalculates under three realistic load profiles across 11 models, covering list price, cached input, peak and off-peak, and post-tokenizer position. The finding is not which model is cheapest, it is that no model is cheapest, only cheapest for your particular load: any comparison that ignores input-output ratio, cache hit rate, and content type is comparing list prices, not costs. Chinese model prices come from a page-by-page check of official pricing pages on 2026-08-24, re-confirmed on 08-28; overseas prices from a 2026-08-31 roundup. Conflicts are flagged per line. No live benchmarking was performed.

Aug 31, 20269 min read
Hardcore Reviews

DeepSeek-V4-Pro Takes On Claude and GPT-5.6: A Frontier Coding Model Comparison

A comparison of 4 frontier coding models: DeepSeek-V4-Pro-0813 (1M context / Anthropic compat / ¥6/M / native in China) / Claude Opus 5 (~200K / most mature ecosystem) / GPT-5.6-Cyber (reasoning) / Gemini 3 (long context, multimodal). Two comparison tables (capability/pricing), selection guide, 5 FAQs. Representative comparison, not a personal benchmark; competitors per official sites.

Aug 13, 20269 min read