Computer-Use AI Agents: The Best Open-Source & Closed-Source Tools in 2026

vComputer-use AI agents are AI systems that can operate a computer in much the same way as a human ‒ by interacting with graphical user interfaces (GUIs), rather than relying only on APIs.

They can read what is on the screen, click buttons, type text, move the cursor, navigate websites, open desktop applications, fill out forms, navigate graphical user interfaces and perform other computer interactions. All of this allows them to execute multi-step workflows across different apps.

Unlike traditional AI agents, which typically call APIs or use structured tools, computer-use agents can work directly with websites, desktop applications, operating systems, and mobile devices. Agents of that kind are usually assembled with frameworks and SDKs for building AI agents, which handle tool calling, memory and orchestration instead of screen control.

They perceive the current state through screenshots, visual inputs, or accessibility data, then plan the next action and repeat this process until the task is complete. Our AI agents vocabulary defines this perception–planning–action loop. It makes cross-application automation possible even when dedicated integrations are unavailable.

Computer-Use AI Agents: Features and Platforms

Prices are in USD per month unless noted. Usage limits apply; free software may still require paid model inference or hosting.

Where it runs means the controlled computer or browser, not necessarily where the model performs inference. A local app can still send screenshots to a cloud model. Desktop/mobile clients can also be remote controls for a cloud agent.

Open-source Computer-use AI Agents

License and release status

Licenses and release status, checked November 8, 2026. Apps, SDKs, and models may have separate versions and terms.

License note: Skyvern’s AGPL-3.0 license can require source disclosure when modified software is offered over a network. Hosted services, model weights, and third-party dependencies may have separate terms, so teams should review the complete dependency chain before production use.

1. UI-TARS

UI-TARS is ByteDance’s family of multimodal computer-use models. It can understand what’s happening on the screen and operate graphical interfaces using mouse and keyboard actions. UI-TARS Desktop turns a configured model into an app for controlling your local Windows or macOS computer and browser. Model inference may run locally or through a hosted endpoint; the former free Remote Operator service has been discontinued. Agent TARS is the broader agent stack. It brings GUI interaction and vision to terminals, computers, browsers, and other products through a CLI and Web UI, with support for real-world tools via MCP.

• UI-TARS GitHub

• UI-TARS desktop GitHub

• UI-TARS Paper

• UI-TARS 2 Paper

2. Agent S3

Agent S3 is an open-source framework that lets AI operate computers running Linux, macOS, or Windows. It uses multimodal models, visual grounding, reflection, and past experience to work through complex, multi-step tasks across desktop apps. It can also run Python and Bash code when a task is easier to handle programmatically.

• Agent S3 GitHub

• Agent S3 Paper

• Agent S3 Blog post

3. Browser Use

Browser Use helps AI agents use a browser much like a person would opening pages, clicking buttons, typing, and filling out forms. Developers can connect it to different language models and build agents for data extraction, form filling, website testing, and longer web workflows. You can run it locally and customize it yourself, or use Browser Use’s hosted cloud agent for more complex tasks.

• Browser Use GitHub

• Browser Use Website

• Browser Use Docs

4. CUA

CUA provides open-source desktop drivers, sandboxes, and evaluation tools for computer-use agents. Cua Driver controls macOS, Windows, and Linux applications; the current SDK supplies isolated desktops that your own agent drives.

Cua Spaces runs desktops on computers you own or VMs in your own cloud account, rather than renting managed Cua computers. The core tools are MIT-licensed; Spaces and Driver Bench are source-available under FSL-1.1-MIT. Optional components have separate terms. Spaces has free and paid plans, and the older cua-agent framework is deprecated. Components and licenses; migration details; deployment and pricing.

• CUA GitHub

• CUA Website

• CUA Docs

5. UFO³

UFO³ is Microsoft’s framework for coordinating GUI agents across multiple devices and platforms, including Windows, Linux, and Android. It breaks complex workflows into dynamic task graphs, assigns each task to the right device, and can run independent steps in parallel while adapting the plan based on results.

UFO³ builds on UFO², Microsoft’s Windows-focused desktop agent, which controls applications through a mix of GUI actions and native Windows APIs. So UFO² handles automation on a single Windows device, while UFO³ connects those device agents into larger cross-device workflows.

• UFO³ GitHub

• UFO³ Docs

6. Stagehand

Stagehand is a framework for automating browsers with both natural language and code. Developers can use AI to navigate unfamiliar pages, switch to precise code for predictable steps, run multi-step agents, and extract structured data. Stagehand can preview and cache AI-generated actions, replay them without calling a model, and bring AI back in if a website changes and the automation breaks.

It’s mainly a developer framework for building reliable browser automations, rather than a finished consumer agent.

• Stagehand GitHub

• Stagehand Website

• Stagehand Docs

7. Skyvern

Skyvern is a browser automation platform that adds AI capabilities to Playwright. It uses vision models and multiple agents to understand websites, plan actions, and complete workflows such as filling out forms, downloading files, and extracting data even on unfamiliar sites or when layouts change.

Developers can use its SDK, while non-technical users can build automations through a no-code interface. Skyvern also offers a managed cloud version with proxies, CAPTCHA solving, and parallel browser sessions.

• Skyvern GitHub

• Skyvern Website

• Skyvern Docs

8. OpenAdapt

OpenAdapt is a system for automating important tasks that can only be done through a user interface. You demonstrate the task once, and OpenAdapt turns it into an inspectable workflow that can run across browsers, Windows, macOS, Linux, RDP, and Citrix/VDI. Its main focus is dependable, verifiable execution, not letting an agent freely improvise. Healthy runs are deterministic and use no generative-model calls. OpenAdapt checks whether important actions produced the intended result and pauses for human review when it’s uncertain.

• OpenAdapt GitHub

• OpenAdapt Website

• OpenAdapt Docs

9. Agent-E

Agent-E is an agent system for automating tasks on a computer, currently focused on browser workflows. It is built on the AG2 framework and lets users control a browser with natural-language instructions, searching websites, filling out web forms, comparing products, finding information, and managing web apps such as Jira.

It uses separate agents for planning and browser navigation and can run through either a chat-style browser interface or an API.

• Agent-E GitHub

• Agent-E Website

10. AgentCPM-GUI

AgentCPM-GUI is a GUI-agent model built for Android automation; inference can run on your own machine or server. The 8-billion-parameter model reads smartphone screenshots, follows natural-language instructions, and works through multi-step tasks by tapping, typing, swiping, and using navigation controls.

It can operate both Chinese- and English-language apps, with specialized training on more than 30 popular Chinese apps. The reinforcement-tuned reasoning helps it plan before choosing each action.

• AgentCPM-GUI GitHub

• AgentCPM-GUI Models

11. nanoMuse

nanoMuse is an open-source personal agent that connects tasks across your own devices. With its Hands feature enabled, it can work through Android apps and desktop windows on Windows, macOS, and Linux, alongside browsers, files, and command-line tools. Its iPhone and iPad apps can send tasks to other devices but cannot control the iOS screen. It offers a limited free model allowance, and after that, you supply your own model key or supported paid subscription.

• nanoMuse GitHub

• nanoMuse Model Setup

12. OpenGCUA (research preview)

OpenGCUA is Alibaba’s open-source toolkit for agents that combine screen control with code execution. The agent can choose between screenshots, accessibility data, shell commands, and app APIs, rather than clicking through every step. Its October 2026 research preview includes a macOS agent and a cross-device SDK for developers. Running the macOS agent requires macOS 14 or later, Python, and a model API key. The Apache-2.0 software is free, but model inference may cost money.

• OpenGCUA GitHub

• OpenGCUA Paper

Proprietary Computer-use AI Agents

1. ChatGPT Work

ChatGPT Work is OpenAI’s general-purpose work agent for tasks involving files, connected services, browsers, and desktop applications. Web and mobile work runs in the cloud; the desktop app can also use local files and applications with permission. Check the execution mode and available tools rather than assuming the device you chat from is the device doing the work.

OpenAI has also announced that Work capabilities will move into the main ChatGPT experience, with shared ChatGPT/Codex token usage for Plus, Pro, and Business starting in December 2026.

Plus, one more note, ChatGPT Atlas is no longer an active option: OpenAI retired the standalone browser in August 2026 as browser-agent capabilities moved into ChatGPT and Codex.

• ChatGPT Work Website

• ChatGPT Work and Codex environments

2. Claude (chat and Cowork)

Claude now brings chat and the former Cowork experience into one conversation, following a rollout that began on September 16, 2026. It can answer questions or take on longer tasks involving files, websites, and productivity tools. Cowork tasks run on Anthropic’s servers, and authorized access to local files, browsers, or applications still depends on Claude Desktop being open. Desktop computer control has its own platform and plan restrictions.

• Claude Cowork Website

• Claude Cowork execution and availability details.

3. Gemini in Chrome with Auto Browse

Gemini in Chrome with auto browse can carry out multi-step web tasks in your local Chrome session or a separate remote cloud browser. Basic assistance and autonomous browsing have different access rules: auto browse requires an eligible Google AI Pro or Ultra plan, age 18+, and English. Don’t forget to check the instructions for your device rather than assuming one global rollout.

• Gemini in Chrome Website

4. Perplexity Personal Computer

Perplexity Personal Computer connects cloud-based Perplexity Computer to your own Windows or macOS machine. Both desktop apps can work with local files, but installed-app control is currently available on macOS, not Windows. You can start or monitor tasks from your phone; local work still depends on the computer being online. Supported Macs can use hybrid cloud/local inference; Portable Computer can run tasks locally on supported AMD Windows or NVIDIA Windows/Linux hardware.

• Perplexity Personal Computer Website

• Perplexity Personal Computer Docs

5. Genspark Claw

Genspark Claw can run on a dedicated cloud computer or locally through its macOS and Windows desktop app. It can operate browsers, applications, and files, and combines computer use with skills for research, documents, spreadsheets, presentations, media, and scheduled workflows. Local mode can have broad file-system access. A selected working folder is not a security boundary, so restrict system permissions and back up important files. The local app does not require an additional Cloud Computer subscription, but it still consumes account credits.

• Genspark Claw Website

• Genspark Claw Docs

6. Amazon Nova Act

Amazon Nova Act is an AWS service and Python SDK for building, deploying, and managing browser agents at scale. Developers can combine natural-language instructions with Python code to create more predictable workflows for search, extraction, form filling, and transactions. The hosted service is proprietary; the SDK repository is available under Apache 2.0.

Nova Act can run multiple sessions in parallel, preserve browser sessions, and hand a task to a human when approval or direct intervention is needed. Workflows can be developed locally, deployed to AWS, and monitored through the AWS Console.

• Amazon Nova Act GitHub

• Amazon Nova Act Website

7. Manus Browser Operator

Manus Browser Operator is a browser extension that turns your existing browser into an active workspace for Manus. It can work inside websites where you’re already logged in, use the premium tools you already have access to, and complete complex tasks through your real accounts and sessions.

• Manus Browser Operator Docs

8. Grok Bot

Grok Bot is the beta work-agent product from SpaceXAI and Cursor, launched on August 11, 2026. Its agents use a persistent cloud computer with a browser, files, and terminal, so work can continue when your laptop is closed. Desktop and mobile apps are control surfaces. Bots can share context and coordinate, with approval controls for consequential actions.

The October 27 beta update added shopping cards, Shop Pay checkout, order tracking, and subscription-management templates. These expand the workflows the cloud agent can handle, but purchases still require the relevant approval.

• Grok Bot Introduction

• Grok Bot Docs

• Grok Bot Changelog

Best Computer-Use Models and APIs

These tools help developers build agents that act on screenshots. Your application must execute actions, capture the next screen, and enforce permissions. Cloud APIs handle model inference; downloadable models need your own inference setup.

OpenAI computer use

OpenAI’s Responses API lets models such as GPT-6.1 Sol turn screenshots into mouse and keyboard actions. Developers connect those actions to a browser or desktop to automate forms, navigation, and multi-step tasks. It fits applications that need computer control alongside other API tools. Model usage and the execution environment have separate costs.

• OpenAI computer use: Docs

• OpenAI GPT-6.1 Model announcement

Anthropic computer use

Anthropic’s computer-use tool lets Claude inspect screenshots, click, type, scroll, and zoom through desktop workflows. The model runs in the cloud; your application executes its actions in an environment you control. Anthropic provides a Docker-based reference implementation for getting started. API usage is billed at the selected model’s token rates.

• Anthropic computer use Docs

• Computer use, the Skills API, and the Files API Announcement

Gemini Computer Use

Gemini Computer Use turns screenshots into actions for browser, desktop, and mobile interfaces. Google recommends Gemini 3.8 Flash, while developers provide the execution layer and approval controls. It is useful for building GUI automation into an application. Computer Use requires paid API usage, even when the model’s other features have a free tier.

• Gemini Computer Use Docs

• Gemini 3.8 Flash Model

UI-TARS Desktop

UI-TARS Desktop connects a vision-language model to Windows or macOS so it can operate apps and browsers. It can use ByteDance’s UI-TARS-1.5-7B, an Apache-2.0 model that interprets screenshots and predicts GUI actions. The Desktop app is the execution layer, not the model itself. Both are free to download, and inference can run locally or through a paid hosted endpoint.

• UI-TARS Desktop Docs

• UI-TARS Desktop Model

AgentCPM-GUI

AgentCPM-GUI is an eight-billion-parameter model for Android automation. It reads phone screenshots and generates taps, swipes, typing, and navigation actions for Chinese- and English-language apps. It suits mobile-agent research and custom workflows, with developers supplying inference and device control. The Apache-2.0 weights are free to download; running them still requires compute.

• AgentCPM-GUI Docs

• AgentCPM-GUI Model

Computer-Use AI Agents vs Browser Agents

Browser agents operate only inside web browsers, where they can navigate websites, click buttons, fill out forms, extract information, and complete online workflows. Computer-use AI agents can interact with entire operating systems, including desktop applications, file systems, terminals, browsers, and, in some cases, mobile devices. This allows them to complete workflows that span multiple applications, for example: collecting data from a website, editing a spreadsheet, saving files locally, and sending the final report through a messaging or productivity app.

In other words, most browser agents are a subset of computer-use agents, but computer-use agents are not limited to the web.

How to Choose a Computer-Use AI Agent

Start with the environment and the amount of control the workflow needs—not the leaderboard score. A strong benchmark result does not guarantee reliable operation on your websites, permissions, or desktop setup — our LLM benchmarks guide covers what agent and coding evaluations actually measure, and where contamination and saturation make them unreliable.

• For browser automation in code: start with Stagehand when you want to mix deterministic Playwright steps with AI, or Browser Use when you want a broader model-driven browser-agent framework.

• For managed or no-code web workflows: consider Skyvern or Amazon Nova Act, then compare deployment, human approval, observability, and pricing requirements.

• For desktop and cross-platform agent development: compare CUA, Agent S3, and UI-TARS. UFO³ is the more specialized option for Windows-centered or multi-device orchestration. For self-hosted setups: if you plan to run the model and the agent loop on your own hardware, you will need the surrounding stack too — a local model runner, a vector store for memory, a Docker sandbox and a task scheduler. We collected the GitHub repos to build your own local agent.

• For repeatable, controlled UI workflows: OpenAdapt is designed around inspectable demonstrations, verification, and deterministic healthy runs rather than free-form improvisation.

• For Android research: AgentCPM-GUI is the specialized mobile option in this list.

• For end-user knowledge work: ChatGPT Work, Claude, Perplexity Personal Computer, Genspark Claw, Manus, and Grok Bot trade developer control for integrated apps, files, browsers, and approval flows.

Before committing, test the agent on your own workflow using completion rate, recovery after failure, latency, cost, permission boundaries, and the quality of the audit trail. For self-hosted software, include the license and the maintenance signal in the decision.

When Computer-Use AI Agents Don’t Work Well

GUI control is valuable when an API is unavailable, but it is usually less precise and more fragile than direct integration. Common failure modes include:

• Long workflows: small navigation or interpretation errors compound across dozens of steps.

• Dynamic or visually ambiguous interfaces: canvas-based controls, virtual desktops, pop-ups, responsive layouts, and changing labels can confuse grounding.

• Authentication and anti-bot checks: CAPTCHAs, expired sessions, multifactor authentication, and site policies can stop unattended execution.

• High-stakes actions: payments, deletion, account changes, legal submissions, and sensitive communications require explicit approval and a recoverable audit trail.

• Strictly repeatable processes: if an API, database operation, or deterministic RPA flow exists, it is often faster, cheaper, and easier to test than visual control.

• Untrusted content: webpages and documents can contain prompt-injection instructions. Limit credentials, isolate the environment, and require approval before consequential actions.

The practical pattern is hybrid: use APIs and deterministic code for stable steps, computer use for gaps in the workflow, and human approval for irreversible decisions.

FAQ

What is a computer-use AI agent?

A computer-use AI agent is an AI system that interacts with graphical user interfaces (GUIs) the same way a human does. Instead of relying only on APIs, it can read the screen, click buttons, type text, move the cursor, navigate websites, and operate desktop or mobile applications to complete multi-step tasks.

What is the difference between computer-use agents and traditional AI agents?

Traditional AI agents primarily interact with software through APIs and structured tools. Computer-use agents work directly with graphical interfaces using screenshots, accessibility information, or visual inputs, allowing them to automate applications even when no API is available.

What are the best open-source computer-use AI agents?

There is no single best open-source computer-use agent. Stagehand and Browser Use are strong browser-first starting points; CUA, Agent S3, and UI-TARS target broader computer control; UFO³ focuses on multi-device orchestration; OpenAdapt emphasizes repeatable, inspectable workflows; and AgentCPM-GUI targets Android. The right choice depends on environment, reliability, license, and maintenance requirements.

What can computer-use AI agents do?

Computer-use agents can browse websites, fill out forms, operate desktop software, manage files, automate repetitive workflows, extract information, navigate operating systems, perform browser testing, and execute complex tasks across multiple applications with minimal human intervention.

Are computer-use AI agents the same as browser agents?

No. Browser agents specialize in web-based tasks such as browsing, form filling, and web automation. Computer-use agents are broader—they can control browsers as well as desktop applications, operating systems, file managers, productivity software, and, in some cases, mobile devices.

Is there an AI agent that can use my computer?

Yes. UI-TARS Desktop and frameworks such as Agent S3 can control a local desktop, while tools such as Manus Browser Operator work inside your existing browser. Some commercial assistants also offer local access. Check the supported OS and permissions: controlling your computer does not necessarily mean the model runs locally or that screenshots stay on your device.

Are there any open-source models for computer use?

Yes. Downloadable options include UI-TARS-1.5-7B and AgentCPM-GUI, whose official model cards list Apache-2.0. UI-TARS Desktop is a separate application, not the model itself. These releases allow self-hosting, but you still need inference hardware, a device-control layer, and appropriate safeguards. Free weights do not mean free compute.

More tool comparisons from Turing Post

• Best AI Image Generators: 11 Text-to-Image Models Compared

• 13 Frameworks and SDKs for Building AI Agents

• 13 Open-Source Tools for Foundation Model Deployment