Desktop AI Agents Compared: Copilot Cowork, Claude and ChatGPT Work Governance on Windows

Farmer's main point holds up: the products work in a similar way, but they differ in where the work runs, what the agent can reach and how much control administrators get. Several details have changed since the products launched, so the vendor documentation adds some useful corrections.

The common working model​

TechTarget describes a pattern most of these products share:

• The user describes the result they want.
• The agent writes a plan, reads files and calls tools.
• It produces documents, spreadsheets or presentations over several minutes or hours.
• Most products can split a task across subagents or run it on a schedule, and they pause to ask the user for approval.

Farmer gives this timeline. Anthropic released Claude Cowork as a research preview in January 2026 and made it commercially available in April. Microsoft unveiled Copilot Cowork in March and released it in June. OpenAI launched ChatGPT Work in July. Most of the market appeared within about seven months.

Summary: these are agents that run multi-step jobs and deliver finished files. For IT, the differences that matter are where the agent runs, what it can reach and what admins can control.

Claude Cowork: a VM on your PC​

TechTarget says Claude Cowork runs on the user's machine. The desktop app starts a sandboxed Linux VM, mounts only the folders the user grants and sends model calls to Anthropic's servers. Inside those folders it can read, edit and create files, including Word, Excel, PowerPoint and PDF output. A beta computer-use feature can control the screen, browser and other applications. Connectors cover Microsoft 365, Google apps, Slack and Jira.

Anthropic's release notes confirm the main architecture. They describe the early preview as running locally in an isolated VM, with direct access to local files and MCP integrations. They also give two dates that matter to Windows shops:

• April 2026: Cowork became generally available on macOS and Windows through Claude Desktop. Enterprise plans gained role-based access controls, so admins can turn Cowork on for specific groups, including groups synced through SCIM.
• March 2026: computer use arrived as a research preview for Pro and Max users. Claude can open files, point, click and navigate the screen.

TechTarget lists some limits:

• Tasks belong to one user and can't be shared while they run.
• A local task stops when the machine sleeps.
• A Cowork task can use 50 to 100 times as many tokens as a chat message.

Anthropic's July 7 release notes partly address the sleep problem. Cowork came to web and mobile, with sessions running remotely in beta, so work continues after you close the laptop and scheduled tasks run with no device online. The rollout started with the Max plan. Local tasks still depend on the PC staying awake, but a cloud option now exists.

The same July notes include a change Microsoft 365 admins should know about. The Claude Microsoft 365 connector gained write tools for drafting and sending email, managing calendars and editing files in OneDrive and SharePoint. A Microsoft Entra administrator has to consent to the updated permissions before anyone can use them.

On safety, TechTarget says security researchers tested Cowork more heavily than its rivals and Anthropic fixed the defects they found. Cowork now asks before each action unless the user switches to an automatic mode, and deleting files needs a separate permission. Anthropic's September 16 announcement merged Cowork and chat. Cowork features are now available from any conversation, rolling out first to Pro and Max, and Enterprise admins are promised at least 30 days' notice before their organizations change.

ChatGPT Work: Codex without the coding screen​

TechTarget says OpenAI built its agent on Codex, which started as an open-source command-line coding agent. By mid-2026 Codex had more than 5 million weekly users, and more than a million of them used it for work other than writing software. ChatGPT Work runs on the Codex architecture with the coding interface removed.

TechTarget describes two modes:

• Cloud: the task runs in an isolated VM with its own browser and continues when your PC is off.
• Local: the task runs on your machine and has full use of it, but it can't be moved to the cloud afterward.

OpenAI's ChatGPT Business release notes add the Windows details. The new ChatGPT desktop app is available for macOS and Windows and combines Chat, Work and Codex. On the desktop, Work can use local files and desktop apps with the user's permission. Codex also supports Computer Use on Windows, where it can see, click and type in Windows applications, and a later release added per-app access controls for Computer Use on Windows.

For Microsoft 365 integration, OpenAI says write actions in its Microsoft apps, such as drafting Outlook email, stay disabled until a workspace admin enables them. Some tenants may also need Entra approval for the updated scopes. Anthropic's connector has the same requirement, so expect Entra consent requests from both vendors.

TechTarget also describes Codex's sandbox modes: read-only, workspace-write and full access. OpenAI's own documentation advises against full access. Network access is off by default, and an optional auto-review step uses a second model to check actions that need approval.

Copilot Cowork: in the tenant, not on the desktop​

This is the product most WindowsForum readers will be asked about. TechTarget notes that Copilot Cowork uses the same agent technology as Claude Cowork, because Microsoft licensed it from Anthropic. The difference is location: Copilot Cowork runs inside the Microsoft 365 tenant, not on the PC. Tasks run in a sandboxed cloud environment, use Outlook, Teams, SharePoint, OneDrive and Excel through Microsoft's Work IQ, and save their output to OneDrive.

Microsoft's GA announcement in June listed the controls admins will care about:

• Cowork is off by default, and admins choose who gets it.
• Spending limits can be set for the tenant, groups or individual users, with custom alerts.
• Activity flows through the audit log, DSPM, eDiscovery, Insider Risk Management and Communication Compliance. At GA, DLP was listed as "coming soon."
• Pay-as-you-go pricing is $0.01 per Copilot Credit. P3 offers a discount for committing to usage in advance.

TechTarget says the cost of a task is unknown until it has run. That is partly true: Microsoft calculates the bill from model use, context retrieval, tool calls and runtime, so there's no fixed price per task. However, Microsoft says users see each task's cost in credits while it runs, and admins can cap spending. Microsoft also claims Copilot Cowork was 30–40% cheaper per prompt than Claude Cowork with its Microsoft 365 connector. That figure comes from Microsoft's own internal testing, so treat it as a vendor claim, not an independent benchmark.

Local browser use: the exception to "cloud only"​

TechTarget says Copilot Cowork can't work with local files or apps, except through local browser use. Microsoft's What's New page shows that feature moving from Frontier to general availability in its August 2026 entries. Microsoft's documentation describes how it works: Cowork can complete browser automation in Microsoft Edge on your device, using the sites you're already signed in to, and it shows its work in a hidden Edge tab so you can stay in the conversation while it acts on your behalf.

The security model is the important part. The agent has exactly the same access to sites that you have when you browse by hand — no more and no less. Credentials, cookies, and session tokens stay on your device. Microsoft adds that because Cowork drives your local Edge, it inherits every site restriction your admin already enforces—web filtering, Conditional Access, browser-management policy, and Microsoft Purview data loss prevention.

Matching the user's access is a guarantee and also a risk. A user with broad SharePoint, Dynamics 365 or browser-session access can ask the agent to work across all of it.

To turn the feature on:

1. In the Microsoft 365 admin center, go to Copilot > Settings > View All > Cowork settings.
2. Under Allow browser access, check the Allow Cowork to use the Microsoft Edge browser to perform tasks on behalf of users option.
3. Choose who gets it. Message Center notice MC1474461 says admins can configure local browser access for: No users, All users, Specific groups. That control rolled out worldwide mid to late September 2026, is off by default and requires admin activation.
4. For control at the endpoint, you can assign policies to user groups to enable browser automation through Edge with Edge 152, released the week of August 24, 2026, and tenant admins can use the Edge policy to control the Edge instances in which Cowork is allowed to use the browser.

If you can't find the setting, Microsoft renamed it. The Cowork settings panel is renamed to Browser use in Microsoft Copilot; the change, rolling out in mid-October 2026, does not affect functionality or configurations.

Users who don't have Edge aren't completely blocked. Cowork still completes every step of a request that doesn't need the browser and tells them which step needs Edge.

The model options also differ from what TechTarget describes. TechTarget says Copilot Cowork ran on Claude models at GA, with OpenAI models available through Frontier, which matches Microsoft's June blog. Microsoft Learn's September entries add Claude Fable 5.1 and GPT 6 Astra to the model selector. Microsoft also says the Fable preview model is off by default, needs an admin to enable it, and requires the model provider to keep prompts and responses. Check that last point with your compliance team before enabling it.

Google, Manus, Perplexity and open source​

The following details come from TechTarget and were not independently checked against vendor documentation:

• Gemini Spark is a tab in the Gemini macOS app for personal Google accounts on AI Pro or Ultra. Workspace and school accounts aren't supported. It's in beta, U.S.-only, and has no published enterprise governance guidance. TechTarget says it isn't a candidate for most organizations, and Windows shops have no client to deploy anyway.
• Manus 2.0, released in September 2026, runs tasks on a cloud computer the user can buy and also has a local desktop agent. Its Cue app gives each personal agent an email address, a phone number and a wallet with a spending limit.
• Perplexity Computer adds desktop tasks to Perplexity's research engine but does little with local files.
• Open-source frameworks such as Open Interpreter, ByteDance's UI-TARS and Simular's Agent S suit organizations with strict data-sovereignty rules, especially when paired with locally hosted models. TechTarget warns that these agents act with the local user's full privileges, so building proper sandboxing takes significant engineering.

How to choose​

TechTarget suggests three questions, and the vendor documentation adds detail to each:

Farmer notes that many organizations end up running two agents: Copilot Cowork for work across the enterprise where oversight matters, and Claude or ChatGPT Work for individual work on local files. That makes sense, but it also means two sets of Entra consents, two billing models and two audit trails to watch.

What to do before users install these themselves​

• Check Entra app consents. Both Anthropic's and OpenAI's Microsoft 365 write features need admin approval. Check what users have already requested.
• Set Cowork spending caps before enabling it. Copilot Credits are consumed with every task.
• Pilot browser use with a small security group, as the MC1474461 guidance recommends, and include Edge policy in the pilot.
• Review oversharing in SharePoint. An agent that inherits user permissions will expose any overly broad access.
• Decide on autonomy settings centrally. The per-action approval prompts on these agents protect you only if users leave them turned on.

Analysis​

Farmer argues that organizations should focus on building their own governance rather than committing long-term to one vendor, and the evidence supports that. In ten months, Claude Cowork went from a macOS preview to Windows GA to a feature built into every conversation. Copilot Cowork added models from two vendors and a local browser feature. OpenAI merged Codex and ChatGPT into one Windows app. Any comparison table will be out of date within a quarter, but sound policies on permissions, spending and auditing will keep working whichever agent wins.

Microsoft's approach of keeping work in the cloud and reaching the device only through a governed Edge tab is the most cautious design for a managed Windows fleet. Claude and ChatGPT Work can do more with local files and apps, but that means the agent clicks around Windows with the user's privileges. Whether that suits your organization depends on how far you trust the approval prompts and how carefully you've limited each user's access.