Agents on the desktop: Claude Cowork and its equivalents

An overview of desktop agents

Use the table below to learn more about the most popular desktop agents for enterprise use:

Let's explore these agents in more detail.

Anthropic: Claude

Anthropic was the first major vendor to market with a personal agent: Claude Cowork. It runs on the user's machine and works directly on local files. To do this, the desktop app starts a sandboxed Linux VM, mounts the folders the user grants access to and sends model calls to Anthropic's servers.

Within those folders, Cowork reads, edits and creates files and produces Word, Excel, PowerPoint and PDF outputs. It can divide a task among subagents and run scheduled tasks. In beta, there is a computer-use feature that can control the screen, browser and other applications. Connectors integrate with Microsoft 365, Google apps, Slack and Jira, and the plugins bundle skills and connectors for particular roles.

There are some limitations of this design. Tasks belong to a single user and can't be shared while running. A local task will also stop if the laptop or desktop goes to sleep. A Cowork task can also use 50 to 100 times as many tokens as a chat message.

As the earliest agent of its kind, security researchers tested the product more rigorously than any competitor's, and Anthropic addressed defects the researchers discovered. Now, the agent asks before each action unless the user switches to an automatic mode. Access is also limited to folders the user selects, and deleting files needs separate, specific permissions.

Recently, Anthropic reconfigured Cowork to merge it with chat functionality. As a result, Cowork's capabilities can now appear in any conversation.

OpenAI

OpenAI's desktop agent grew out of Codex, a tool for programmers that began as an open source command-line coding agent. When it was integrated with the desktop app to manage multiple coding agents simultaneously, usage took off. By mid-2026, Codex had more than 5 million weekly users, over a million of them doing work other than writing software. In July, OpenAI released ChatGPT Work, which runs on the Codex architecture but with the coding interface removed. The vendor also merged the Codex app into a new ChatGPT desktop app that offers chat, Work and Codex on every plan.

ChatGPT Work offers two modes. In the cloud, the task runs in an isolated VM with its own browser and can continue when the user's machine is off. When it's running locally, the agent runs on the user's computer, with full use of the computer while it's running, but the task can't then be moved to the cloud. Users can schedule work to act once, on a timetable or when an event occurs.

The governance model is quite explicit, as it was inherited from a system for developers who run commands. Codex operates in a sandbox that limits what a command can do with three modes: read-only, workspace-write and full access. The documentation advises against the latter. Network access is off by default. For further governance, an optional auto-review step uses a second model to check actions that need approval.

Microsoft: Copilot Cowork

Copilot Cowork uses the same agent technology as Claude but runs inside the Microsoft 365 tenant rather than on the desktop. Microsoft licensed Anthropic's technology, and the product ran on Claude models at general availability, with OpenAI models available through the Frontier program, letting customers choose between models.

The agent runs in a sandboxed cloud environment, not on a local machine, so a task continues when the user closes the laptop and can be picked up on another device, including a mobile device. It is grounded in Outlook, Teams, SharePoint, OneDrive and Excel through what Microsoft calls Work IQ, and it writes its outputs back into OneDrive. Plugins connect it to services like Jira, Salesforce, ServiceNow, SAP and Workday. Event-driven tasks can start a task when a matching email or a Teams mention arrives.

Copilot Cowork can't work on local files or local applications, except for local browser use, which enables the agent to complete web tasks in Microsoft Edge on the user's device using the user's existing sign-ins.

Note that the product is off by default and requires a Microsoft 365 Copilot license. It bills each task separately through Copilot Credits based on model use, tool calls and runtime, so the cost of a task is unknown until after it has run.

Google, Manus and Perplexity

Google's entry, Gemini Spark, is somewhat narrower in focus. It appears as a tab in the Gemini macOS app for users with a personal Google account on the AI Pro or Ultra tier. Workspace and school accounts are unsupported. Spark acts on local folders that the user links in the sidebar, sorts and renames files, and it can build Workspace spreadsheets from local documents on a schedule. For now, it is in beta and limited to the U.S. Google hasn't published enterprise governance guidance yet, so it isn't a candidate for most organizations.

There are other products worth knowing about that have had less market impact so far, including the following:

• Manus released Manus 2.0 in September 2026. It runs tasks on a cloud computer the user can buy, has a desktop agent for local use and can start tasks on events such as a new email. Its Cue app gives each personal agent an email address, phone number and even a wallet with a spending limit.
• Perplexity Computer adds desktop tasks to Perplexity's research engine but does little with local files.
• Open source computer-use agent frameworks are a useful option for organizations with strict on-premises data governance, for reasons of security or data sovereignty. Examples include Open Interpreter, ByteDance's UI-TARS and Simular's Agent S. For even stricter governance, users often pair these agents with locally hosted models rather than external APIs. Because these agents execute actions with the local user's full privileges, deployment requires significant engineering to implement comprehensive sandboxing that prevents unauthorized execution or data loss.

Selecting a desktop agent

Three questions are critical when choosing computer-use agents:

1. Where does the work run? Do you need to access local files, or files in Microsoft 365 or Google Workspace? For many enterprises, their work is already in Outlook, Teams and SharePoint, so Microsoft Copilot Cowork reaches it with the organization's permissions applied. But individual users are also looking for access to their local files to boost their personal work and productivity.
2. How much autonomy is the organization prepared to grant? In general, a computer-use agent should ask before acting. Some agents let users switch that off, while other features, such as OpenAI's auto-review and Microsoft's tenant controls, set defaults at the enterprise level without relying on users' judgment.
3. How will it be paid for? Claude is a flat subscription with a shared usage allowance. Microsoft and OpenAI both meter cost per task.

Many organizations run two types of computer-use agents. For example, Microsoft Copilot Cowork suits coordination across the enterprise, where oversight matters, but Claude or ChatGPT Work suits individual work on local files, where reach matters more. Other agents might provide users with cost or flexibility benefits or prevent them from experiencing vendor lock-in.

A practical approach to deployment is a phased, risk-managed rollout. Run a controlled pilot, tie agent autonomy to ROI and governance and keep vendor decisions flexible. The strategic priority should be building organizational governance, rather than committing long-term to a specific vendor platform.

Donald Farmer is a data strategist with 30+ years of experience, including as a product team leader at Microsoft and Qlik. He advises global clients on data, analytics, AI and innovation strategy, with expertise spanning from tech giants to startups. He lives in an experimental woodland home near Seattle.