What are Cowork apps? AI coworkers that use your computer

A cowork app has three traits that set it apart from a chat assistant:

1. It takes an outcome, not a prompt. "Turn these receipts into an expense report" rather than "summarise this PDF".
2. It acts. It opens files, drives a browser, calls connected apps and writes new documents, instead of handing you text to paste.
3. It runs long. Tasks take minutes or hours, often in the background, and can repeat on a schedule.

Anthropic's own description is a good benchmark: Cowork "uses the same agentic architecture that powers Claude Code, with no terminal required". That's the key to the category. Coding agents proved that a model in a loop, with tools and a sandbox, can finish real work. Cowork apps point that loop at spreadsheets, inboxes and slide decks.

They sit between two neighbours. Coding agents such as Claude Code and Codex do the same thing for software. Bots, as we define them in Bots vs agents, go a step further: a persistent agent with its own computer, memory, schedule and identity. A cowork app with scheduled cloud tasks is already halfway to a bot.

The category may not stay a separate category for long. On 16 September 2026, Anthropic announced that Claude Cowork and chat are now one Claude, and ChatGPT now has Chat and Work modes side by side. The likely end state is an assistant that decides for itself whether you need an answer or a coworker.

Capabilities vary by product, so here's what each major one documents, grouped by the job.

Work with your files. On desktop, Claude reads and writes files only in folders you connect, and produces Excel spreadsheets with working formulas, decks and formatted documents (Anthropic). ChatGPT Work uses local files in the desktop app, with plugins for Google Drive and SharePoint (OpenAI). Kimi Work mounts local folders, and Manus's My Computer runs terminal commands in folders you authorise.

Drive a browser. Cowork gained a built-in browser on 26 August 2026 that runs apart from your own browser and logins. ChatGPT Work's desktop app has a built-in browser too. Kimi Work uses a browser extension for multi-step web tasks, and Copilot Cowork can browse through a local Edge browser for customers in Microsoft's Frontier programme.

Use your other tools. Cowork connects to MCP connectors, and its plugins bundle skills, connectors and sub-agents into a single install. ChatGPT Work's plugins reach Slack, Teams, email, calendars and CRMs. Copilot Cowork is grounded in Microsoft 365 through Work IQ. Gemini Spark connects natively to Gmail, Calendar, Drive and Docs, with each connection off until you turn it on. Perplexity says its Mac app coordinates more than 400 connectors.

Run on a schedule. Cowork's scheduled tasks run remotely hourly, daily, weekly or on weekdays, even while your computer is asleep. ChatGPT's Scheduled Tasks can run once, repeat, trigger on an event or monitor for changes. Kimi Work has a built-in cron engine with a keep-awake option, Gemini Spark offers time-based and conditional schedules, and Manus 2.0 automations can start when a new email, calendar event, Slack message or Notion update arrives.

Split work across sub-agents. Cowork breaks complex work into subtasks and runs parallel workstreams. Moonshot says Kimi's Agent Swarm coordinates up to 300 sub-agents. Perplexity Computer creates sub-agents for research, documents and data, and lets you run dozens of Computers in parallel. TraeWork can dispatch tasks concurrently to its cloud and to several of your computers.

Operate your computer. In the desktop app, ChatGPT Work can operate your computer in the background, clicking, typing and moving files. Computer use is still a preview feature in Cowork. Manus's My Computer can launch and control local apps.

This table covers the products we could verify on official pages. Access details were checked on 8 October 2026, and we've left out anything we couldn't confirm. Prices are in US dollars.

A few notes behind the table:

• Anthropic announced Cowork on 12 January 2026 for macOS, running its work in a virtual machine using Apple's Virtualization Framework (InfoQ). It began rolling out to web and mobile in beta in July, and on 6 October 2026, "new Cowork tasks on Pro and Max plans run in the cloud".
• OpenAI launched ChatGPT Work on 9 July 2026 with a merged desktop app that puts Chat, Work and Codex in one place, and says Work can "stay with a project for hours if needed". For teams, workspace agents are shared, Codex-powered agents that run in the cloud on a schedule or in Slack.
• Microsoft built Copilot Cowork with Anthropic, integrating "the technology behind Claude Cowork" into Microsoft 365. At general availability on 16 June 2026 it ran on Anthropic models, with model choice for customers.
• Moonshot AI ships several agent products: OK Computer (September 2025), Agent Swarm, Kimi Work on the desktop and Kimi Claw, a browser-hosted version of the open-source OpenClaw agent that entered public beta in February 2026.
• ByteDance says TraeWork "builds upon the original TRAE IDE SOLO mode", and that when a connected computer goes offline, tasks switch to the cloud automatically.
• SpaceXAI's Grok Bot is the most bot-like entry. Announced in beta in August, it's now generally available for Enterprise and for teams on Cursor and Grok plans. SpaceXAI says your bots share one persistent cloud computer, including "files, browser, logins", and that "isolation is per user, not per Grok Bot". That makes handoffs easy, and it's worth knowing when you decide what a bot may sign into.
• Google doesn't use the word "cowork", but Gemini Spark is the same idea pushed further: Google says it works 24/7 even when your devices are off, and on 29 July 2026 Google announced it was rolling out in Australia to Google AI Pro subscribers over the following weeks.

The progress is real, but so are the limits. Before you plan a rollout, know these:

• Much of it is still beta. Computer use is still a preview feature in Cowork, Cowork on web and mobile is in beta for Team plans, ChatGPT's Sites is in public beta and Manus's Cue is in early access. Expect behaviour and limits to change.
• Local access needs the desktop app running. Anthropic's cloud sessions reach your folders only while the Claude desktop app is open, and ChatGPT's synced tasks fall back to a cloud container without local files if your computer is offline.
• Long tasks cost more. Anthropic says multi-step tasks use more of your usage than a quick question, and OpenAI says more complex Work tasks may use more of your plan's included usage.
• Collaboration is thin. Cowork sessions can't be shared, though individual outputs can. Most of these tools are still built for one person delegating, not a team sharing an agent.
• Enterprise controls are catching up. At general availability in June, Microsoft listed Data Loss Prevention for Copilot Cowork as coming soon.
• Security is never finished. Anthropic says the chance of a prompt injection attack is "still non-zero" even with its safeguards.

None of these is a deal-breaker. They're the reasons to start small.

The better products take isolation seriously. Anthropic's architecture overview says each cloud session gets its own sandbox that's discarded when the session ends, cannot reach private or internal addresses, sends all traffic "through a mandatory proxy" that only reaches allow-listed destinations, and keeps connector tokens out of the sandbox entirely. Microsoft's Copilot Cowork is off by default, has spend limits at tenant, group and user level, inherits sensitivity labels and works with the audit log and eDiscovery. OpenAI's Agent Security settings let admins set requirements members can't override. Gemini Spark is "designed to check with you before taking major actions", Kimi Work asks before modifying files or running code locally, and Manus asks before sensitive terminal commands.

The risks are just as real, and the vendors are candid about them. Anthropic's Cowork safety guide says prompt injection becomes a risk when Claude can both read information from outside your trusted boundary and take actions that could compromise you, that "computer use has no sandbox between Claude and what's on your screen", and that network egress settings don't cover its web fetch, web search or MCP tools. OpenAI notes that local work sync doesn't support Zero Data Retention. None of that is a reason to avoid these apps. It's a reason to roll them out on purpose.

Our governance checklist for a business rollout:

1. Name an owner. Australia's Guidance for AI Adoption starts with "Decide who is accountable", and ends with "Maintain human control".
2. Start off by default. Enable the app for a pilot group, not the whole organisation.
3. Scope the access. A dedicated working folder, only the connectors the task needs, and no banking or healthcare portals in reach.
4. Choose the approval mode deliberately. Manual approval for sensitive files, unfamiliar tools and anything hard to undo.
5. Know which browser it uses. A separate built-in browser is a smaller blast radius than an extension acting inside your signed-in profile.
6. Log it. Use what the platform offers, such as Anthropic's Compliance API and OpenTelemetry export, or the Microsoft 365 audit log, and send it to your SIEM.
7. Cap the spend. Long tasks and automatic-approval modes consume usage quickly.
8. Vet plugins and skills. Treat them like software dependencies. Anthropic's Enterprise plan can scan them on install.
9. Check the data path. Know where tasks run, what's retained and whether that meets your privacy and contractual obligations.

Demos are designed to impress. Evaluate with your own work instead:

1. Pick three real, recurring tasks: one document-heavy, one research task and one data clean-up.
2. Shortlist two apps that fit your stack. If your organisation lives in Microsoft 365 or Google Workspace, start there.
3. Run a two-week pilot with a handful of people and a dedicated working folder.
4. Measure four things: finished-output quality, the number of times a person had to step in, time saved and cost per task.
5. Score governance alongside quality: where it runs, its permission model, admin controls, audit trail, connectors and model choice.
6. Decide what stays manual. Some tasks should stay with people. That's a finding, not a failure.

The app that fits your systems and your governance usually beats the one with the best launch video.

We build our own agent tooling first and run it under our own products, so we've learnt the hard parts first-hand. Our open-source Ultrathink plans work and asks clarifying questions before an agent starts. Agent Substrate gives every agent shared memory, its own sandbox and an auditable ledger. Agent Swarm runs the software lifecycle behind fail-closed gates with human escalation, and Omega gives one developer a configurable programming desk. KanbanOS brings them together on a board humans can operate.

If you want agents wired into how your team ships, with human approval where it counts, that's what Ship a Product is for. We'll help you choose the first workflows, design the gates and get something real into production. Book a meeting, and bring the task you'd hand to an AI coworker first. For more on where this is heading, read Bots vs agents or browse the blog.

• Anthropic: Claude Cowork, Get started with Claude Cowork, Cowork on web, desktop and mobile, architecture overview, Use Claude Cowork safely, scheduled tasks, Cowork on web and mobile, Cowork and chat are now one Claude and pricing
• OpenAI: ChatGPT Work announcement, ChatGPT Work and Codex, Agent Security and local work sync and workspace agents
• Microsoft: Copilot Cowork announcement and general availability
• Google: Gemini Spark and Gemini Spark in Australia
• Moonshot AI: Kimi Work and Kimi Agent overview
• ByteDance: What is TraeWork?
• Manus: Manus 2.0, My Computer and My Computer permissions
• Perplexity: Introducing Perplexity Computer and Personal Computer on Mac
• SpaceXAI (formerly xAI): Grok Bot and Introducing Grok Bot
• Coverage: Claude Cowork launch, InfoQ; Kimi Claw launch, Dataconomy
• Governance: Guidance for AI Adoption: Foundations, National AI Centre