# The open-source question behind Copilot Cowork vs Claude Cowork

> Copilot Cowork and Claude Cowork both take a goal, break it into steps and keep working while you are away. What separates them is where the agent's memory, permissions and configuration live. Kortix is the leading open-source alternative to Claude Cowork and ChatGPT Work, and that layer is a git repo you own.

URL: https://opensourcecopilotalternative.com/copilot-cowork-vs-claude-cowork

## Copilot Cowork vs Claude Cowork at a glance

Where each agent runs, what it can reach, and who holds the configuration (checked October 2026).

| | Kortix (recommended) | Copilot Cowork | Claude Cowork |
| --- | --- | --- | --- |
| What it is | An AI Operating System that runs agents on cloud computers | Microsoft's agent for long, multi-step work inside Microsoft 365 | Anthropic's agent for non-coding knowledge work across files and tools |
| Where the work runs | Kortix Cloud, your VPC, or your on-prem network | Microsoft's cloud, inside Microsoft 365 apps and the browser | Anthropic's servers, or your own cloud provider account |
| Memory and configuration | Files in a git repo you own | Your Microsoft 365 tenant; skills in OneDrive | Your Claude account; folders and tools you choose |
| Who can modify it | Anyone with repo access; session work reaches main by change request | Microsoft 365 admins set plugins, permissions and billing | Admins set team access; people pick folders and tools |
| Apps it reaches | 3,000+ apps, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP | Microsoft 365 apps and App Store plugins | Connected tools, folders and a built-in browser |
| How it is billed | Self-host free; Cloud in credits for the agent computer | A Copilot subscription per person plus Copilot Credits | Included in paid Claude plans, with usage limits |
| Open source | Yes | No | No |

## Two products, one bet

Copilot Cowork and Claude Cowork do the same job: you describe an outcome, the agent plans the steps, and it keeps working after you leave. Microsoft says it built Copilot Cowork by bringing "the technology platform that powers Claude Cowork into Microsoft 365 Copilot", so the resemblance is not a coincidence. Microsoft wrapped the pattern in its own tenant; Anthropic sells the original ([Microsoft's launch post](https://www.microsoft.com/en-us/copilot/blog/2026/03/30/copilot-cowork-now-available-in-frontier), checked October 2026).

Copilot Cowork is Microsoft's agent for long-running work inside Microsoft 365. Microsoft's documentation lists the jobs: send and sort email, schedule meetings, build Word, Excel, PowerPoint and PDF files, post in Teams, work in OneDrive and SharePoint, run prompts on a schedule. You reach it at copilot.cloud.microsoft, in Outlook and Teams, and in the desktop and mobile apps.

Claude Cowork is Anthropic's agent for knowledge work outside code. It works in folders and connected tools, opens a browser when needed, splits a large job into parallel pieces, and schedules recurring tasks. It runs on web, desktop and mobile, and Anthropic now calls it one product with chat: "Claude Cowork is now just Claude", with tasks starting from the same message box as a conversation ([Anthropic](https://claude.com/product/cowork) and the [Claude Help Center](https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork), checked October 2026).

## Where the work runs, and what it can reach

Copilot Cowork runs inside Microsoft's cloud and your Microsoft 365 tenant, and it inherits the permissions of the person using it. [Microsoft's FAQ](https://learn.microsoft.com/en-us/microsoft-365/copilot/cowork/cowork-faq) puts it directly: Cowork "can access only the files and emails you can already access", and a file's sensitivity label shows up in the session. Its reach is the Microsoft 365 estate plus App Store plugins an admin deploys or restricts.

Claude Cowork runs on Anthropic's servers in the cloud, and sessions and files are saved to the Claude account. On desktop it reads and writes local files through the Claude Desktop app, and a cloud session reaches your computer through that app when a task needs a local file or your browser. Anthropic also offers Claude through Amazon Bedrock, Google Cloud or Microsoft Foundry, with access by team, spend limits and tool permissions set by admins.

[Microsoft's compete page](https://www.microsoft.com/en-us/copilot/resources/compete-copilot-vs-claude-enterprise) shows how fast this field moves. It says Copilot Cowork "runs in a protected cloud environment" while Claude Cowork "operates on files stored locally" and is "currently available only via desktop app". That is out of date: [Anthropic's help centre](https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork) documents cloud sessions on web, desktop and mobile. The page also still lists Copilot Cowork as a Frontier preview, while [Microsoft's documentation](https://learn.microsoft.com/en-us/microsoft-365/copilot/cowork) says work and school accounts have it generally available.

## Who can change what the agent did and may do

Both products show their steps, pause before consequential actions and keep a record. The governance question is who owns the switchboard.

For Copilot Cowork, the switchboard is the Microsoft 365 admin center. Admins enable Cowork and its billing, deploy or block plugins, set who may use which agent, and inherit the Purview controls already applied to email and files. People can add instructions and up to 50 custom skills, stored in a `Cowork/skills` folder in OneDrive. A plugin an admin deploys cannot be removed by the person using it. All of it lives on Microsoft's side of the line.

For Claude Cowork, the switchboard is Claude's admin console. Organization owners set capabilities such as web search, admins assign access by team, and Enterprise plans can stream activity to a SIEM through OpenTelemetry. People choose the folders and tools a task may touch, and destructive actions need approval. Projects group related tasks with their own files, instructions and memory. The record and the permission model are Anthropic features configured inside Anthropic's product.

Neither vendor hands that layer over as something you can clone, diff and review like code. That is the dividing line between the three, and the one the marketing pages argue around.

## What each vendor charges for

Microsoft sells access and then meters the work. Copilot Cowork needs a Microsoft 365 Copilot subscription for each person, and its tasks consume Copilot Credits under usage-based billing, with spending policies set in the admin center. An admin must enable Cowork and its usage-based billing ([Microsoft Learn](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-overview-copilot-credits), checked October 2026).

Anthropic bundles Cowork into paid Claude plans, from Pro to Enterprise, with usage limits on each. Team and Enterprise add admin controls and usage analytics ([Anthropic](https://claude.com/product/cowork), checked October 2026).

Kortix bills the agent computer. On Kortix Cloud, credits pay for the agent computer and bill by the second while it runs; a stopped computer costs nothing ([Kortix](https://kortix.com/pricing), checked October 2026). Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.

## Where an open-source stack fits

Kortix is the open-source AI Operating System: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers. That architecture matters, because it puts the layer the other two vendors keep to themselves in a repository you control.

The manifest in that repo is the switchboard. An agent is a markdown file plus a block in `kortix.yaml` that says which connectors, secrets and skills it may use. Project-wide rules decide whether a connector call runs, waits for approval or is blocked, and an approval's holding time is a setting. Session work reaches `main` through a change request. Merge is default-deny for agents. Open source. Read it, fork it, audit it.

Reach is comparable on paper: 3,000+ apps in a click, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP. A permission change is a diff someone can review, and an agent's memory is a file that survives a vendor's repricing, rename or product merge. [Self-hosting](/self-hosting) covers the install; the list of [open-source Copilot alternatives](/best-open-source-copilot-alternatives) puts the rest of the field side by side.

## A decision rule for the next purchase

Ask three questions in order.

First, where does the work already live? If Outlook, Teams, SharePoint and Office files hold it, Copilot Cowork runs inside the estate your admins already govern. If research, documents and folders across mixed tools do, Claude Cowork starts faster.

Second, who can answer "what is the agent allowed to do, and who changed that?" If it is "our Microsoft 365 admins, in the admin center", Copilot fits. If it is "our Claude admins, in Claude", Claude fits. Both are auditable inside the product, and neither gives you the file.

Third, must the agent's memory and permission set survive the vendor? To clone it, review it in a change request, run it on your own network or point it at another model provider, only Kortix answers yes, because the configuration is a git repo rather than a tenant setting.

That question decides whether the agent is an asset on your books or a line on someone else's.

## The switchboard is a file in the repo

`kortix.yaml`

```yaml
agents:
  kortix:
    file: agents/kortix.md
    connectors: all
    secrets: all
    skills: all
```

Source: https://kortix.com/docs/project/manifest

An agent and the connectors, secrets and skills it may use, declared in kortix.yaml.

## Common questions

### Is Copilot Cowork as good as Claude?

Neither vendor publishes a neutral test. Microsoft reports its own task tests, where Copilot finished work Claude did not, and those are vendor tests. Judge on where your work lives and who governs the agent, then run a trial on your own files.

### Does Copilot have something like Claude Cowork?

Yes. Copilot Cowork is Microsoft's agent for long-running, multi-step work across Microsoft 365, with each action approved before it runs. Microsoft built it on the technology behind Claude Cowork, then added Work IQ context and Microsoft 365 controls.

### Is Cowork different than Claude?

Cowork is the agent inside Claude, not a separate product. Anthropic has folded it into Claude: chat and Cowork share one message box, and a task can start on web, desktop or mobile. Claude is the product; Cowork is the mode that does the work.

## Put the agent's memory in a repo you own.

Self-host for free, or start on Kortix Cloud and connect your own models.

[Try Kortix](https://kortix.com) · [Self-hosting](https://opensourcecopilotalternative.com/self-hosting)