Is Copilot Cowork open source? No, and there is no self-host option

Is Copilot Cowork open source?

No. Microsoft publishes no source for Copilot Cowork, and its documentation offers no install, no repository and no self-hosted build. It is a feature of Microsoft 365 Copilot, served from Microsoft's cloud to accounts inside a Microsoft 365 tenant.

Microsoft's general-availability post sets the terms in one line: Cowork requires the Microsoft 365 Copilot User Subscription, the per-seat plan known as the USL, and that seat comes before anything runs. Every task is then billed in Copilot Credits, priced from four inputs Microsoft names as model use, context retrieval, tool calls and runtime, with pay-as-you-go at $0.01 per credit. The subscription sits underneath and a meter runs on top. Nothing in that stack is a file you can read.

Where the work and the configuration actually live

Microsoft's own documentation, checked October 2026, answers both parts. The March launch post says Cowork "runs in a protected, sandboxed cloud environment". The Cowork FAQ lists the consequence as a design limit: "Cowork can't access or edit files stored locally on your device. It works with files in OneDrive and SharePoint." Browser tasks are the exception, and those run in Edge on the machine in front of you, with your own sign-ins.

The configuration sits where the files sit. Custom skills are SKILL.md files, and they belong in /Documents/Cowork/skills/ inside OneDrive, fifty at most. The parts look like files a team could version. The folder belongs to a tenant, and the tenant belongs to Microsoft, which is where the version history stops.

Access runs through the same gate. Microsoft's admin guide states that access "is granted only through a spending policy that selects Cowork", and that a spending policy is "an access control, not only a budget". Cowork also arrives switched off until an administrator turns it on. One console holds the on switch, the skill folder, the model list and the audit trail.

The ownership answer is in the plan list

Three product decisions set what a tenant can do with Cowork. It arrives switched off until an administrator enables usage-based billing. Access is granted through a spending policy that selects Cowork. And the model list is an admin setting: an organization can turn off the Anthropic model family and keep the GPT models, or the reverse.

At general availability Cowork ran on Anthropic's models, the Claude Opus and Sonnet families among them, with OpenAI's GPT models available in Frontier. Which model does your work is therefore a tenant setting, not something a team configures in its own repository with its own keys. Add per-task credits on a runtime priced by model use, context retrieval, tool calls and time, and the picture completes. Capacity, governance and model choice are three dials in one console, and none of the three belongs to the team using it.

Microsoft chose that shape on purpose and argues for it well: predictable governance inside a boundary your auditors already recognize. The effect is that the ownership question gets answered before a buyer sees the product.

Two products called Cowork, and the question that survives

Anthropic merged Claude Cowork and chat into one Claude on 16 September 2026, so Cowork is now a capability inside Claude rather than an app of its own; Pro, Max and Ultra plans get it first, with more plans to follow. Microsoft had already integrated "the technology behind Claude Cowork into Microsoft 365 Copilot" at launch. Two vendors ship the same approach to delegated, long-running work, and both ship it closed.

The question that survives is narrower and easier to answer. Can your team hold the system that does the work? Kortix answers with a repository: its agents, skills, company memory and connectors are files in a git repo the team owns. Any model provider with your own keys. Where Copilot Cowork's model list is a toggle someone else administers, Kortix's is a key your team holds.

When Copilot Cowork is the better buy

It depends on which risk you are pricing. If your company already runs Microsoft 365, keeps its data inside that boundary and answers to an auditor who knows Purview, Copilot Cowork fits. Sensitivity labels carry through a session, prompts and outputs land in the audit log, and the policies that govern email govern the agent too. Nobody has to run a server, hold a sandbox key or take a call at 2am. For that team the closed design is the feature, and the meter is the honest price of not operating anything.

If ownership means the work and its configuration can leave with you when the contract ends, the same design is the problem. Two questions are worth asking before signing: where the skills folder lives, and who can switch it off.

The open-source alternative, and what you can run today

Kortix is the open-source alternative, and the part of it that matters is the boundary. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Read the docs for the install; the self-hosting guide covers what a Linux box needs before the first start.

Today you can clone the repository and read how a session starts, how skills load and where memory is stored, before an agent touches real work. If the field is still open, the comparison puts the options in one table, and Claude Cowork vs Kortix takes the Anthropic side apart. Teams with a data-residency question usually want whether Claude Cowork is safe settled first.