Microsoft Copilot vs Claude: the open-source layer that outlasts both

Copilot and Claude are bought differently, and that decides more than the feature list

A search for "microsoft copilot vs claude" usually means Microsoft 365 Copilot against Anthropic's Claude. GitHub Copilot is a different product with a different bill, so it stays out. They are not rival editions of one purchase. Microsoft sells Copilot as an add-on to a Microsoft 365 plan the tenant already runs; its business pricing page describes adding Copilot to an existing subscription and says Copilot Chat comes with eligible plans at no extra cost. Claude is sold the other way round. Anthropic's pricing page lists individual, Team, and Enterprise plans, and its Claude Platform prices models per token for teams that build on the API. The layer neither one hands over is where Kortix 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 difference sets what each vendor keeps. Copilot's value comes from the tenancy: Work IQ grounds answers in mail, chats, documents, and meetings, and Microsoft says prompts and responses stay inside the Microsoft 365 service boundary under Purview, retention, and eDiscovery settings. Claude's value sits in the model and the surfaces built on it: chat, Cowork, Code, Skills, and an API with tool use, file handling, and a context window of up to a million tokens. One is an assistant attached to a system of record a company already runs. The other is a model a team attaches to systems it chooses.

Who each one suits

Copilot suits an organisation already living in Word, Excel, Outlook, and Teams, where IT has retention and data-loss rules tuned, and where the ask is a first draft, a meeting recap, a formula, or a summary of a long thread. The assistant arrives where the document already is, and governance arrives with it because the data stays inside the boundary the tenant set. The limit follows from the same design: the assistant is useful in proportion to how much of the company's work already sits in Microsoft 365.

Claude suits a team whose work is mostly language and code, where model quality decides the output, and where someone will do the wiring. Anthropic's enterprise page describes connectors an admin approves, project memory, skills deployed across an organisation, and an API for products and agents. It reads as a platform for a team building its own workflow, not a finished suite inside an office app. Claude has Word, Excel, PowerPoint, and Outlook add-ins, so the overlap is real. The office suite is the product for Microsoft and an add-on for Anthropic.

The settings both vendors keep

Microsoft keeps Copilot's controls in the Microsoft 365 admin center, where usage, spending, agent creation, and analytics sit, and a model choice sits in the app. Copilot Memory sits inside Work IQ. Anthropic lists what an enterprise admin holds: single sign-on, provisioning, role-based access, connector approval, audit logs, retention, and organisation-wide skills. In both, an admin configures the product from a console, and the memory sits in the vendor's service.

That matters on the day the subscription stops. A team that has wired agents, rules, and memory into a console has to rebuild them somewhere else. A team whose agents are files can copy the repository and point it at a new model. Which console is nicer to use matters less than which one you can leave.

Microsoft's own comparison page makes its case and should be read that way. It argues that Copilot completes multi-step work "across Microsoft 365 apps and within Microsoft 365 security and governance boundaries", and that in its tests Claude's add-ins did not apply Purview protections or find the right document. That is Microsoft testing Microsoft inside Microsoft's boundary. The claim worth keeping from it is narrower: in this pair, the boundary is the product. Checked October 2026.

Where the agents run

Both vendors now sell agent work, and both run it on their own computers. Microsoft puts agent work in Copilot Studio inside the tenant and bills advanced work in Copilot Credits. Anthropic offers hosted environments for agents on the Claude Platform, with vaults and memory, alongside the API. Neither hands your team the machine, the harness, or the agent's files.

Kortix fills that layer. Open source. Read it, fork it, audit it. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. Any model provider with your own keys. Session work reaches main through a change request. Merge is default-deny for agents.

That changes what a buyer can check. The runtime has a location the team chooses. The model is a setting, so another provider is a configuration change rather than a migration. And each agent is text in a repository, so the security review can read it, a change arrives as a diff, and an auditor can be handed a clone.

A decision guide

Three questions settle most of this.

1. Where does the work start? If it starts inside a Word document, an Outlook thread, or a Teams meeting, and the tenant already enforces Purview, buy Copilot seats for the people in those apps, measure a quarter, then widen. The base-plan rule and caps are on Microsoft Copilot plans, and Microsoft 365 Copilot pricing covers how the bill is assembled.
2. Who will do the wiring? If the answer is a platform team and the output is a product or an internal tool, Claude is the stronger starting point: plans for people, an API for the build, and add-ins where the team still works in Office.
3. Who holds the file? If the answer is neither vendor, add a layer you own before the agent fleet grows. Start with one workflow, put the agent, its skills, and its memory in a repository, and connect a model you already pay for. The open-source alternatives round-up shows what that layer looks like next to the hosted products.

The pair can coexist. Copilot for the document work, Claude for the building, and a repository for the definitions. What neither fixes is the exit. Somewhere in the stack, a team should be able to answer "where does the configuration live" with a path rather than a console URL. If the answer is a path, the model becomes a component. If it is a console, the vendor owns the road out.