Microsoft Copilot is open source in two editor plugins

Published in two editor plugins, closed everywhere else

Microsoft's Copilot is open source in exactly two places, checked October 2026: the Copilot chat extension that runs inside VS Code, and the Microsoft-published GitHub Copilot plugin for Eclipse. Both have source you can read and fork. Copilot in Word, Excel, PowerPoint and Teams, Copilot Chat at copilot.cloud.microsoft, Copilot Studio, and the service that answers a completion request in your editor are hosted Microsoft products. No source to read, nothing to run on your own hardware.

That split is the whole answer, and it matters more than the phrasing of it. A company asking it usually wants to inspect, host and modify the thing that reads its documents and code. On three of the four surfaces, it cannot.

One name over four products

"Copilot" covers four products that share a brand and little else, and the source question has a different answer for each. Microsoft 365 Copilot answers questions about your mail, files and meetings. Copilot Studio builds agents and workflows. GitHub Copilot writes code inside an editor, a pull request or a command line. Each has its own plan, admin console and answer to what you may read. What Microsoft Copilot is covers the family. The source question is the one that splits it.

Microsoft 365 Copilot and Copilot Studio: closed services

Microsoft's own overview describes Copilot by what powers it: grounding in web and organizational data through Microsoft Graph and Work IQ, and access scoped by user permissions. None of that is a program you install. You open it at copilot.cloud.microsoft, in the Copilot app, or inside Word, Excel, PowerPoint, OneNote and Teams. An administrator decides which Copilot plan a tenant gets and which agents people may use.

The plan a tenant holds changes what the assistant can see. The basic Copilot Chat experience reads the web and only the content you paste into the prompt. The Microsoft 365 Copilot add-on grounds answers in your own tenant data and appears inside the Office apps. Without the add-on, Copilot cannot summarize a document it was not given.

Copilot Studio has the same shape one level lower. You build an agent or a workflow in a browser at copilotstudio.microsoft.com, connect it to data through connectors, then publish it to Teams, a website or Microsoft Copilot. Studio's own overview calls it a graphical, low-code studio for building and managing AI-powered agents and workflows. You own the definition you typed and the conversations it held. The engine that interprets the instructions, picks the tools and calls the model is Microsoft's, and it runs in Microsoft's cloud.

Both products carry real governance: data protection terms, tenant boundaries, admin reporting. Your security team still cannot read the code that touches the mailbox, and a review that stops at that question ends in no.

GitHub Copilot: a published client, a closed service

GitHub Copilot is the surface that makes the question worth asking, because part of it genuinely is published. The repository microsoft/vscode-copilot-chat holds the Copilot chat extension for VS Code. You can read the editor client: how the chat panel requests a completion, which tools it wires up, how it reaches MCP servers, how a session is drawn on screen. The README notes that the project moved into the main VS Code repository and is now archived.

What the published client does not contain is the service. No models, no routing between them, no code that assembles and serves a suggestion. The same README says an active GitHub Copilot subscription is required, and the extension releases in lockstep with the editor because the newest models arrive through Microsoft's service.

The Eclipse plugin tells the same story. Microsoft's announcement gives the reason as transparency and community contribution, and points readers at the published plugin repository. Its FAQ adds that the change does not affect a Copilot subscription: you still need a GitHub account with Copilot access.

Two clients are published, and both are clients. Reading them shows how the editor asks, not how the answer is made. A developer can audit what leaves the editor and where the trust boundaries sit; a company that wanted to run the assistant on its own infrastructure gains nothing.

What "open source" has to mean to run the whole thing

Sort the question into four tests. Can you read the code that handles your data? Can you run the system yourself, on your own network if your review demands it? Can you change behavior and keep the change? Can you walk away with your configuration intact?

The editor extensions pass the first test and fail the other three, because the service behind them is Microsoft's and the terms you accept belong to the product. Microsoft 365 Copilot and Copilot Studio fail the first test outright, which makes the rest moot. That is a delivery model, with support, compliance attestations and an admin console. It is not open source, and no Copilot tenant becomes open source by being configured differently.

A company that needs all four tests to pass is choosing a different system rather than weighing editions of Copilot. That is the ground self-hosting covers, and the ground the open source Copilot alternatives are judged on.

Where Kortix stands

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.

Run the four tests against it. Open source. Read it, fork it, audit it. A project is a git repo, so the agents, the skills, the memory and the connector configuration are files you can diff, review and carry out with you. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. One isolated sandbox per session. Each session has its own isolated machine and branch. Session work reaches main through a change request. Merge is default-deny for agents.

The difference from Copilot sits at the level of the repository. The part that reads your files, the part that decides what goes to a model, and the part that stores what came back all live in the checkout you cloned. A company can answer its own security questionnaire from that source rather than from a vendor page.