Claude Cowork vs Open Cowork: the open-source Kortix answer for teams

What Open Cowork actually ships

Open Cowork is an Electron desktop app whose code is public and free to fork. Its authors call it the open-source implementation of Claude Cowork. The readme lists what that includes: one-click installers for Windows (.exe) and macOS on Apple Silicon (.dmg), and Linux only by building from source. A graphical front end wraps Claude Code, OpenAI, Gemini, DeepSeek and other models behind an API key you supply. Skills for PPTX, DOCX, XLSX and PDF ship with it. MCP connectors reach browsers, Notion and other desktop apps. Computer use drives a screen, and Feishu or Slack can send instructions from a phone.

For one person with a folder of files and a key, that is a working answer to a permission problem. Install it, choose a workspace folder, paste the key, and the agent reads and writes where you point it. The readme's note on data is plain: the app runs locally, the files stay in the workspace, and the only external calls go to the model API you configured.

Isolation comes in levels. With WSL2 on Windows, commands run inside a Linux virtual machine. With Lima on macOS, an Ubuntu machine. Without either, commands run natively under path-based restrictions, so the boundary is the app's own file resolver rather than a machine.

A setup that lives on one laptop

Settings belong to the installation. The readme puts them in the app's config store and its sessions in a local database. The workspace is the folder you chose. Nothing there describes a shared file that lists which agents exist, or a copy of the setup a colleague can pull.

The cost appears on the second day of a second person. A colleague installs the same app and starts from a blank one: default skills, no history, none of the instructions you spent a month sharpening. There is no record of which folders you had cleared it to touch. Your prompts and your folder choices never existed as anything another machine could read.

Remote control does not close that gap. Feishu or Slack can hand a task to the desk machine, which assumes the desk machine is awake, online and yours.

Who reviews the edit

Open Cowork does stop and ask. Its permission rules default to ask before a write, an edit or a shell command, so a dialog waits at the machine before anything lands (the defaults in its source). That is a real check, and it matches the readme's warning to be careful when authorising file changes or deletions.

What it does not add is a second person. The prompt appears on the machine that asked for the work, and whoever sits there answers it. Nothing arrives as a diff for someone else to read, and the approval leaves no record another machine can open later. If the agent rewrites a budget spreadsheet badly at eleven at night, the only sign-off came from the person who asked, in the same window.

For a solo operator that trade is fair. Plenty of people want exactly this: no server, no account, no per-seat bill, their own files, their own key. It stops fitting when a job repeats every week and three people take turns. The fourth run inherits no memory of the first three, and the person at the keyboard approves work they never asked for.

What changes when the setup is a repo

Kortix puts the configuration in git and moves the computer off the desk. A project is a repository. kortix.yaml at the root declares the machine image, connectors, triggers and the secret names and grants. Agent files, skill folders and memory sit beside it as plain text. Someone who joins clones the project and starts with the same agents, the same connectors and the same permissions on their first morning.

Sessions run elsewhere. 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. What a reviewer opens is a diff with a branch behind it, and the record of who approved what outlives the person who wrote it. For a team that wants the machine inside its own network, self-hosting is the other half of the arrangement.

Models stay a choice the project makes. Any model provider with your own keys works. The repository holds credential names instead of values, so a key can be rotated without editing an agent.

Where Claude Cowork ended up

Anthropic merged Cowork into Claude on 16 September 2026 (the announcement). The capability moved next to chat, where a task can start in a conversation and arrive as a document or a deck. The product page sells it inside the paid plans: Pro at $17 a month billed annually, Max at $100, Ultra at $200, Team at $20 a seat, Enterprise quoted (checked October 2026).

Hosting is Anthropic's call. On Pro, Max and Ultra, sessions run in Anthropic's cloud. On Team and Enterprise, an owner's "Run Cowork in the cloud" toggle picks the cloud or a local virtual machine on each member's computer. Either way the models are Anthropic's. A company that wants the whole thing in its own account has no self-host to reach for.

The gap for a team runs in the mirror direction. A closed seat keeps the instructions, the skills and the record inside a product the company rents by the month. The permission question is settled by paying rather than by reading code. Open Cowork removes the payment question and leaves the state on a laptop. Kortix is the arrangement where both sit in a repository the company holds, so it takes the first column and the recommendation in the comparison table.

Choosing between the three

Take Open Cowork when the work is yours: personal files, your own key, one machine, nobody waiting on the output. It costs nothing to install, it runs on hardware you already have, and each change stops at a prompt on that machine.

Take Claude Cowork when you would rather Anthropic operate the thing. The assistant is polished, it keeps working after the lid closes, and the bill is a plan you may already pay. What you give up is the right to read the code, to host it, or to hand the setup to someone outside the seat.

Take Kortix when the agents are company property. A second person can read them, and a reviewer can see what changed. The same instructions serve the tenth run as well as the first. If you are still weighing the field, the alternatives page and the OpenWork comparison cover the other open-source options. Whether Cowork is open source settles the question for Anthropic's product.

The other projects a buyer weighs

A reader choosing between Claude Cowork and Open Cowork is choosing among five projects that publish their code. Two facts separate them: whether the project is open source, and the machine the work runs on. Kortix heads the comparison and is the recommendation when a second person has to read the work before it lands.

Four of the five are open source throughout. Open Cowork's readme describes "a free, open-source AI agent desktop application for Windows and macOS" that keeps the files on your machine (readme). CoWork OS calls itself "a free, open-source AI desktop app" on its own site, and installs as a desktop app on macOS and Windows or a command-line package on Linux (coworkosapp.com). CoWork by Cosyntic is open source and "Self-hosted or cloud", and runs as a shared workspace rather than an installed app (cosyntic.com/cowork). MindsHub is open source; its work runs in a desktop app on your computer or in its hosted app (github.com/mindsdb/mindshub, mindshub.ai/agents).

OpenWork needs a second sentence. It is open source outside its enterprise directory, and the console inside that directory needs a subscription, free for organizations of up to five people (README). The desktop app runs on macOS, Windows and Linux, and cloud is optional.

The pattern is worth more than any single row. An open-source project tells you what you may do with the code. It says nothing about where the work runs, or who else can see it. Two open-source projects here can put the same task on your laptop or in a vendor's account. Open Cowork and CoWork OS keep it on the machine. CoWork by Cosyntic and MindsHub offer a shared or hosted surface, and OpenWork sells the console that governs a team. For a team, Kortix is the pick: the agents, their skills and the connectors are files a new colleague reads on the first morning.