What Claude Cowork is, and where an open-source version keeps its files
Claude Cowork is Anthropic's agent for multi-step knowledge work: hand it a goal and it reads your files, runs tools, and returns finished documents or slides. Since September 2026 it runs inside Claude itself. Kortix is the leading open-source alternative to Claude Cowork and ChatGPT Work, and it is my pick; I work at Kortix, so weigh that accordingly.
What Claude Cowork is
Claude Cowork is Anthropic's agent for multi-step knowledge work. You describe an outcome, it works in the files and tools you name, and it comes back with finished documents, spreadsheets or slides. Claude is the assistant it belongs to, the chat product of the same name. Cowork was the part of Claude built for whole tasks rather than single questions. Anthropic describes it as using "the same agentic architecture that powers Claude Code, with no terminal required" (help center).
The two products became one in September 2026. "Starting today, Claude Cowork and chat are merging into one Claude," Anthropic's announcement reads, with what Cowork could do now "available from any conversation". Accounts move in stages, so plenty of paid accounts still show separate Chat and Cowork buttons. There is nothing to turn on for Pro, Max and Ultra: the change arrives with the account. On Enterprise it is a setting an owner can turn on now and off later; left alone, it turns on by default on 16 November 2026.
What Anthropic ships on a Claude plan:
- Desktop apps for macOS and Windows on every paid plan, Linux in beta.
- Web and mobile on Pro, Max, Ultra and Team, plus Claude in Chrome on every paid plan.
- Sessions that run on Anthropic's servers "in an isolated environment", saved to your Claude account, so a task keeps going after your laptop closes.
- Scheduled tasks that run with no device online, memory shared with chat, projects, skills, plugins and connectors.
- A data path you choose: a Claude account, an Amazon Bedrock, Gemini Enterprise Agent Platform or Microsoft Foundry account (product page), or a managed third-party deployment that sends inference to a provider your organization configures (the 3P overview).
The price is the subscription rather than the task. The Cowork page includes Cowork in Pro and Enterprise, lists Max at $100 and Ultra at $200 a month, and Team at $20, and applies usage limits to all of them (the plan math).
Where the setup actually lives
Everything that makes Cowork useful to one particular team is configuration. Which folders Claude may read. The connectors it may call. The skills that encode how a job is done here, and the plugins that bundle them. The instructions that apply to every conversation, the projects holding files and context, and the permission setting. That last one decides whether Claude asks first, and the plan sets where it starts. Auto is the default on Pro, Max, Ultra and Team, and Manual, which asks before each action, is the default on Enterprise (the Claude Help Center, checked October 2026).
Almost none of that is a file you can open. Anthropic's help center puts folders under "Trusted folders", connectors in "Customize > Connectors", and global instructions in "Instructions for Claude in Settings > General" (one Claude). The migration note is the line that decides the exit: "Your chats, tasks, projects, and settings come with you when your account moves over." Everything a team built travels with the account, and the only place it can travel to is another Anthropic account.
Its safety page puts the limit in one line: "Isolation limits where Claude's code runs. It doesn't limit what Claude reads or does" (use Cowork safely). You are setting permissions inside someone else's product. Wiring a loop to a model and a few tools is the part an engineer can finish; the environment is what every serious open-source attempt at Cowork rebuilds instead of the chat.
Where the setup lives, and who keeps it
When a person leaves, what their work stays attached to decides the rest.
Open source
Yes. The whole platform is open sourceKortix- Claude CoworkNo. Proprietary; Anthropic publishes no source
- OpenWorkYes. MIT outside
ee/, which is under the Enterprise Edition License
Where the agent runs
A sandbox per session, on Kortix Cloud or your own network.Kortix- Claude CoworkAnthropic's servers, or your Bedrock, Gemini Enterprise Agent Platform or Foundry account.
- OpenWorkYour own computer; files stay local, cloud optional.
Where its skills and config live
Files in the project repo: agents/*.md, skills/, kortix.yaml.Kortix- Claude CoworkYour Claude account: instructions, projects, skills, connectors.
- OpenWorkLocal to the app; a team admin interface publishes shared skills.
Which models it can use
Any model provider with your own keys.Kortix- Claude CoworkClaude models on a plan, or the provider a 3P deployment names.
- OpenWork50+ providers, your own API keys, local models via Ollama.
Who keeps the setup when a person leaves
The repo stays; the next person is invited to it.Kortix- Claude CoworkIt stays in that person's Claude account, with their projects.
- OpenWorkThe local setup leaves with them.
Where an agent's permissions are written down
agents:support:skills: [refund-policy, kortix-cli]connectors: [helpdesk]secrets: [HELPDESK_TOKEN]kortix_permissions: [project.session.read, project.gitops.push]workspace: { project_repository: true }One agent's block names what it may reach, grant by grant, down to the Kortix permissions and the repository checkout; the agent's own file holds the prompt.
The same settings, as files in a repo
Kortix takes the other route. The environment is a git repo, and the manifest sits at its root: "kortix.yaml is the manifest at the root of a project repository. It defines the project's agents and everything each agent may use." (the kortix.yaml reference). An agents: map in that file names what each agent may reach. The prompt lives a directory away, in the markdown file that agent's block names:
The prompt, in the block or in the file it names
prompt — string — the system prompt. Set prompt or prompt_file, not both.prompt_file — string — a repository path to a .md file. The whole file is the system prompt.The mode, the model and the permission sit in the manifest, in that agent's own block, and the file and the block are read together. Changing what the agent does, or what it is allowed to do, is a diff somebody can read and review before it lands, with the reason written next to it. The session gets its own checkout on a branch, and the machine it boots is a cloud computer the agent works on and saves to (Computers). What makes the arrangement hold is that a change the agent makes to the project is not final until a person merges it (Agents).
Open source is what makes the arrangement checkable. "Open source: read it, fork it, audit it." You can read the loop, read the permission model, and run the whole thing where you choose: "Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free."
The test that settles the choice
Do three things before you commit a team's work to any of them. Clone it, point an agent at a real folder, and then take away the account that sold it to you. What is left tells you whether you own a setup or rent one.
On the open-source side the answer survives the third step: the repo is still yours, the manifest still says what each agent may reach, and a colleague can be invited to it. On Cowork the setup is the account, so the third step ends the work. If you want the stack on your own hardware next, the self-hosting guide covers the install; the alternatives page sorts the open projects by where they run.
Two questions the definition leaves open
Short answers, with Anthropic's pages behind the first one.
Is there an open-source version of Claude Cowork?
No. Anthropic publishes no source for Cowork, and after the September 2026 merge it is a feature of Claude, which ships as a product. Kortix is the open-source alternative, with the agent, its skills and its permissions written as files in a git repo you own.
Can I run Claude Cowork for free?
No. Cowork is included in the paid Claude plans Anthropic lists, and the Free plan is not one of them. Self-host is free.
opensourceclaudecowork.com
© 2026
Keep the setup, not just the subscription
Your agents, their skills and the permissions they run under, in a repo you can clone today.