What happened to Claude Cowork, and what open source keeps
Anthropic merged Claude Cowork into Claude on 16 September 2026, so the Cowork toggle is going away and Claude now decides when a chat becomes a longer job. Kortix is the open-source CompanyOS, my pick: your agents, skills, company memory and connectors in one git repo you own; I work at Kortix, so weigh that accordingly. Details checked 29 October 2026.
The merge, dated
Claude Cowork was Anthropic's mode for work that outlasted a single answer. A request became a job: it read your files, used your connectors, and came back as a document, a spreadsheet or a deck. On 16 September 2026 Anthropic stopped keeping that in a separate place and folded it into Claude itself (VentureBeat).
- 12 January 2026. Cowork arrived as a research preview for Max subscribers in the macOS desktop app. It was an agentic workspace built on the architecture behind Claude Code, with no terminal (VentureBeat).
- By 16 September 2026. Every paid plan carried Cowork, on desktop, web and mobile, and Enterprise accounts had it where an admin enabled it (VentureBeat, 16 September 2026).
- 16 September 2026. "Starting today, Claude Cowork and chat are merging into one Claude," Anthropic's announcement reads (Anthropic). Existing chats, projects, artifacts, connectors and skills carry over, the Cowork name will eventually disappear, and Claude Code stays a separate product (VentureBeat).
- From 16 September 2026. Pro, Max and Ultra accounts move first on web, desktop and mobile, with more plans to follow. An Enterprise owner controls the timing: "Enable now" turns it on early, "Keep disabled" keeps it off, and it can be turned off again later. For them the unified experience is available where an admin turns it on: off by default in existing organizations, on by default from 16 November 2026 (the Enterprise admin guide).
- 6 October 2026. New Cowork tasks on Pro, Max and Ultra start in the cloud. Work on files stored on your own computer still needs the desktop app open (ClickForest, checked 3 October 2026).
- 23 October 2026. Anthropic's update: "the unified experience is now generally available for Claude Enterprise". Docs and Slides are new, "in beta on paid plans, and Enterprise admins choose when to turn them on". Pro, Max and Ultra have "nothing to turn on"; Team and Free follow (Anthropic, checked 29 October 2026).
What did not change
The work still runs through Anthropic's product, and the app, the harness and the approval step stay Anthropic's in every arrangement. Anthropic's help centre says Cowork "runs your tasks in the cloud" and that "Claude's work runs on Anthropic's servers, in an isolated environment". Sessions and files are saved to the Claude account (Claude Help Center). An organization can point Claude Desktop at its own inference endpoint instead. Conversation content then goes to Bedrock, Gemini Enterprise Agent Platform or Foundry, and history stays on the device (Claude docs). Either arrangement ships a client and a service: Anthropic publishes no Cowork source, and no edition of it runs as a service the team operates.
Permissions stay inside the product as well. A Manual or Auto setting in the message box decides whether Claude asks before each action (Claude Help Center). On a managed desktop deployment, an MDM profile delivers the configuration (Claude docs). A rulebook like that is a screen to click and, at larger companies, a policy file pushed to laptops. Nobody on the team reads it as a diff before it takes effect.
The agent now chooses the moment a chat becomes a job. That handover used to be the person's, a toggle under the input box. After the merge Claude makes it from the prompt, and scheduled tasks keep the job running with no device online (Claude Help Center). For a person this is simpler. For a team, it moves one more decision behind a screen the team does not own. The question of which jobs ran under what permissions then arrives after the fact.
Who runs the job, where the permissions are written and where the record sits all stayed inside Anthropic's product, exactly where the merge found them.
What survives the next rename
The merge carried Cowork's chats, projects, artifacts, connectors and skills into Claude, and the separate Cowork surface goes away as the rollout reaches each account. A team that reopens the question "which Cowork-like product do we buy" will reopen it again at the next announcement. The files sit in each member's Claude account, the grants sit in the organization's settings, and the session record comes from the Compliance API on Enterprise, because Anthropic's audit logs do not carry Cowork activity (Claude Academy).
Kortix is built so that layer is a repository. An agent is one agents.<name> block in kortix.yaml (behavior and grants) plus the prompt file it names. The file holds the prompt; the block holds the model, the mode, the permission and the connector, secret and skill grants. Nothing about it lives in a dashboard; the pair is the whole agent. Each session works on its own branch, a private copy of the repository. Its work reaches the shared repository only when a person merges its change request (Computers).
kortix.yaml registers what exists in the project, and each agent enables what it needs by name. An omitted skills, subagents, commands, extensions, connectors, secrets, apps or kortix_permissions grant resolves to none, while tools defaults to "*" with a kortix validate warning, permission to allow, and codemode to on (the kortix.yaml reference). Work then lands the way code lands. Session work reaches main through a change request, and merge is default-deny for agents. An admin can grant project.gitops.merge to change that.
The repo is readable and the permission is a diff a colleague reviews. Every action is recorded. Reading, exporting and streaming the audit log depend on the plan. Who may merge is written down too. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.
If you are still working out what the product was, what Claude Cowork is covers it. The alternatives page sorts the open projects by where they run, and moving a team off Cowork covers the handover.
Four options, six questions
What each option still hands you after a rename: the files, the record, the exit.
Open source
YesYes, the whole platformKortix- Claude, with Cowork inside itNoNo
- OpenWorkYesYes, the desktop app
- ChatGPT WorkNoNo
Where the work runs
Kortix Cloud, your VPC or on-prem, one sandbox per sessionKortix- Claude, with Cowork inside itAnthropic's servers, or your Bedrock, Gemini Enterprise Agent Platform or Foundry endpoint
- OpenWorkYour own machine; cloud workers optional
- ChatGPT WorkOpenAI's cloud, with a desktop app on your machine
Where permissions are written
One agents.<name> block in kortix.yamlKortix- Claude, with Cowork inside itA Manual or Auto setting per conversation
- OpenWorkLocal app settings; Enterprise desktop policies
- ChatGPT WorkAdmin Console, by workspace and group
Audit trail
Every action is recorded. Reading, exporting and streaming the audit log depend on the planKortix- Claude, with Cowork inside itCompliance API transcripts for Enterprise; no Cowork events in the audit log
- OpenWorkAudit log and spend observability on Enterprise; every agent action attributable and exportable to a SIEM
- ChatGPT WorkA Compliance API for Enterprise and Edu
Self-host
YesYes, on your own networkKortix- Claude, with Cowork inside itNoNo
- OpenWorkYesYes, or a managed private instance
- ChatGPT WorkNoNo
Model choice
Any model, your keysKortix- Claude, with Cowork inside itClaude models, or a configured provider
- OpenWork50+ providers, local models
- ChatGPT WorkOpenAI's models
What one agent block holds
default_agent: assistantagents:assistant:description: Answers questions about the project.prompt: You answer questions about this project. Read the files before you answer.skills: "*" # every skill; omitted: no skillsworkspace: { project_repository: true } # omitted: no repository in /workspaceWhat the reference calls its smallest useful manifest: a default agent and one agent block carrying a description, a prompt, a skills grant and a workspace grant. The block can name a prompt file in place of the prompt, but not both.
Questions the merge left behind
Short answers, with the dates Anthropic gives.
Does Claude Cowork still exist?
Yes, as a capability inside Claude. On Pro, Max and Ultra its work now happens inside ordinary Claude conversations, and the toggle disappears as the rollout reaches each account; on Team and Enterprise plans the merged experience arrives when an owner turns it on. Anthropic says the Cowork name will eventually disappear, while the work it did stays.
What happened to Claude Cowork?
Anthropic merged it into Claude on 16 September 2026, so one conversation can answer a question and then carry on as a longer task. Chats, projects, artifacts, connectors and skills carried over, and Claude now decides when a request becomes a job.
Did I lose my Cowork projects and files?
No. Anthropic says existing Cowork tasks, projects, connectors, skills, artifacts and files come with the account, and those tasks open as they did before (Claude Help Center). On desktop, the folder Cowork saved work to is still listed as the storage folder in Settings.
Why can't I see the new Claude yet?
Because the rollout is staged. Pro, Max and Ultra go first on web, desktop and mobile over the coming weeks, and accounts on the same plan can move at different times. On a Team or Enterprise plan, the merged experience is a setting an owner turns on, so ask yours. On Pro, Max and Ultra nothing needs switching on: if the message box still shows Chat and Cowork, the account is not in the wave yet.
Is Claude Code going away too?
No. Anthropic keeps Claude Code separate for developers working from terminals and IDEs against a codebase, and the September announcement covered Cowork and chat only (VentureBeat).
Do I still need Claude Code if I don't code?
No. Cowork already used the same agent architecture as Claude Code without the terminal, and after the merge that work happens in the chat a non-developer already uses.
opensourceclaudecowork.com
© 2026
Clone the layer before the next rename
Your agents, their permissions and the merge gate live in a repo you can clone today.