Kortix vs Claude Cowork: Cowork went everywhere, Kortix stayed open source
Cowork became part of Claude in September 2026; Kortix stays open source, self-hostable, and built to run a company's agent work.
Cowork is now one excellent agent inside Claude itself. Kortix gives that same working power to a whole company, governed and self-hostable.
Claude Cowork earned its reputation as a strong single-player agent, and open-source Kortix is the CompanyOS for teams that need to own the system. On September 16, 2026, Anthropic folded Cowork into Claude itself: what it does now starts from any conversation. The rollout starts with Pro, Max, and Ultra plans on web, desktop, and mobile; Team (in beta) and Enterprise organizations see it where Chat and Cowork unified is turned on. What changed is reach. The terms are the ones Cowork always had: closed source, Claude models only, no self-hosting. The choice in front of a team is starker now: rent a feature, or own the system.
Cowork is now part of Claude
Anthropic's release notes for September 16, 2026 carry the heading "Claude Cowork comes to every conversation." Chat and Cowork are one surface now: a quick question and a bigger job, like a report or a spreadsheet, happen in the same thread. Chats, Cowork tasks, projects, connectors, and skills carry over. For an existing organization the Chat and Cowork unified setting stays off until November 16, 2026, when it turns on by itself and can still be turned off afterwards. An account that has moved cannot go back to separate Chat and Cowork options.
Two changes earlier in 2026 led into the merge. Since July 7, Cowork sessions have run remotely in the cloud. Work continues after the laptop closes, and scheduled tasks run with no device online. Since August 25, memory spans chat and Cowork in the cloud; Free, Pro, Max, and Ultra plans get it by default, while on Team and Enterprise plans it stays off until each member turns it on.
The case for Cowork
- It returns finished work. Give it a goal and it searches, reads your files, and runs code. What comes back is documents, spreadsheets, and presentations you can open and edit.
- A permission model you set. When approval checks are on, Claude shows its plan and waits for your approval before anything significant, and deleting anything needs your approval. The merged experience adds a permission setting: Auto is the default on Pro, Max, Ultra, and Team plans and keeps working behind automated safety checks; Manual is the default on Enterprise plans and asks before each action. Either one can be changed at any time.
- It works while you're away. Tasks keep running in the cloud after the laptop closes, and a scheduled task runs on any cadence, unattended.
- Extend it with skills, plugins, and connectors. Plugins bundle your tools, knowledge, and workflows into one install; on Team and Enterprise plans, admins manage them for everyone.
- Enterprise observability. On Enterprise plans, Cowork activity streams to your SIEM through OpenTelemetry, with usage analytics and an Analytics API beside it.
Where it stops being enough
- Closed, and not hostable. Cowork runs in Anthropic's cloud, or in your own Amazon Bedrock, Google Cloud, or Microsoft Foundry account. Either way it is Anthropic's software on Anthropic's terms; there is no version your team hosts.
- Claude models only. Anthropic's product pages document no bring-your-own-model path for Cowork. Every task runs on a Claude model, at the rate your plan sets.
- One usage pool. Everything you do with Claude counts toward your plan's usage limits, and agentic tasks draw that pool down faster than chat does.
- Configuration lives in Anthropic's product. Skills, plugins, connectors, and memory are settings in Claude's apps and state in Anthropic's cloud. A company's way of working exists there, where nobody on the team can version it, diff it, or audit it.
For one person, that is a fair deal. Plans that include Cowork start at $17 a month billed annually, or $20 month to month, checked October 2026. The terms start to bite the moment the tool has to serve an entire company instead of one desk.
The other answer to those terms
Kortix was built on a different bet. An agent's behavior, its skills, the company's shared memory, and the rules for each connector are files in one git repository. Teach one agent something and every teammate and every other agent inherits it. The setup accumulates rather than starting over with each new hire. Every change has an author, a diff, and a reviewer, so a new way of working gets the same scrutiny a code change gets.
The repository is also the governance mechanism. An agent opens a change request, and a person decides what lands on main; by default agents hold no merge rights, though an admin can grant that one capability.
Nothing about how the platform works is hidden from the teams that run it. Open source. Read it, fork it, audit it. Scale comes out of the same files: thousands of agents in parallel on one config, each on its own cloud computer.
Where the data has to stay decides where the system runs. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.
Who picks the model, and who pays
Cowork runs on Claude models, so the engine choice is made for you. Anthropic's API list prices put Opus output at $20 per million tokens. The frontier Fable family runs $50, checked October 2026. A Cowork plan hides those numbers behind usage limits, and agentic work drains a limit fast.
Kortix runs on the opposite arrangement. Any model provider with your own keys. Or the ChatGPT plan you already pay for. Open-weight models put real numbers on the gap. Z.ai's price page lists GLM-5.2 output at $4.4 per million tokens. DeepSeek's deepseek-flash prices 1M output tokens at $0.6 off-peak and $1.2 at peak, both checked October 2026. Send the routine volume to a cheap model and reserve frontier tokens for the steps that justify them.
The same workload, a fraction of the bill, on computers your company owns.
Row by row
Who each one is for
you want one polished assistant for yourself or a small team, and you accept Anthropic's models, Anthropic's terms, and one usage pool.
you want the same working power run as company infrastructure, with agents in parallel, any provider with your keys, and a setup that lives in your own repository.
One limit worth naming on our side: Kortix agents work on their own cloud computers. Cowork, on Pro, Max, and Ultra plans, can also take over your screen in a beta computer-use mode, checked October 2026. If your work lives in local files and desktop apps, that direct reach can decide the question by itself. Running both is a real pattern, too: a power user keeps Cowork for personal tasks while the company's agent fleet runs on Kortix, where people review what it does.
Two related posts, if you want to go deeper: ChatGPT Work vs an open-source alternative your company can own covers the other hosted option, and Personal AI agents vs a company OS draws the line between a personal assistant and a company system. When you are ready to try the platform itself, Getting started with open-source Kortix takes four commands.
Put a fleet of agents to work under rules you set. Connect your tools, give one a real piece of work, and approve the change request it opens when the job is done. Trying it costs nothing, and so does self-hosting. Try Kortix