Open source AI agent platform vs Claude Cowork: settings or repo

Cowork is good at one person's desk

Anthropic merged Cowork and chat into one Claude on 2026-09-16. The help center says the distinction between the two products is gone: what you could previously only do in Cowork is available from any conversation. What comes back is finished work. Hand it five interview notes and it returns the themes with a quote for each. Hand it an analysis and it returns a ten slide deck you can open in PowerPoint. Give it a folder and it reads, organizes and writes files there. Long tasks keep running in the cloud after you close the laptop. Scheduled tasks run with the machine off, and the default permission setting asks before each action.

For one analyst with a laptop full of half-finished work, that is a good deal. The plan list covers it from a personal plan upward. The narrower question is whose repo the work lives in when a second person, or a reviewer, needs it.

Two cases where a personal setup stops fitting

The first is a colleague who needs the same agent. In Cowork, what you teach it belongs to your account: skills are something you give it, and the instructions for Claude apply to your conversations. Anthropic's plan table lists organization wide skills deployment on Team and Enterprise, and project sharing from Team upward. The documented path to handing someone your setup is a plan change and an admin. After that, the setup still lives inside Anthropic's apps.

The second is a reviewer asking what the agent did in March. Audit logs and the compliance API appear on the plan table for Enterprise plans only. The help center's advice for consequential work shows the same shape: stay close and review what Claude does. That is sound advice for one person watching one task. It is thin cover for a reviewer who arrives a quarter later.

What a repo changes about both cases

A Kortix project is a git repo with kortix.yaml at its root. Agents are markdown files, skills are folders of markdown and scripts, and memory is a directory. The manifest declares which agents exist and what each one may reach. The docs put it in one sentence: agents, skills, memory, tools, triggers and connectors are files in the project repo.

That one property settles the first case. Hand a teammate the repo and they have the agent, its skills and its grants, because those are the files. When you teach one agent something new, you commit it, and everyone who pulls inherits it. Nothing has to be redeployed from an admin console. Nothing walks out of the door with the person who wrote it.

The second case follows from where sessions run. One isolated sandbox per session. Each session has its own isolated machine and branch. The sandbox clones the repo and works on a new branch. The agent commits, pushes and opens a change request, and a person with the merge permission reviews it before it reaches the default branch. The review trail and the audit trail are the same object: a branch, a diff and a merge record, sitting next to the work itself.

Because reach is declared in files, a change to what an agent may touch arrives as a commit with an author on it. That is the part a settings screen cannot give you. It is also the part a security review asks for.

What ownership costs

Hosting is a choice you make and then operate. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. It moves the update hour, the database backups and the key that encrypts every stored secret onto your side of the line. The self-hosted versus managed page prices that trade in agent-computer hours, which is the honest unit for it.

Model spend is the second bill. Any model provider with your own keys. That is the point of the arrangement, and also the work. Somebody has to choose which model runs which step and watch what it costs. The pricing page splits sandbox, memory and model cost, so the number is visible before you commit.

The third cost is a person. Someone owns the repo, the grants and the review queue, or the agents rot quietly. Where no one will take that role, Cowork's settings screen is the cheaper answer. The team should buy it without ceremony and revisit the repo when the second person asks for the same agent.

Where the two sit together

Keeping Claude for your own desk work while the shared agents live in a repo is a real arrangement. Nothing about either arrangement forbids it. The two answer different questions, one about a person's afternoon and one about what the company keeps.

Three questions that decide it

Name two people who need the same agent. Ask where the skill a departing colleague built will be next month. Ask what a reviewer opens to see what an agent did in March. When the answers are "my account", "gone" and "a support ticket", the decision in front of you is about a repo. Cowork was built to answer a different question, and it answers that one well.