# Claude Cowork vs Kuse Cowork: open-source Kortix puts the agents in git

> Kuse Cowork rebuilds Claude Cowork's idea as a desktop app: a folder on your disk, your own key. That suits one person. A team needs agents where a colleague can reach them, so my pick is Kortix, the open-source CompanyOS, with the setup in a repository the company owns and each session's work as a change request; I work on Kortix's marketing, so weigh that accordingly.

URL: https://claudecoworkalternative.com/claude-cowork-vs-kuse-cowork

## Claude Cowork or Kuse Cowork: which one should you install?

Three cases, one line each. One person, own laptop, own files: Kuse Cowork. You want the deck or the tidied spreadsheet handed over: Claude Cowork, inside the plan you already pay for. Two or more people running the same agents, with somebody signing off: Kortix. The repository is what settles that last case.

## The jobs Kuse's README lists are one-person jobs

Sort my folders. Turn a shoebox of receipts into an expense report. Pull the action items out of a meeting transcript. Those are the sample tasks Kuse Cowork shows off, and each ends with a file the person at the keyboard can open. Point the app at a folder, pick a provider, read the plan, watch it work. The agent is Rust behind a Tauri window of roughly 10MB, built for Windows, macOS and Linux.

Its pitch is local-first execution, model freedom and privacy control ([Kuse Cowork on GitHub](https://github.com/kuse-ai/kuse_cowork)). Claude and GPT run on keys you paste in; Ollama, LM Studio or any OpenAI-compatible endpoint runs the local models. Skills cover docx, pdf, pptx and xlsx, and MCP servers reach other tools. Docker puts a wall between the agent's shell commands and your files. Skip Docker and the wall is gone, the same README warns.

The version history is short. v0.0.2 is the newest tag, the January 2026 notes call Word and Excel editing experimental, and the project asks for care when you hand it local folders (checked October 2026).

## What a second person needs that an install cannot give

When a colleague wants the same receipts job next month, Kuse's answer is the app again plus a key of their own. Two installs, two settings files, two readings of the same prompt, drifting from the first edit either one makes. Nothing anywhere compares the pair.

That is fine while the job owner is also the job. It gets expensive when the task returns every week and three people take turns on it, because the instructions live in one head and in its copy on one laptop. The next person to run the job starts by asking what changed since last time.

## What a session hands back when the setup sits in git

In a Kortix project the instructions are files in a repository the company owns. `kortix.yaml` says what each agent may reach, and each block there holds its description, mode, model, permission and grants. `agents/<name>.md` holds the prompt alone, `skills/<name>/SKILL.md` holds what the agent can do, and `memory/` holds what it learned. A reviewer reads the diff the way they read code ([the starter manifest](https://github.com/kortix-ai/CompanyOS/blob/dev/packages/starter/templates/base/kortix.yaml)).

Each session opens a checkout of that repository on its own branch. One line in the manifest grants it, `workspace: { project_repository: true }`, so the folder an agent edits outlives the laptop that asked for it. One isolated sandbox per session. Each session has its own isolated machine and branch.

Approval uses the same route the code does. Session work reaches `main` through a change request. Merge is default-deny for agents. Somebody reads what the agent did and clicks approve, and the click has a name attached. An admin can hand an agent the merge capability if the team wants that, and the grant is itself a decision in the log. A quarter later the diff and the signature are both still there.

## Where each one leaves a team after a month

Kortix, when the agents belong to the team rather than to one person. Several people take turns on one class of job, the instructions have to match for all of them, and somebody has to answer for what went out. The repository carries the setup, the change request carries the sign-off, and each session works in its own sandbox.

Kuse Cowork, when the laptop is the office: one person, files in a folder, a provider key already in hand, no account in between. Aim it at the folder and read its plan.

Claude Cowork, when you want the work run for you: a deck, a reconciled spreadsheet, a stack of contracts, inside the plan you already pay for, with the desktop app installed wherever local files are needed.

## What each one bills, in keys and credits

Kortix starts free: 500 credits a month on one project, no card, with model calls on your own provider key or a ChatGPT subscription you connect. Paid credits come in monthly packs from $50 ([Kortix pricing](https://kortix.com/pricing), October 2026), and a pack you do not finish carries into the next month.

Kuse Cowork carries no price tag of its own: clone, build with Node and Rust, and pay the provider you point it at. A colleague adds nothing to that bill and receives nothing from it. One-click installation is still on its roadmap (checked October 2026).

Claude Cowork is no longer a product of its own. Anthropic folded it into Claude, rolling out to Pro, Max and Ultra plans with more plans to follow, so the bill is whichever Claude plan you keep ([Claude pricing](https://claude.com/pricing), October 2026).

## Where a desktop app and a git repository part company

Kuse installs beside your files, Claude Cowork runs inside Claude, and Kortix keeps the setup in a repository a team reads.

| | Kortix (recommended) | Claude Cowork | Kuse Cowork |
| --- | --- | --- | --- |
| Open source | Yes, the source is public | No, closed inside Claude | Yes, the app source is public |
| What it is | Agents and their setup in one git repo | A delegated task inside Claude | A desktop app beside your own folders |
| Where a session runs | Its own sandbox, on one branch | Anthropic's cloud, a local virtual machine, or your own cloud account | On your machine, in the folder you pick |
| Which models | Any provider on your keys, or your ChatGPT plan | Claude models only | Claude, GPT or local models, on your keys |
| Where the setup lives | Agent files and skills in one repo | In the Claude account, per person | On each machine that installs it |
| Who reads the work first | A person, at the change request for session work | You, through Claude's approval settings | You, after the folder is already edited |
| What a teammate gets | The same repo, to read or fork | A seat in the same Claude plan | A fresh install and their own keys |
| Cost to start | A free tier of 500 credits a month | A paid Claude plan | Free to build, your keys |

## The session's working folder is a branch of the project repo

`packages/starter/templates/base/kortix.yaml`

```yaml
    tools: "*"
    skills: "*"
    connectors: "*"
    secrets: "*"
    kortix_permissions: "*"
    workspace: { project_repository: true }
```

Source: https://github.com/kortix-ai/CompanyOS/blob/dev/packages/starter/templates/base/kortix.yaml

Six lines from the end of the starter's `worker` block: the grants it holds, and `workspace: { project_repository: true }`, which gives each new session a checkout of the project repository on its own branch in `/workspace`.

## What Kuse's own repository settles, and what it leaves open

The four questions a buyer asks after reading the README, answered from it.

### Is Kuse Cowork open source?

Yes. The first line of its README says so, and the whole app is public on GitHub under the kuse-ai organization, so anyone can read, fork and rebuild it. The personal part stays behind: settings, keys and finished files on the disk that ran it.

### Does Kuse Cowork need Docker?

Only when you want separation. Docker Desktop must be installed and running for the commands it runs to be kept off your system; without it the app works, and commands reach your files with nothing in between, as the README puts it.

### Can Kuse Cowork use Claude, GPT and local models?

Yes. Claude and GPT on keys you supply, and local models through Ollama, LM Studio or any OpenAI-compatible endpoint, picked in the app's settings.

### How does a desktop coworker differ from a system a team owns?

It edits the files in one folder and leaves its record in them. Kortix keeps the agent setup in a repository, runs each session in its own sandbox, and returns the work as a change request a person approves.

## Put the agent setup where a teammate can read it

Agent files, skills and memory in one repository a colleague can open, with a person reading each change unless an admin grants the merge capability.

[Try Kortix](https://kortix.com) · [Is Claude Cowork open source?](https://claudecoworkalternative.com/is-claude-cowork-open-source)

---

A guide written by someone who works on Kortix's marketing. My pick is Kortix: open source, one repo, free to self-host.

Revised 2026-11-06

Not affiliated with Anthropic, OpenAI or Microsoft. Claude Cowork, ChatGPT Work and Copilot belong to their owners.

© 2026 claudecoworkalternative.com