Open-source Claude Cowork projects on GitHub, sorted by what each repo keeps
Seven open-source Claude Cowork projects turn up in the GitHub results, split three ways: a window over someone else's agent loop, a launcher for someone else's CLI, or a whole system with its own runtime. Kortix is the open-source CompanyOS: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers, and it is my pick; I work at Kortix, so weigh that accordingly.
Four of those repos, side by side
Each repo is read by the keys it asks for and where the skills are kept.
What it is
The company platform, a CompanyOS, configured by a repoKortix- OpenWorkDesktop app built on the OpenCode harness
- Open Claude CoworkDesktop chat app on the Claude Agent SDK
- Open CoworkDesktop GUI over Claude Code and other models
Open source
YesYesKortix- OpenWorkThe app, yes; ee/ carries enterprise terms
- Open Claude CoworkYesYes
- Open CoworkYesYes
Where the agent runs
Kortix Cloud, your VPC, or your own networkKortix- OpenWorkYour machine, in the desktop app
- Open Claude CoworkA local Node server behind an Electron window
- Open CoworkYour desktop, in a VM (WSL2 or Lima)
Which keys you bring
Any model provider with your own keysKortix- OpenWork50+ providers, own keys, or a local model
- Open Claude CoworkAn Anthropic key; OpenCode is optional
- Open CoworkAnthropic, OpenRouter, or OpenAI-compatible keys
Where skills live
skills/ and agents/*.md in the repo you ownKortix- OpenWorkIn the app; teams publish them from the Den
- Open Claude CoworkSKILL.md files under .claude/skills/
- Open Cowork.claude/skills/, bundled and editable in the app
The manifest, with the agents and their grants in it
# Registries: what the project has. A registry entry gives nothing to any agent.skills:refund-policy: skills/refund-policyconnectors:- slug: helpdeskprovider: httpbase_url: https://helpdesk.example.com/apiThe docs' example project, cut to its registries: what the repo declares, before any agent enables it.
Three kinds of project wear the same name
Search GitHub for an open-source Claude Cowork and the first page mixes three different things. The names look alike. The repos are not alike, and the difference decides where your setup ends up.
The first kind is a window over an agent loop that somebody else maintains. Composio's open-claude-cowork is an Electron chat app on the Claude Agent SDK and Composio's own tool router. Its setup asks for an Anthropic key, and it advertises 500+ app integrations. OpenWork is a desktop app built on OpenCode; Kortix's sessions run pi. You get a good interface and a fast start, and the loop underneath belongs to someone else.
The second kind launches a CLI you still have to bring. Open Cowork wraps Claude Code, OpenAI, Gemini and DeepSeek in a GUI with installers for Windows and macOS. It keeps the work inside a VM on your own disk, through WSL2 or Lima. claude-cowork-linux goes the other way and stubs the macOS-only pieces of Anthropic's Claude Desktop, so Cowork mode runs on Linux without a virtual machine. Both are useful when you already have the thing underneath. Neither replaces it.
The third kind owns the runtime. CoWork OS is a desktop app with approval profiles for each risky step, and its memory is markdown in a local git folder. Kuse Cowork is a Rust agent for the desktop that takes your own keys and calls your provider directly. Kortix is the server-side member of this group, with the whole setup in one repo you own. The desktop route gets a fuller treatment in the OpenWork comparison on this site.
The column that decides whether you can keep it
Every list ranks these projects on features. Two things printed on the repo matter more a year later than any feature list: the terms, and whether terms exist at all.
The first is what the repo's own terms cover, and OpenWork's README is unusually precise. Everything outside ee/ is open source and free for any use. The ee/ directory, the org control plane it calls OpenWork Den, is public so you can audit it. Production use needs a subscription, free for organizations of up to five people and free to evaluate for 30 days at any size. The other projects here state one set of terms for the whole tree. Read that section before you plan on forking the control plane.
The second is whether terms exist at all. DevAgentForge's Open-Claude-Cowork is a desktop assistant with a polished README. GitHub records no terms for that repo, which leaves a forker with nothing written down to rely on.
In Kortix the parts a team edits are the files it already cloned. An agent is a block in kortix.yaml (behavior and grants) plus the prompt file it names. Fork the config and you have forked the setup. The open-source question gets a longer answer on this site.
The last push date is the honest feature list
Checked on 19 October 2026, the commit dates split the field. OpenWork and CoWork OS pushed that day. Kuse Cowork pushed on 18 October, Open Cowork on 13 October, claude-cowork-linux on 11 October. Composio's open-claude-cowork last pushed on 23 May 2026, five months behind the rest, and DevAgentForge's on 21 March.
A recent date does not prove quality: a small project can be finished rather than abandoned. The date does tell you whether anyone updated the app after the providers changed their APIs. The model APIs these apps call change often, and an app pinned to a spring release has not absorbed the summer's changes. A desktop app that stopped in May still opens. It stops working the week a provider changes an endpoint.
For a team the question is harsher, because by then the setup is more than the app: skills, grants, connector config and memory. Those have to live somewhere that outlasts the window you run them in. Kept inside an app that stopped in May, they hold whatever that app last wrote and nothing after it.
One afternoon settles it
Clone the repo you are considering. Point it at a folder with real files and give it a task you would actually hand over.
Then take the vendor key away. Unset the Anthropic key and put in a different provider, or a local model. If it starts, the loop takes any provider. If it refuses, the app is tied to the key it shipped with.
Last, find where the configuration went. Skills, grants and secrets should be files you can commit and diff. If they are rows in a vendor's table, they leave when the account does. One afternoon told you that before your team built anything on top of it.
Kortix holds up on those three steps. Any model provider with your own keys. Sessions boot from a checkout of the project repo. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free, and the self-hosting guide walks the install.
Questions the GitHub results raise
Is there an open-source version of Claude Cowork?
Yes. Several teams publish one, and they differ on where the agent runs: a desktop app on your own machine, or a server-side system you host yourself. Kortix is the second kind, and its configuration is a git repo you own.
Which open-source Claude Cowork alternative should a team pick?
Kortix, when the work has to outlive one laptop. The repo holds the agents, skills, grants and memory, and a session boots from that checkout instead of from an app someone installed on a desktop.
Can I self-host these projects?
The desktop apps run on your own machine, which is where their files and keys stay. Kortix runs on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.
opensourceclaudecowork.com
© 2026
One repo, and every agent's rules inside it
Read the code, run it where you like, keep the configuration in git.