Is Claude Cowork open source? How to self-host an open-source stack you own

There is an open source answer to Claude Cowork, and it is Kortix: the open-source AI Management System you self-host, with any model and your own keys. Every agent, skill, memory file and connector lives in one git repo you own. Anthropic's own Cowork is closed source, runs on Claude models only, and cannot be hosted on your own infrastructure.

What an open-source Cowork actually means

Claude Cowork is Anthropic's agentic workspace. Give it a goal and it works across your files and connected tools, then hands back finished work for your review (Claude Cowork). It is a strong product for one person's desk, and it is not a thing you can fork.

An open-source Claude Cowork is a different object. It is a self-hosted agent platform that gives a company the same loop, an agent that plans, calls tools and returns finished work, while the code, the models and the machines underneath are yours. For that to hold, the source has to be readable and changeable, the model has to be your choice with your keys, sessions have to run on infrastructure you control, and a person has to read the diff before work reaches the main branch.

Several open-source platforms meet those conditions, each with a different center of gravity. Kortix is the one we built for a whole company, and it covers the loop end to end: one repo, an isolated machine per session, and a change request between an agent's work and the main branch.

OpenHands runs coding agents from a self-hosted control center that can point at local, Docker or VM backends (OpenHands). Open WebUI is a self-hosted AI platform that runs offline and adds an agent runtime with a real terminal and filesystem (Open WebUI). Dify builds agents, agentic workflows and chatbots you can publish as apps, with a Community Edition you run on your own infrastructure (Dify).

Where Claude Cowork stops for a self-hoster

Cowork is closed, and it runs where Anthropic runs it. Four facts answer the question, all from Anthropic's own pages.

• There is no source. Anthropic sells Cowork as a product on its Pro, Max, Team and Enterprise plans (product page); it publishes no repository, licence or self-host option.
• The models are Anthropic's. Cowork is Claude, with no bring-your-own-key or per-agent model choice (Cowork docs).
• The agent runs on Anthropic's infrastructure. Sessions run in the cloud by default, with the agent loop and code execution in a temporary sandbox on Anthropic-managed servers; the local option is the Claude Desktop app on a member's device (architecture overview).
• Your cloud account routes inference, not hosting. You can use your own Amazon Bedrock, Google Cloud or Microsoft Foundry account, which changes where the model calls run, but the platform stays Anthropic's app (product page).

None of that is a fault. Cowork is built to be one polished product for one person's work, and it is good at that. It also means there is no Cowork you can install on a server you control.

Claude Cowork versus an open-source stack

The two paths are built for different owners. One is a product you subscribe to; the other is a stack you run.

The self-hosting checklist

Running a Cowork-style stack on your own infrastructure comes down to four setup steps.

• Take the repo. Install the CLI and scaffold a project, and kortix init creates kortix.yaml plus your agents/, skills/ and memory files. The repo is the company.
• Point it at a model. Add an API key from any provider, sign in with a subscription you already pay for, or enter your own OpenAI-compatible endpoint.
• Start it on your own machine. One Docker Compose stack runs on a laptop, a VPS, your VPC or on-prem, or you use managed cloud instead. kortix self-host start brings the box up (Kortix on GitHub).
• Turn on the gate. Merge is default-deny for agents, so every session opens a change request you read before it lands. Approval gates are off until you set them.

Kortix is open source (Elastic License 2.0): self-host, read and modify the code. The same install gives your agents 3,000+ apps through one scoped token, plus MCP, OpenAPI, GraphQL and raw HTTP, with connector credentials kept server-side.

Getting started

The fastest path runs on Kortix Cloud, and the same commands build the repo you would self-host later.

curl -fsSL https://kortix.com/install | bash   # install the CLI
kortix init                                    # scaffold kortix.yaml, agents/ and skills/
kortix ship                                    # push the repo and bring it live
kortix cr ls                                   # review what an agent proposes; merge to keep

Every session boots its own isolated Linux machine with your repo and tools already on it, and runs the agent on its own branch. When the work is done, the agent opens a change request, and you read the diff and merge to keep it. Setup details are in the docs at kortix.com/docs.

If you want the company platform, read how it compares with a single desktop assistant at Kortix vs Claude Cowork, or where a personal agent stops and a company platform begins at personal AI agents vs a company OS.

Own the whole thing, from the repo to the machine each session runs on: Get started with open-source Kortix.