The self-hosting part is genuinely solved. OpenHands runs on the machine in front of you, in Docker, or in your own cloud, and a developer can have it working in an afternoon. If your only requirement is "the coding agent runs on hardware I control", you can stop reading and go install it.
The questions start when the work leaves the repository.
What a coding agent is built to do
A coding agent is very good at one loop: read a codebase, change it, show you the diff, take your correction. Everything it needs is already in the repo and the conversation.
The moment you ask the same agent to answer a question in Slack while you are asleep, pull a number out of a billing system, remember what the company decided three months ago, and let a colleague who is not an engineer run the same task without rebuilding it, you have left the loop the harness was designed for. Not because it is badly built. A single workspace plus a conversation is simply the wrong shape for work that has to be shared, scheduled and governed.
Four questions that decide it
Where do the agents live? If each person configures their own, the company learns nothing. If agents, skills and memory are files in one repository, what one person teaches is there for everyone on the next pull.
Who can act, and on what? An agent that can send mail, move money or change a customer record needs a permission model that is written down, versioned, and enforced somewhere other than a checkbox in one person's app.
What does it run on? One agent on your laptop is easy. A hundred of them at once, each on its own machine, is an infrastructure question.
Can you change the model? If the agent only runs one vendor's models at that vendor's prices, the cost of your own automation is set by somebody else.
Where Kortix fits
Kortix is the open-source AI Operating System. It answers those four questions by making everything a file in a git repository you own:
- Agents and skills are markdown. Company memory is plain files. Connectors, triggers and permissions live in
kortix.yaml. - Every session gets its own isolated machine and its own branch. Its work reaches
mainthrough a change request a person reads as a diff, and merge is default-deny for agents. - Bring your own keys for any model provider, or use the ChatGPT plan you already pay for.
- Connector credentials are brokered server-side and never enter the machine the agent is running on.
- Self-host is free on a laptop, a VPS, in your VPC or on-prem. There is also a managed cloud if you would rather not run it.
The honest split
Keep the coding harness for repositories. That is what it is for, and it is good at it. Run the company's other work on something that was built for many agents, many people and a paper trail. Most teams end up with both, and that is a reasonable place to land.
- Full comparison: OpenHands vs Kortix
- The other routes, side by side: OpenHands alternatives
- Kortix: kortix.com and Kortix on GitHub
Top comments (0)