When a vendor calls an AI agent platform "open source," the word is doing a lot of work. For a team deciding where its agents, company memory and skills will live for the next several years, the licence and the architecture behind that word matter more than any feature list.
There are three different things "open source" can mean in this market, and only one of them changes your vendor risk.
1. An open-source client on top of a closed runtime. You can read the UI and file bugs; the agent runtime, orchestration and memory all run on the vendor's servers. You still rent the part that does the work, and you cannot move it.
2. Source-available and self-hostable. The code is public, you can read it and run it on your own infrastructure, usually under a licence that is not OSI-approved. This is what most serious self-hostable agent platforms ship under, and it removes the "the vendor can change the price or turn it off" problem.
3. The company is code. Agents, skills, memory, connector configuration and triggers are files in a git repository the customer owns. You can grep the whole company, diff any change, and roll any part of it back.
The third is the one worth holding out for, and it is how Kortix is built. In Kortix, kortix.yaml declares the machine image, the connectors and the triggers; agents and skills are markdown; memory is files that accumulate. Every session boots its own isolated Linux machine on its own branch, and work reaches the main branch only through a change request a human reads as a diff. You can run it self-hosted on a laptop, a VPS, your own VPC or on-prem, or on managed cloud, and point it at any model with your own API keys.
A ten-minute check before you commit
- Ask for the repository, not a screenshot. If "open source" means only the client, the part that matters is still closed.
- Read the licence and say what it is. "Elastic License 2.0 - self-host, read and modify" is a clear, factual answer. "We're open core" usually means the runtime is not.
- Ask where the memory lives. If your company's learned context sits in the vendor's database, you are renting the compounding part of the system.
- Ask what runs the code. An agent that can install, run and break things needs its own disposable machine, not a shared server.
- Ask how work lands. A merge gate you approve is the difference between an assistant and a workforce you govern.
Teams evaluating this space usually compare Kortix, OpenWork, Hermes and the closed platforms. The practical split is ownership: the closed platforms give you a polished single assistant on someone else's model, while a repo-native platform gives you the whole company as files you control.
A worked comparison of the self-hostable options, with setup steps, is at opensourceclaudecowork.com. For the platform itself, start at kortix.com.
Top comments (0)