AI coding tools have changed how software gets written. The next wave of agent harnesses are not toys — they are the way a generation of developers will build for the next decade. The productivity gap between someone with these tools and someone without them is not subtle. It is the difference between shipping a feature this afternoon and shipping it next week.
And yet entire industries are sitting it out.
Banks. Hospitals. Defense contractors. Government agencies. Law firms. Critical infrastructure. R&D teams sitting on patents. Anywhere the rule is "your data does not leave the perimeter," the developer doesn't get the agent. They watch peers at unregulated companies move twice as fast, and their CTO and CISO have a difficult conversation about it once a quarter.
The standard answer is "we'll have an enterprise tier eventually." But the enterprise tier of a hosted agent is still a hosted agent. The data still leaves. The records still live in someone else's database. The model still belongs to someone else. Adding a certification to the bottom of the page does not change where the bytes go.
The wedge
So we are building Lumium, and the wedge is small and sharp on purpose.
Lumium is a workspace that runs on your machine, on your servers, or inside your self-hosted environment. Projects hold conversations, artifacts, work items, knowledge, and the agents that work alongside your people. It uses the model providers you choose, including local ones. Every action it takes lands in a ledger on storage you control. The console and the runtime are one process; read every line if you want, fork it if you must. We do not phone home. We do not proxy your prompts. We do not store your provider keys on infrastructure we operate.
The product is not better because we wrote a smarter agent loop than the cloud vendors. The agent loop is good — we put a lot of work into it — but that's not why this matters. The product is better for the people who currently have nothing because it ships where they are: behind their firewall, inside their VPC, on their laptop, on a sealed network with the cable cut. The deployment shape is the differentiation.
Three things we will not compromise on
First: self-hosted by default. The console and the runtime are one Node process. State lives in the database and object store you already back up. Register a model served on your own network and nothing needs to leave the building. This is not a feature flag. It is the deployment shape.
Second: model-neutral. Bring OpenAI when you have approval to use it. Bring Anthropic if you prefer. Bring Cohere, or Gemini, or Mistral, or Llama running on a box in your closet. The agent loop does not care, the policy does not care, and the audit log records the choice. You should not have to pick a vendor to pick a tool.
Third: every action is recorded. Every provider call, every tool invocation, every policy decision, every file touched. Append-only, in the database you already back up, with a sequence that makes tampering visible. The ability to ask "what did the agent actually do" is not a paid add-on. It is the price of running an agent where someone is responsible.
What this means for the people we are building for
If you are a developer at one of the companies that hasn't been able to use any of this: you should not have to wait. There is a category of work that requires the context of your repository, the output of your test suite, the contents of your internal docs, and the operations of your infrastructure. It will get done in three hours instead of three days, and you do not have to send any of it to a vendor.
If you are a security or compliance leader: you should not have to keep saying "no." The architecture you have been waiting for — agents that run inside your boundary, with records you can read line by line — is shipping. The hard work of saying "yes, with these controls" should be available to you.
If you are a CTO trying to bridge those two: there is a path that does not require choosing between productivity and the policies that keep you in business.
What this is not
Lumium is not for everyone. SaaS coding agents are excellent at what they do and the developers who can use it should keep using it. Lumium is the runtime underneath the agent for the developers who can't.
Lumium is not a model. We do not train foundation models. We are building the kernel that lets the model do useful work in the places models alone cannot reach.
Lumium is not yet finished. Today's release is v0.1. The console is real, the runtime is real, the ledger is real, and the agent loop is covered by tests. We have miles to go on billing, a public CLI and SDK, outbound chat delivery, and a wider connector set. The roadmap is public. The brand is honest.
Where to start
Clone the repository onto a machine you control and start a conversation against a model on your own network. Watch the ledger fill up in real time and notice that nothing left the box. Then think about what your team could do if every project had that, every day, on the data they are not allowed to send anywhere else.
That is what we are building. That is who we are building it for.
— The Lumium team