The default playbook for developer tools in 2026 is obvious: build a hosted service, gate it with a free tier, convert on usage. It's a great model — for the vendor. For the team running it, every request goes through someone else's infrastructure, someone else's compliance story, someone else's pricing page.
We went the other way. KB Labs is self-hosted first — you install it with npm, it runs in your infrastructure, and the data never leaves your network.
The three reasons, in order of importance
1. Trust is cheaper to earn when you give up control
Enterprise sales cycles exist because companies need to verify that a vendor won't mishandle their data. Self-hosted removes the question entirely. There's nothing to verify — your code, your servers, your logs. The security review is "read the source."
This matters especially for AI tooling. LLM calls carry proprietary context — codebases, business logic, internal documentation. Asking teams to send that through a third-party hosted service is a hard sell to any security team. Self-hosted means the LLM adapter routes to whichever provider you choose, through your gateway, logged in your telemetry.
2. Operational independence
When a SaaS goes down, your CI goes down. When a SaaS changes pricing, your budget changes. When a SaaS sunsets a feature, your workflow breaks. These aren't hypothetical — they're Tuesday.
Self-hosted means your platform availability is your problem — which sounds worse until you realize it's also your solution. You can debug it, fix it, and deploy the fix without filing a support ticket.
3. Adapters make self-hosted viable
The usual objection: "self-hosted means you maintain the infrastructure." True — but KB Labs is built on contract-based adapters. The platform doesn't care whether your cache is Redis or in-memory, whether your LLM is OpenAI or a local model, whether your vector store is Qdrant or Chroma.
This means a team of two can run KB Labs with in-memory adapters on a single box, and a team of two hundred can run it with Redis, Postgres, and a GPU cluster — same codebase, same config shape, different adapter lines.
What about the hosted version?
We'll ship one. But it will always be "the same thing, hosted for convenience" — not "the real product with a self-hosted afterthought." The open-source core is the product. Hosting is a service on top.
If the hosted version disappeared tomorrow, every user would still have a working platform. That's the test.