Architecture

Published July 2026 · 6 min read

Why we run every client on a single-tenant VPS

When we started OmniPaws, the obvious architecture was the same one every AI SaaS uses: a shared multi-tenant cloud, a vector database everyone shares, and a single LLM API key. Cheap to run. Easy to demo. Catastrophic for the customers who actually have data they care about.

So we did the opposite. Every OmniPaws deployment runs on its own VPS — dedicated CPU, dedicated disk, dedicated LLM endpoint, dedicated firewall. There is no shared database. There is no shared object store. Your agent and your data live alone.

It sounds expensive, but the math is interesting. Modern Linux servers with LUKS full-disk encryption cost roughly $30–60/month for the kind of workloads an SMB actually runs. Add a managed LLM endpoint and you're still under $200/month of infrastructure — well below the $1,000+/month a single ops hire costs.

Single-tenant also turns a SOC 2 audit into a conversation instead of a fire drill. When the auditor asks 'where does customer X's data live', the answer is 'on this one server, encrypted, with these SSH keys', and you move on. Try answering that question on a shared multi-tenant cloud.

There are trade-offs — we can't auto-scale across customers the way a true SaaS can, and on-boarding takes a few hours longer because we provision a real machine, not a container. But for the customers we built OmniPaws for, those trade-offs are the entire point.

← Resources