knowledge = data + memory

We already pay for Claude subscriptions at Hedgy, so I built a company brain that doesn't spend extra tokens. Setoku is a self-hosted MCP server (a ClickHouse data lake plus a knowledge layer about that data) that gives your AI the data it needs to understand your business/life. No inference happens in the box itself, all the thinking runs in the AI you already pay for, so the total bill is a cheap VPS and there's no new vendor holding a copy of your data.

Ours ingests spending and account data from Mercury, Vercel logs, Render logs, a few Slack channels, and GitHub activity. I honestly hooked up the log drains just to test with more data, but the extra context made agentic debugging faster and more reliable and now I couldn't go back.

The knowledge layer is the part I find more interesting. Anyone can point Claude Code at a Postgres, that's exactly how this started. The problem is every session rediscovers the same tribal knowledge, stuff like "this is how we count a user as active" or "check column X when computing LTV". The MCP nudges the AI to ask about and record gotchas as it finds them so the knowledge builds up instead of getting re-derived every time. Last week "how many hires have we made" turned out to have three different answers depending on how you count (some placements fell through, our old dashboard silently dropped a few real ones). Claude worked through it with my cofounder once, recorded the right definition, and now every session gets the same number. There's a small admin UI for auditing and pruning what it's learned.

We also added a protocol so an LLM can publish a chart or dashboard to the box as an app hooked up to live data. The part I didn't expect: the non-technical folks push apps the hardest. claude.ai + Setoku is the only coding environment they have, and they've been shipping internal tools I never would have thought to build.

Ships as a Docker image plus a set of Claude Code skills that cover server setup and adding connectors. The MCP works with whatever agent you use, setup is just Claude-shaped right now. Our company instance runs on an OVHCloud VPS-3. I also run a family instance on a $5/mo VPS-1 that ingests Gmail and finances from Monarch Money so that Fable can give me grounded financial advice (this is not financial advice!). A fresh setup takes about 20 minutes.

Site: https://setoku.com

Code: https://github.com/Hedgy-Labs/setoku

Happy to answer questions or help anyone get an instance running. Thanks for checking it out!

-peter