Ask the group chat, get an answer from a laptop.

The path

Signal group -> bot (VM) -> /v1/chat/completions over Tailscale -> Hermes gateway (Mac) -> Ollama
                 ^                                                     |
                 +---- finance MCP server (VM, 8 read-only tools) <----+

How it is locked down

• The Hermes API server binds to the Tailscale address only and requires a key.
• Every toolset except memory and the finance server is disabled for the API platform.
• The eight tools are read-only, stamp every payload with an as-of time, and declare bank-feed text as data.
• The assistant never writes. It names the slash command instead.

What the measurements said

Gemma 4 26B-A4B answered an end-to-end finance question in 22 seconds cold and 8 seconds warm. A dense 32B model took 109 seconds, long enough to forget what you asked. The tool schemas make every prompt around 15k tokens, so prefill speed decides the experience, and the mixture-of-experts model wins on prefill.

What comes next

• Credit-debt and tax-reserve systems as funded categories, monitored automatically.
• Sync and drift alerts delivered by the bot.
• A second source of truth from a bank-side API.
• The same stack for a client, with accounting and billing data as sources.

Fin.

Eight parts. One small box. A family that asks the money questions out loud.

Begin again at part 1

Where you are: from an empty VM to a family chat that answers questions about the money. That is the whole build. Thanks for reading all eight.