Alexa — AI skill for Claude Code

Alexa

A local agent that lives on one Windows desktop, runs on a single 6 GB graphics card, and treats the internet as a door it has to unlock — not a place it sends your work.

**Alexa is what it answers to; `sunday` is what it is made of.** The window title, the executable and the installer say Alexa, because that is the name a person says out loud, and it comes from `[assistant] name` in the config. The repository, the Python package, the sidecar process and `%LOCALAPPDATA%\Sunday` say sunday, deliberately: renaming the data directory would orphan the memory store without saying so, leaving a fresh empty one beside it that looks like it is working. So every command below is `sunday`-something, and that is not a leftover.

What this actually is


Not a voice assistant with a chat box bolted on. It is the same shape as the agent runtimes people have started keeping open beside their editor — a model in a loop that decides which tool to call, calls it, reads what came back and decides again, up to twelve calls in one turn, with memory read before it starts and written after it finishes. Voice is one of two ways in. Typing is the other, and neither is a mode you switch between.

What is different is where it runs. **The thinking is local**: one small model, `qwen3.5:4b` through Ollama, on your own card. No prompt, no file, no calendar entry and no piece of mail is sent anywhere. There is exactly one remote call in the whole program — DeepSeek summarising a page the assistant fetched — and it sits behind an airlock that composes the outgoing query in a context your private data was never allowed into. That is the design, not a setting:

A local model can use your private things and the public web in the same breath, because the query that goes outside is composed in a room your private things were never allowed into.

Reading anything credential-shaped shuts the