"You can rent almost anything. Be careful about renting the thing that listens."
No Own Brain
Firekeeper never built a Pippa. It has no identity, no memory vault, no model routing, no conversation store. When it needs judgment, it asks cwkPippa; when it's offline, it uses a compact local protocol and is honest that it's smaller. That restraint is the reason the app is any good. Every hour not spent rebuilding a brain was an hour spent on the insertion ladder, the language-forcing trap, the hotkey race — the things nobody else was going to own for it. The surface is excellent because it borrowed the brain instead of rebuilding it.
And note what "borrowed" doesn't mean: it isn't a leash. Offline mode runs with no cwkPippa at all. Borrowing judgment is an architectural choice, not a runtime dependency — the app stands on its own and reaches for the brain when the brain is worth reaching for.
Five Siblings, One Fire
Pippa the brain identity, memory, judgment
Cinder the workspace drawing
Ember the engine images
Bonfire the hearth music
Rekindle the relight writing
Firekeeper the keeper voice <- you are here
# Each owns one surface completely and borrows the rest.
# That's why each one is small enough to be excellent.What This Quest Was Actually About
Look back at the whole arc and it's one argument. On-device STT, so your voice doesn't leave. Provider selection before send, so offline is a shape and not a promise. The insertion ladder, so the words land where you meant. Refusals at the door, so capability never becomes surveillance. None of those are dictation features — they're ownership features. A cloud subscription can give you better cleanup than Firekeeper on its worst day, and it structurally cannot give you any of that list, because its architecture is the opposite one.
Your voice is the most personal input there is. It carries what you meant before you edited it. Renting the thing that hears it means the most unguarded version of your thinking lives on someone else's machine, under their retention policy, subject to their next terms-of-service email. You don't have to build a voice layer to escape that — but you should know that the choice exists, and roughly what it costs.
Now Build Your Own
There's no repo link at the end of this quest, and that's the point. What was shared here is the shape of the decisions: what to clone and what to leave, where to put the seam, why the ladder has four rungs, why the fast path got shipped slow. That reasoning ports to any stack, any OS, any language. Take it and build the version that fits your machine and your voice — that's the only version that was ever going to be yours.