It rains in
every
language

Bring your own soul.

Creature runs language models — yours. It runs them in-process via MLX on Apple Silicon, so the model lives inside the binary rather than behind a local server. No Ollama, no LM Studio, no daemon to babysit. Or point it at any OpenAI-compatible endpoint and use whatever you already pay for.

With in-process models, nothing leaves your machine. Not the code, not the question, not a telemetry ping. That isn't a privacy policy — it's an architecture. There's no network call to opt out of.

A mind that sleeps on it.

Creature has two cognitive slots, and you fill both. The conscious slot reasons, explains and converses. The unconscious slot does code and logic. They're different models, chosen separately, and creature calibrate runs a handshake that syncs whatever pair you plugged in.

Which is why "what's the best model" is the wrong question here. Put a big remote reasoner in the conscious slot and a small fast local coder underneath it. Or run both local and go offline entirely, on a plane, forever. The defaults below are Apache-2.0 and arrive on first use.

Grey is not green.

Every other tool tells you your code is fine by failing to mention that it isn't. Creature can't. Green is granted only by the configured checker's affirmative exit 0, and it says exactly how deep that proof goes: parse, syntax, lint, or typecheck. A syntax green never masquerades as a typecheck green. A missing checker grants nothing.

So there are three colours, and the third one matters most. Unknown is a real state that stays on the screen. Press the probe: nothing turns green until its checker says so, and a broken leaf takes its parents down with it.

Green can't be worn bare, either. Internally every green carries the triple ⟨grammar · unit · condition · proof strength⟩ — which grammar vouched for it, over what unit, under what condition, and whether the result was parsing, syntax, lint, or typecheck. A green with nothing behind it is unrepresentable.

One meaning, many grammars.

Look up. The storm falling through this page is the same handful of human meanings — language, understand, hello — in fifty tongues at once, and not one of them is the original. That's the whole idea, and it's older than software: a meaning is not the words you happened to say it in.

Code is the same. A for loop in Rust and a for loop in Elixir are related thoughts in two grammars. Creature carries 43 of them offline — each with a sourced checker recipe and an explicit proof depth — and treats none as the centre.

Every grammar, its own verdict.

A Swift add and a Python add may share a structural shape. That makes them candidates, not equivalents. Neither may borrow the other's green. A cross-language link exists only when one current, unambiguous translation path proves the relevant facet — calls, types, declarations, control flow — under the active conditions.

Creature keeps each native fact immutable, records where every translation rule came from, composes indirect conversions through a versioned dialect graph, and invalidates stale recipes. Ambiguous or missing paths stay unresolved. The prior art is in the mechanism; the evidence remains Creature's own.