The voice above.
Conversation, reasoning and explanation. Bring an MLX-format model or point the slot at any OpenAI-compatible endpoint.
Default · Qwen2.5-3B-Instruct-4bit
Replaceable · Local or remote
Replaceable · Local or remote
01 · The vessel
Two minds.
Two minds.
One creature.
Creature has two cognitive slots: conscious for reasoning, explanation and conversation; unconscious for coding and logic. Use local models, remote endpoints, or a mix. Calibration teaches the pair how to work together.
The work below.
Coding and logic. It can live beside the conscious model on your machine, elsewhere on your network, or behind an endpoint you control.
Default · Qwen2.5-Coder-3B-Instruct-4bit
Replaceable · Local or remote
Replaceable · Local or remote
Calibration handshake
02 · Proof
Green has to
Green has to
be earned.
Creature does not treat recognition as proof. A file becomes green only when its checker runs successfully at a labelled depth. A missing tool is UNKNOWN—not a quiet pass and not a borrowed result from another language.
Colour is the consequence, never the claim.
Every checker recipe records its condition, strength and source. Compatibility is not transitive: one language never borrows another language’s green.
Pass
The checker ran.The file earned green at the recorded proof depth.
Fail
The checker objected.The failure remains visible and local to the invocation.
Unknown
No proof was inferred.The checker is absent, unsupported, or could not establish a result.
03 · Forty-three languages
Does it
Does it
hold?
Creature catalogues exactly 43 language checker recipes. Swift and Python receive native AST-quality indexing; the other languages use the shared structural indexer while their installed checker determines the proof depth they can earn.
Reading the canonical capability manifest…
Recipe ≠ installed tool ≠ live pass
No capability claim has been inferred. The generated manifest is required.
Green-compatible means Creature knows the command and conditions required to test the language. The checker still has to be installed, and every file still has to earn its own result.
04 · Where it runs
Run where
Run where
you choose.
Local-first is a choice Creature can prove, not an absolute network promise. In-process models stay on your machine; remote slots intentionally use the endpoint you configure.
In process.
MLX model inference runs inside Creature itself—no Ollama, LM Studio or separate model server required.
Build · chat · index verifiedBring the address.
Either cognitive slot can connect to an OpenAI-compatible endpoint, including local servers or hosted services you select.
Local · remote · mixedContainer proven.
Creature builds and runs on Linux, where it drives a local Ollama instance instead of performing MLX inference itself.
Build · run verifiedBarely tested.
CI proves the binary compiles and starts. Chat, indexing and the full test suite are not yet established there.
Compile · start only
05 · Pre-v1
Early enough
Early enough
to be honest.
Creature is free, open, and useful today. It is also rough. If you try it now, you are a tester rather than a customer—and the failures you report are part of the work.
Works today
- Runs its own MLX model in-process on Apple Silicon.
- Chats and indexes a codebase across 43 catalogued languages.
- Checks workspaces with explicit PASS, FAIL and UNKNOWN states.
- Builds and runs on Linux with a local Ollama backend.
Still rough
- The IDE editor is read-only and has no syntax highlighting.
- The macOS app is not notarised or signed with an Apple Developer ID.
- The install paths are young and have already exposed packaging failures.
- Windows evidence stops at compile and binary start.
06 · Open creature
Open code. Open rough edges.
Creature is licensed under AGPLv3: free to use, modify and share, with the same openness required when a modified version is run as a network service. Contributions are welcome. So are precise bug reports from people willing to find the edges.
The floor · 10,911 m
Bring your own soul.
Choose the models. Choose where they run. Keep the proof attached to the claim.