Monet is a studio named for the painter. Impressionists built a picture from many small strokes; a language model builds an answer from many small tokens. We make the instruments for working with models, and next, the art made with them.
The five instruments sit on the path of a single model call, in the order a request meets them. Each works on its own. In every mark, one stroke is painted red: it is the one thing that instrument decides or protects.
ButterflyCoding harness
A coding agent built around the prompt cache. It keeps cache hits high, compacts its context rarely, and checks its own work before it calls a task done.
CocoonCredential firewall
Keeps provider keys out of the agent. The agent holds a session token; the real key is attached only after the request passes policy, so the secret never sits in the agent's context.
NimbleModel router
One OpenAI-compatible endpoint that chooses the model and provider for each request, and remembers enough about a session to keep its prompt cache warm.
InverLLM gateway
A prepaid, OpenAI-compatible gateway. Every workload gets its own key, and every request settles to an exact cost in a ledger that never reprices the past.
CicadaToken ledger
One command that reads what is left on every model API account you use, all at once, instead of one billing console per provider.
The canvas
The instruments come first so the art can follow: work where the model holds the brush and every stroke is a token. The painting at the top of this page is a first sketch of that idea. Nothing is hung yet.
To talk about any of this, write to tio@monet.sh.