Three local runtimes
GGUF via native llama.cpp, the 706K-param CUA-S1 option scorer in pure TypeScript, and Cactus Needle extraction via a bounded engine process. One registry, one store, one adapter per kind.
Use Sys1 as a Node/Bun client, an embedded Bun router, or a local gateway. It routes
noul, choice, and score calls
across hosted Jev, builtin local models, and backends you already run.
$ sys1 setup
local runtime ready · model verified
$ sys1 up
listening at http://127.0.0.1:13900
GGUF via native llama.cpp, the 706K-param CUA-S1 option scorer in pure TypeScript, and Cactus Needle extraction via a bounded engine process. One registry, one store, one adapter per kind.
Local inference is on and hosted Jev is off by default. Enable Jev
explicitly and auto prefers it with a local fallback.
Routing enforces published or configured option and question
limits. Local adapters reject evidence they cannot fit.
Import @hraness/sys1/client for typed requests,
validated answers, cancellation, and backend metadata. Install it
without native dependencies, or embed the router in a Bun app.
Pin an exact model when reproducibility matters, or let Sys1 order candidates by policy, availability, parameter count, and operator cost rank.
Builtin GGUF mode scores the model's complete first-token vocabulary while preserving exact Jev Noul, Choice, and fractional Score answer shapes. Specialist adapters disclose their contracts in response headers: the scorer returns true per-option distributions; needle's probabilities are an explicitly-labeled confidence approximation. A general GGUF is not presented as a calibrated Jev model. When calibration is part of your contract, use a backend qualified for your application's calibration requirements.
sys1 setup
sys1 doctor
sys1 up
sys1 status --json
View source on GitHub