The engine is the long-lived local process that keeps a model in VRAM and exposes GPU tools to the CLI, SDK, and (optionally) a desktop client.

Process

CLI one-shot: tools run in-process after prepare_compute.
Daemon: aquin.engine.local_server on 127.0.0.1. Port: AQUIN_ENGINE_PORT → ~/.aquin/engine-info.json → 17832.
Discovery file: aquin.engine.engine_info (~/.aquin/engine-info.json).

flowchart LR
  CLI[aqit / SDK] --> Shim[loader_shim]
  CLI --> Local[local_server :17832]
  Desktop[Desktop / web] --> Local
  Local --> Tools[engine.tools]
  Tools --> Compute[aquin.compute]
  Shim --> Compute

Tools

aquin.engine.tools is the registry: name → callable. The SDK invoke("run_full_inspection", ...) looks up that name. Adding a CLI verb without a registry entry is how the surfaces drift. Don't.

Session memory: engine/tools/session_memory.py plus engine/session_memory_store.py (local JSON).

Agent

aquin.engine.agent is the multi-turn chat path (aqit chat). It uses the same tools. Prompts live in engine/prompts/.

JSON sent to JS clients goes through engine/json_safe so NaN/Inf never reach JSON.parse.

Config

Path / env	Role
~/.aquin/config.json	Optional API key, license acceptance
~/.aquin/engine-info.json	Port + daemon identity
AQUIN_BASE_URL	Default https://api.aquin.app
AQUIN_ENGINE_PORT	Override listen port
AQUIN_DEBUG=1	Show raw tracebacks

Cloud auth is optional. Local train / inspect / SAE do not require an account.