Local agent runtime for Gemma 4 on llama.cpp. Unsandboxed by design — code_run is real Python; shell runs against an allowlist. You are the watchdog. Ships a working jinja template for Gemma 4 (llama.cpp has no built-in one yet) and a GBNF grammar that enforces tool calls at the logit level. 39 tools — semantic memory, browser automation, cron scheduler, and hot-reload skills (the model writes its own). Runs on 7GB VRAM. MIT, plain Python, no framework dependency.