Two CLIs over a local llama.cpp stack. `llm` sets up and runs the inference server: it picks a GGUF, sizes the context against the card's real VRAM -- sliding-window, recurrent and MTP draft caches included -- builds llama.cpp, ik_llama.cpp or PrismML's llama.cpp fork when no binary exists, and pushes the active model into whatever agent configs already exist. `mimic` launches Claude Code or Codex against that server through their own documented provider settings, each in an isolated config dir, so a local model answers instead of a hosted one. Linux only, and it downloads no models.

Required Ruby Version

>= 3.2

Authors

raykin

Versions

  1. 0.1.0
    1. source

      1. 0.1.0 source November 16, 2026 (95.5 KB)

Pushed by

SHA 256 checksum