miii-cli

About

miii exists for engineers whose code cannot leave the machine: it talks only to local inference servers — Ollama by default, plus any OpenAI-compatible local endpoint like llama.cpp or LM Studio — so it works offline with no accounts and no per-token billing. The loop plans a task, acts through six permission-gated tools (read, write, edit, glob, grep, bash) confined to the workspace, and observes results before proceeding, with ~/.miii/permissions.json persisting approvals across sessions. Two implementation choices stand out: oversized tool output is written to disk and paged through the model rather than truncated, so long build logs survive intact, and MIII.md plays the role of CLAUDE.md as a per-repo conventions file read every turn. Model choice is hardware-bound — qwen2.5-coder 7b for 8 GB VRAM up to 32b at 48 GB+ — and miii doctor grades installed candidates on real tasks rather than trusting benchmarks. Privacy-sensitive codebases and offline environments are the audience; the project is a one-maintainer MVP at small star count but actively developed.