miii-cli

Cloud coding agents are useful right up until you point them at proprietary code, or the bill arrives. Then you are shipping your source to a third party, juggling an account per vendor, and watching a per-token meter run. And you are locked to whichever model that vendor sells.

miii-cli answers two questions at once: can a genuinely capable coding agent run with nothing leaving the machine, and can the same agent use any model you choose when you do want the cloud?

• Model-agnostic from the core. 14+ providers, Claude, GPT, Gemini, DeepSeek, Grok, Mistral, Kimi, GLM, Groq, OpenRouter, Cerebras, plus local runtimes (Ollama, LM Studio, llama.cpp, vLLM) for Qwen3 Coder, Devstral, gpt-oss and Llama. Bring your own key, or no key at all.
• Fix-until-green with auto-rollback. The agent runs your tests, keeps iterating until they pass, and checkpoints so a bad turn can be rewound instead of hand-reverted.
• `miii doctor` grades models on real agent tasks, so you pick a model by evidence, not by benchmark marketing.
• Permission-gated edits and commands with persistent approval rules, because an agent that can run bash needs a real consent model, not a yes-to-everything prompt.
• Path confinement and lossless output spill, so the agent stays inside the repo you handed it and never silently drops tool output.

It reads code, plans, writes features, runs tests, and verifies its own work, from an Ink terminal UI or a browser app (`miii web`). MCP servers, hooks, and subagents extend it the same way the commercial agents do. `miii provider add anthropic` wires up Claude in one line; skip it and everything runs 100% offline.

Apna code, apna model, apni marzi. Your code, your model, your choice.

It is the clearest statement of how I think about AI products: small surface, real safety model, no lock-in, no hand-waving about privacy. Fifty-six releases in, it is also proof I maintain what I ship.