The open-source, offline alternative to Claude Code, Cursor, and GitHub Copilot.
A private AI pair programmer that runs on your machine with Ollama — no API keys, no cloud.
Private by default. Free forever. Works offline.

🔒 100% local — your code never leaves your machine  ·  💸 Free — no API keys, no per-token billing  ·  ⚡ Offline — runs on your own GPU

ollama pull qwen2.5-coder:14b   # any coding model works
curl -fsSL https://raw.githubusercontent.com/maruakshay/miii-cli/main/install.sh | sh
miii

Windows: irm https://raw.githubusercontent.com/maruakshay/miii-cli/main/install.ps1 | iex  ·  any platform: npm i -g miii-agent  ·  needs Node ≥ 18 and Ollama.

> refactor the auth module to use async/await
> @src/server.ts add rate limiting to all POST routes
> why are my tests failing in utils/parser.ts

miii reads your files, writes the code, runs your tests, and fixes what breaks — planning before it acts, and verifying after. Entirely on your own GPU.

• 🧠 Works with small models — miii repairs the malformed tool calls a 7B model emits instead of burning a turn on each one, and sizes its own prompt to your context window so the room goes to your code.
• 🧪 miii doctor — not every local model can drive an agent. Grades your installed models on real engineering tasks.
• 🖼️ Paste images — Ctrl+V a screenshot to ask why a UI looks broken. Needs a vision model (llava, llama3.2-vision, …).
• 💧 Lossless output spill — a 50K-line test log is never truncated. The full text goes to disk and the model pages through it.
• 🔒 Permission-gated tools — you approve what the agent touches, and see the exact rule before you save it. A saved wildcard never stretches across a command boundary, so approving npm test can't quietly authorize npm test && rm -rf ~.
• 📄 MIII.md — drop one in your repo to teach miii your conventions and commands. Same idea as CLAUDE.md, read every turn.

Picking a model: 8GB VRAM → qwen2.5-coder:7b · 16–24GB → qwen2.5-coder:14b (sweet spot) · 48GB+ → qwen2.5-coder:32b.

{
  "model": "qwen2.5-coder:14b",
  "effort": "medium",
  "providers": {
    "ollama": { "type": "ollama", "baseUrl": "http://localhost:11434" }
  }
}

{
  "model": "qwen2.5-coder-14b",
  "provider": "llamacpp",
  "providers": {
    "llamacpp": { "type": "openai", "baseUrl": "http://localhost:8080" }
  }
}

npm config set prefix "$HOME/.npm-global"
export PATH="$HOME/.npm-global/bin:$PATH"   # add to ~/.bashrc or ~/.zshrc

[This command output was long (412900 bytes), so I'm showing the start and
 end. The full text is saved at ~/.miii/output/9f3a1c.txt — read it with
 read_file offset/limit to see the middle.]

src/
 ├── agent/       # The core reasoning loop, and tool-call repair
 ├── tools/       # read/write/edit/bash/grep/glob/todos + output spill
 ├── prompt/      # System prompt and MIII.md project context
 ├── permissions/ # Approval rules and how they're scoped
 ├── llm/         # Ollama and OpenAI-compatible backends
 ├── session/     # Saved conversations
 ├── ui/          # Ink terminal UI and input handling
 └── config.ts    # Settings and provider resolution

git clone https://github.com/maruakshay/miii-cli.git && cd miii-cli
npm install && npm run dev

npm run build       # production build
npm run typecheck   # type-check src + eval
npm test            # unit tests
npm run eval        # regression gate (powers `miii doctor`)

Does miii work without internet? Yes. Once you've pulled a model with Ollama, miii runs fully offline — no network calls, no account, no cloud.

Is my code sent anywhere? No. Every file read, edit, and inference happens on your machine — privacy is the default, not a setting.

How is miii different from Claude Code, Cursor, or GitHub Copilot? Those are cloud services — metered, account-gated, and they ship your code to a third-party server. miii is open-source, free, and runs entirely on your hardware, with the same terminal-agent workflow.

How is it different from Continue.dev? Continue.dev is an IDE extension. miii is a standalone terminal agent — no editor required.

Which local LLM is best for coding? qwen2.5-coder at the largest size your VRAM allows. Run miii doctor to grade what you have installed.

Do I need a GPU? No, but it helps. Smaller models run on CPU; a GPU makes larger ones fast enough for real work.

MVP. The core agent loop is stable; actively refining tool execution, streaming, and the permission model. PRs welcome.

MIT © maruakshay

Built for engineers who'd rather own their tools than rent them.