Why miii?
- Free and open source. MIT-licensed, with no account, subscription or paid tier.
- Any LLM, one CLI. 14+ providers, and you can switch mid-session with
/provider.
- Private by default. With a local model, your code never leaves
localhost, and there's no telemetry.
- You stay in control. Every file write and shell command asks first, and
/rewind undoes anything.
Quick start
npm i -g miii-agent
Local, offline LLM (free):
ollama pull qwen3-coder:30b
miii
Hosted model (uses keys you already have, like ANTHROPIC_API_KEY or OPENAI_API_KEY):
miii provider add gemini
miii
Then ask it to do things:
> 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
Prefer a browser? miii web opens a desktop-style chat UI with the same agent, sessions and permission rules. It only listens on localhost and needs a one-time key to connect.
[!TIP]
Windows: irm https://raw.githubusercontent.com/maruakshay/miii-cli/main/install.ps1 | iex (needs Node 18+).
Supported LLM providers
| ☁️ Hosted (bring your own API key) | 🏠 Local (free, offline) |
|---|
| Anthropic Claude · OpenAI GPT · Google Gemini · DeepSeek · Mistral · Groq · Cerebras · xAI Grok · OpenRouter · Together | Ollama · LM Studio · llama.cpp · vLLM |
Tool calling is what makes an agent usable. qwen3-coder and devstral are trained for it, and miii repairs malformed tool calls from other models.
Features
- 📋 Plan mode.
/plan researches your code read-only and proposes a plan. Nothing is written until you approve it.
- 🎯 Turn receipts. Each turn ends with ✓, a coloured
+42 −10 diffstat and the test result, and a running total of changes stays on screen.
- 🔁 Fix until green.
miii fix runs your tests, fixes what fails and repeats until they pass. Attempts that don't help are rolled back.
- ⚖️ Done-check. A second, smaller model checks the work is really finished before the agent stops.
- ⟲ Rewind and checkpoints. Undo the agent's file changes and the conversation together.
- ⚡ Mid-turn steering. Type corrections while the agent runs, and it picks them up on its next step.
- 🔔 Notifications. The terminal rings when a long turn finishes or needs your approval.
- 🧩 Subagents, MCP and hooks. Delegate searches, connect GitHub or Postgres over the Model Context Protocol, and run shell hooks on tool events.
- 🖥️ Headless mode for CI.
miii -p "prompt" lets you script it, pipe in git diff, or run it in CI/CD pipelines.
/plan · /models · /provider · /agents · /mcp · /permissions · /rewind ·
/sessions · /memory · /context · /compact · /cost · /review · /init ·
/export · /copy · /settings · /vim · /new · /clear
- Rewards only for real progress. Dopamine responds to unexpected reward and fades for rewards you can predict (Schultz et al., 1997). So the green ✓ appears only when files changed or tests passed. A plain answer gets a quiet check.
- Useful information during the wait. Uncertain outcomes keep the reward system active (Fiorillo et al., 2003). Instead of a bare spinner, the status line shows what the agent is doing, how long it has taken and how many tokens it has written.
- No random rewards. Every signal comes from a real diff, test result or finished turn.
Design inspiration, not a clinical claim.
Pick a local model for your GPU
| VRAM | Model | Size |
|---|
| < 4 GB | qwen2.5-coder:3b | 1.9 GB |
| 8 GB | qwen2.5-coder:7b | 4.7 GB |
| 12–16 GB | qwen2.5-coder:14b | 9.0 GB |
| 24 GB+ | qwen3-coder:30b ⭐ | 18.6 GB |
Apple Silicon works well, since unified memory counts as VRAM.
FAQ
Yes. miii gives you the same plan → edit → run → verify workflow, MIT-licensed and with no subscription. You only pay your model provider, or nothing if you run a local model.
Yes. With Ollama, LM Studio, llama.cpp or vLLM, everything stays on your machine. That suits air-gapped work, code under NDA, or a flight with no wifi. The only call miii makes by itself is an npm version check.
Every write and shell command asks for approval first, and file tools are limited to the working directory. If something does go wrong, /rewind restores both the files and the conversation.
No. Any laptop works with a hosted provider. For local models a GPU helps: qwen2.5-coder:3b runs on a CPU, but slowly.
Sponsor and contribute
miii is built independently in India. Sponsorship pays for the GPU time and API costs of testing it against every model. You can sponsor through GitHub Sponsors.
Issues and PRs are welcome. The web UI lives in web/ (React, Vite, Tailwind), and npm run dev:web starts it.
