miii

🇮🇳 Made in India. Free forever. Works with any model.
A professional-grade coding agent for your terminal — it plans, reads, writes, and verifies code.
Bring your own keys (Claude, GPT, Gemini, DeepSeek) or run it 100% locally on your GPU.

Apna code, apna model, apni marzi.

🆓 Free & open source — no account, no subscription  ·  🔌 Any provider, one tool  ·  ⚡ Install to first edit in a minute

🎯 Why miii?

Most agentic coding tools lock you into a single vendor, a monthly subscription, and bills in dollars.

miii brings the professional "plan → edit → run → verify" workflow to everyone, on their own terms. Use the best hosted models for complex tasks, free tiers for experimentation, or local models for complete privacy. No accounts, no subscriptions—just your code and your choice of AI.

🚀 Get started in 60 seconds

1. Install

npm i -g miii-agent

2. Run

Pick your preferred backend:

Hosted (uses existing env keys like ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.)

miii provider add gemini
miii

Local (100% free, offline, no account needed)

ollama pull qwen3-coder:30b
miii

Switch anytime with /provider inside miii.

3. Code

Talk to it like a teammate:

> 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

It plans before it acts, and runs your test suite before it claims to be done.

[!TIP] Windows Users: Use irm https://raw.githubusercontent.com/maruakshay/miii-cli/main/install.ps1 | iex. Needs Node 18+ and a POSIX shell (git-bash or WSL); otherwise, miii falls back to Node's filesystem calls.

🔌 Your model, your choice

One command wires up any backend: /provider add <name> inside miii, or miii provider add <name> from your shell. The endpoint is built-in; the key comes from the environment variable you already export.

☁️ Hosted (your key)	🏠 Local (free, offline)
Anthropic · OpenAI · Gemini · Groq · OpenRouter · DeepSeek · Mistral · Together · xAI	Ollama · LM Studio · llama.cpp · vLLM

Anything else that speaks the OpenAI protocol works too: /provider add <name> <baseUrl>.

Paying nothing is a real option. miii is free, local models are free, and providers like Gemini, Groq, OpenRouter, and Cerebras offer free tiers — enough to get real work done without a credit card.

	Single-vendor agents	miii
The tool	Subscription	Free, MIT-licensed
Models	One vendor's	14+ providers, or your own GPU
Setup	Account + login	npm i -g miii-agent
Your code	Goes to that vendor	Goes where you choose — or stays local
Offline	No	Yes, with a local model

✨ Power Features

miii delivers the advanced workflow of top-tier agents without the vendor lock-in:

• ⚖️ Decision Box — A second, smaller model verifies if the work is actually complete before the agent stops. No more false positives.
• 📋 Plan Mode — Use /plan for a read-only session. miii researches and proposes a strategy; nothing is written until you approve.
• 🔒 Permission-Gated — You are in control. Approve every file write and shell command. Rules persist in .miii/permissions.json.
• ⟲ /rewind — Undo agent changes and the conversation history. Checkpoints make experimenting risk-free.
• 🧩 Subagents — Delegated task loops handle complex searches in their own context, returning only the final answer.
• 💭 Live Status — One status line shows what the agent is doing, with live timing and token counts. Full thoughts are a ctrl+t away.
• ⚡ Mid-turn Steering — Send corrections or new instructions while the agent is running. miii queues your input and adapts its next step in real-time.
• 🎯 Turn Summaries — Every completed turn ends with a concise summary of files changed and test results. No more scrolling up to see what actually happened.
• 🧠 Model-Aware — miii handles broken tool calls from smaller models and optimizes prompts for various context windows.
• 🔀 Hot-Swapping — Use /provider and /models to switch backends instantly without restarting your session.
• 🔌 MCP Support — Connect to GitHub, Postgres, or your own internal services via Model Context Protocol.
• 🪝 Shell Hooks — Trigger custom commands on tool events (e.g., auto-format on write or block specific paths).
• 🖥️ Headless Mode — Run miii -p "prompt" for scripts, CI/CD, or piping git diffs directly into the agent.
• 🔁 Fix until green — miii fix runs your tests, fixes what fails, and runs them again until they pass. The test result, not the model, decides what's fixed; attempts that don't help are rolled back automatically, so small local models only ever move the code forward. miii fix --check "npx tsc --noEmit" --output-format json for CI.
• 🐚 Shell-Backed Edits — Operations run through the shell for maximum compatibility, with a robust Node.js fallback.

Every slash command

/plan · /models · /provider · /agents · /mcp · /permissions · /rewind · /sessions · /memory · /context · /compact · /cost · /review · /init · /export · /copy · /settings · /vim · /new · /clear

🏠 Going local? Pick a model for your GPU

Budget a few GB above the download size for the context window.

VRAM	Recommended Model	Size	Note
< 4GB	qwen2.5-coder:3b	1.9 GB	Fast, lightweight
8GB	qwen2.5-coder:7b	4.7 GB	Great balance
12–16GB	qwen2.5-coder:14b	9.0 GB	Professional grade
16–24GB	gpt-oss:20b · devstral:24b	~14 GB	High reasoning
24GB+	qwen3-coder:30b ⭐	18.6 GB	State-of-the-art local coding

Pro Tip: Tool-calling is what separates a usable agent from a frustrating one. qwen3-coder and devstral are natively trained for it; miii automatically repairs calls from other models to keep them working.

FAQ

Is miii really free?

Yes. miii itself is MIT-licensed with no paid tier, no account and no sign-up. What you pay for is the model: nothing if it runs on your machine, and whatever your provider charges if you use a hosted one — several of which have free tiers.

Is my code ever uploaded?

With a local provider, your code only ever travels to localhost. No telemetry and no analytics — the one outbound call miii makes on its own is a version check against registry.npmjs.org, which carries nothing but the package name. Add a hosted provider and that provider sees what you send it, which is why local is the default.

Can it break my repo?

Every write and every shell command asks first, and file tools are confined to the working directory. When something does go wrong, /rewind puts the files and the conversation back.

Do I need a GPU?

No. Use a hosted provider and any laptop will do. For fully local models you want one — qwen2.5-coder:3b runs on CPU, but it will be slow and it will need more hand-holding. Apple Silicon works well — unified memory counts as VRAM here.

Why not just use Claude Code / Cursor?

Use them if you can. miii is for when you want the same workflow without being tied to one vendor or one subscription: pick any model, pay only for what that model costs, or go fully offline for air-gapped work, NDA'd code, or a flight with no wifi.

Sponsor miii 💛

miii is built and maintained independently, from India, with no VC money and no paid tier. Sponsorship is what keeps it that way — it pays for GPU and API time to evaluate new models across providers, test hardware, and the hours that go into making every model, big or small, behave like a real coding agent.

• Individuals — GitHub Sponsors, any amount, one-time or monthly.
• Companies — if your team uses miii for private or air-gapped work, sponsoring gets your logo here and a direct line for the features you need. Open an issue or reach out via GitHub.
• Free ways to help — star the repo, share it with your dev community, file issues, and send PRs.

Contributing

Issues and PRs are welcome from anyone, anywhere. Good places to start: model compatibility reports (which models work well on which hardware), bug reports with a repro, and docs improvements.

---

⭐ Star it if you think great coding tools should be free for everyone.
_{Built with ❤️ in India 🇮🇳 · MIT © maruakshay}