Newmiii now works in your browser too — a desktop-style app like Claude Desktop. Run miii webSee it ›

Free forever. Works with any model.

Made in India.
Built for everyone.

A professional-grade coding agent for your terminal or your browser — 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.

View on GitHub
miii — zsh
> ⏎

Most AI tools rent you a cloud.
miii uses your machine.

The model runs on your GPU. Nothing is metered, nothing is uploaded, and it keeps working with the wifi off.

Private by default

Your code never leaves your machine.

Cloud agents ask you to trust your codebase to someone else's server. miii talks to a model on localhost, so nothing touches the wire.

your repo→localhost:11434→your GPU

Cost

$0

No per-token billing. It runs on your hardware.

Offline

Yes.

No keys, no account, no network needed.

A real agent, not a chat box

It edits, runs and checks its own work.

Exact edits instead of blind rewrites, and it runs your tests before it says it's done.

read_fileedit_filewrite_fileglobgreprun_bash

Permission-gated

You approve every change.

Every write and shell command asks first, and file tools stay inside your workspace. "Always" rules are saved per project.

.miii/permissions.json

Runs on

  • Ollamadefault
  • llama.cpplocal
  • LM Studiolocal
  • Any OpenAI-compatible server

Lossless output

Nothing gets cut off.

Huge tool output goes to disk with a preview of the start and end, and the model reads the middle when it needs to.

Model doctor

Know which model works.

miii doctor tests your local models on real engineering tasks.

Up and running in three steps.

Requires Node.js 18 or later and Ollama running locally.

1

Pick a model

ollama pull qwen2.5-coder:14b
2

Install miii

npm install -g miii-agent
3

Talk to it

miii
> @src/server.ts add rate limiting
> why are my tests failing?

Rather have a window?
miii web

The same agent in a desktop-style app in your browser — for when you'd rather click than type. Same sessions, same models, same permission rules as the terminal.

Chat, diffs and approvals

Everything the terminal shows, laid out like a chat app.

Replies stream in, each tool call opens to show the exact command and its output, edits show as real diffs, and every write waits for your click.

One command

Open it from any project.

It starts a local server and opens your browser. Add --port, --no-open or -c to resume your last chat.

miii web
# or: miii --web

Locked to your machine

Yours alone.

It listens on localhost only, and every request needs the one-time key in the link miii prints. Other sites can't reach it.

127.0.0.1#token=…

Everything is a / away.

The same command list you get inside miii. Type to filter, then press enter or click to copy. Press / anywhere on this page to jump here.

    ↑↓ navigate · ⏎ copy · esc clear

    What's new.

    Read live from CHANGELOG.md every time the page loads.

    Loading releases…

    Full changelog

    Questions.

    Is miii really free?

    Yes. miii is MIT-licensed with no paid tier, no account and no sign-up. With a local model you pay nothing at all.

    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.

    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.

    Is there a desktop or web app?

    Yes — run miii web in any project and it opens in your browser: a chat sidebar, streaming replies, real diffs and one-click approvals. It's the same agent and sessions as the terminal, served only on localhost, and every request needs the key in the link it prints.

    Do I need a GPU?

    For fully local models you want one — qwen2.5-coder:3b runs on CPU, but it will be slow and need more hand-holding. Apple Silicon works well: unified memory counts as VRAM here.

    Why not just use Claude Code or Cursor?

    Use them if you can. miii is for when you want the same workflow without being tied to one vendor or one subscription — or you need to go fully offline for air-gapped work, NDA'd code, or a flight with no wifi.

    Own your tools.

    No keys, no billing, no cloud. A coding agent in your terminal in under a minute.

    Star on GitHub