miii

What does miii do?

miii is a terminal coding agent you install with npm i -g miii-agent. The README introduces it as "a free, local-first alternative to Claude Code, Cursor and GitHub Copilot. It plans, edits, runs and tests code." A session reads your repository, proposes a plan when you say /plan, writes files, runs commands, and checks the result with a second smaller model before it stops.

Fourteen model providers are supported in one CLI, split into hosted and local. Hosted: Anthropic Claude, OpenAI GPT, Google Gemini, DeepSeek, Mistral, Groq, Cerebras, xAI Grok, OpenRouter, Together. Local: Ollama, LM Studio, llama.cpp, vLLM. You switch mid-session with /provider. A desktop-style chat UI is available by running miii web; it listens on localhost and asks for a one-time key.

Key facts

• Licence: MIT on the code.
• Platforms: Node.js 18 or newer; Windows (via PowerShell install script), macOS and Linux.
• Providers: 10 hosted (Anthropic, OpenAI, Gemini, DeepSeek, Mistral, Groq, Cerebras, Grok, OpenRouter, Together) and 4 local (Ollama, LM Studio, llama.cpp, vLLM).
• Approvals: every file write and shell command asks for approval. File tools are limited to the working directory.
• /rewind undoes the agent's file changes and the conversation together.
• miii fix runs the tests, fixes failures and repeats until they pass; attempts that do not help are rolled back.
• A "done-check" runs a second smaller model to verify the work is really finished before the agent stops.
• Headless mode: miii -p "prompt" for scripts, CI and git diff piping.
• The only outbound call miii makes on its own is an npm version check.

What does it replace, and where does it fall short?

miii replaces Cursor, Claude Code and GitHub Copilot as a terminal-first coding agent with your own choice of model. Its own README says: "miii gives you the same plan to edit to run to verify workflow, MIT-licensed and with no subscription. You only pay your model provider, or nothing if you run a local model."

Where it falls short: tool calling is what makes an agent usable, and the README itself notes that qwen3-coder and devstral are trained for it while miii has to repair malformed tool calls from other models. A GUI editor's inline-suggestion flow is not here; this is a terminal and a browser-rendered chat on localhost, not a VS Code integration. The web UI only listens on localhost by design, so there is no shared team instance. The maintainer sells no hosted or paid tier; sponsorship on GitHub covers the testing costs.

How does miii run?

Install with npm i -g miii-agent, then run miii in a repository. For a local model, run ollama pull qwen3-coder:30b first and start miii; for a hosted model, run miii provider add gemini and provide the API key, or set an environment variable like ANTHROPIC_API_KEY or OPENAI_API_KEY. On Windows a PowerShell one-liner at the project's install.ps1 installs it; Node 18 or newer is still required. The web UI starts with miii web.

A GPU helps for local models but is not required: qwen2.5-coder:3b runs on a CPU, though slowly. Apple Silicon counts unified memory as VRAM. The README's VRAM table recommends qwen2.5-coder:3b under 4 GB, :7b at 8 GB, :14b at 12 to 16 GB, and qwen3-coder:30b from 24 GB.

Who is miii for?

A developer who wants a terminal coding agent on their own machine, with the choice of a hosted provider or a local model, and no subscription. Someone whose work is under NDA, air-gapped or on a flight benefits from the fully local path. A team that wants a server-side agent across one shared workspace should look at UseAgent on the open-source alternatives to Cursor list.

What limits does the README state?

Tool calling quality varies by model: qwen3-coder and devstral are trained for it, and miii repairs malformed tool calls from other models. For local models a GPU helps and a 3B model on CPU is slow. The web UI only listens on localhost, with a one-time key. The project is independently built in India, and the maintainer notes that sponsorship pays for the GPU time and API costs of testing against every model.

Questions people ask

Is miii open source?

Yes. The code is MIT, an OSI approved licence, with no additional conditions and no CLA on contributions mentioned in the README. Issues and pull requests are accepted. The repository name is miii-cli; the npm package is miii-agent.

Can miii be used commercially?

Yes. MIT allows commercial use, modification and redistribution, provided the licence travels with the code. miii itself is free. You pay your chosen model provider's bill, or nothing if you run a local model with Ollama, LM Studio, llama.cpp or vLLM.

Does miii work offline?

Yes. With a local model through Ollama, LM Studio, llama.cpp or vLLM, every call stays on your machine. The README states that air-gapped work, code under NDA, and a flight with no wifi are supported use cases. The only outbound call miii makes on its own is an npm version check.

How does miii compare with Claude Code and Cursor?

Claude Code is Anthropic's own terminal agent tied to its models; Cursor is a GUI editor tied to its model routing. miii is a terminal agent that works with any of the 14 providers in its list, including local models, and is MIT licensed. The open-source alternatives to Claude Code page sets it beside Codex, Crush, Gemini CLI, opencode and others.

Can miii break my repository?

Every file write and shell command asks for approval first, and file tools are limited to the working directory. If a change goes wrong, /rewind restores the files and the conversation together. The README notes this is a design choice, not an opt-in feature.

Sources

• miii README and LICENSE: github.com/maruakshay/miii-cli, read 2026-10-26. The README is the source for every feature, provider, slash command and limit above.
• Project website: miii.in, read 2026-10-26.