Mi

Product Introduction

1. Definition: Mi is an autonomous coding agent, a single-file JavaScript (Node.js) tool designed for AI-driven software development and automation. It falls under the technical categories of AI coding assistants, agentic AI frameworks, and command-line development tools.
2. Core Value Proposition: Mi exists to provide a lightweight, dependency-free, and highly flexible agentic AI harness that enables developers to automate complex coding workflows using any OpenAI-compatible language model. Its core value is agentic coding in a minimal, 30-line-of-code loop, offering full system access and multi-agent capabilities without complex frameworks.

Main Features

1. Autonomous Agentic Loop: The core engine is a recursive loop that calls a Large Language Model (LLM), checks for requested tool calls, executes those tools (like bash commands or skills), and feeds the results back into the conversation context. This loop continues until the AI agent determines its task is complete and provides a final text response. It works with any OpenAI-compatible API endpoint, including OpenAI, Ollama, LM Studio, LiteLLM, and vLLM.
2. Built-in Tool System with Hot-Loading: Mi ships with two foundational tools and a modular architecture for more. The bash tool grants the agent full system access for Git operations, running compilers, package managers, and file I/O. The skill tool loads and executes markdown playbooks (skills) from a local directory. Developers can add custom tools by dropping .mjs files into a tools/ directory, which are hot-loaded before each LLM call, enabling self-extending functionality.
3. Multi-Agent Workflow Tools (delegate & goal): These tools transform Mi from a single agent into a coordinator for multi-agent systems. The delegate tool spawns a Mi subagent with a specific prompt, streaming its work in real-time. The goal tool orchestrates a series of subagents to achieve a high-level objective, using a persistent progress file to iteratively build upon previous attempts until a specified check command (e.g., npm test) succeeds, enabling complex, multi-step problem-solving.

Problems Solved

1. Pain Point: The complexity and overhead of setting up and managing AI-powered coding agents and development automation scripts. Traditional methods often require extensive framework knowledge, dependency management, and lack flexibility in model choice.
2. Target Audience: The primary user personas are software developers, DevOps engineers, and technical leads seeking to automate coding tasks, legacy code refactoring, test writing, debugging, and repository exploration. It also appeals to AI enthusiasts and researchers experimenting with agentic AI workflows and multi-agent systems in a lightweight environment.
3. Use Cases: Essential scenarios include autonomously debugging a failing test suite by reading logs, modifying code, and re-running tests; refactoring a legacy module to use a new library; exploring an unfamiliar codebase to generate a summary; and creating a new skill/tool to extend the agent's own capabilities, following the "self" skill pattern.

Unique Advantages

1. Differentiation: Unlike monolithic AI coding platforms or complex agent frameworks (e.g., LangChain), Mi is a single JavaScript file with zero external dependencies beyond Node.js built-ins. It offers greater transparency, control, and portability. Compared to scripted automation, Mi provides adaptive, LLM-driven reasoning and tool selection.
2. Key Innovation: The combination of extreme minimalism (a ~30 LOC core loop) with powerful, recursive multi-agent capabilities (delegate, goal) is its key innovation. The use of a persistent progress file for the goal tool allows iterative subagents to maintain context and learn from prior failures, a simple yet effective form of short-term memory and planning uncommon in such lightweight tools.

Frequently Asked Questions (FAQ)

1. How does Mi compare to GitHub Copilot? Mi is an autonomous coding agent that executes commands and writes code directly in your environment, while GitHub Copilot is an AI-powered code completion tool integrated into your IDE. Mi performs multi-step tasks (like debugging and refactoring) autonomously, whereas Copilot assists developers line-by-line.
2. Can I run Mi with a local, open-source LLM? Yes, Mi is designed for compatibility with any OpenAI-compatible API. You can configure it to use local models running on Ollama, LM Studio, or vLLM by setting the OPENAI_BASE_URL and MODEL environment variables, making it a powerful tool for private, offline AI coding.
3. Is it safe to give Mi bash tool access to my system? The bash tool provides full system access, so it should be used with caution. It is intended for use in controlled environments, such as development containers, sandboxes, or on non-critical systems. Users must audit the agent's proposed commands before execution in sensitive contexts.
4. What are Mi "skills" and how do I create one? Skills are markdown playbooks stored in the skills/ or ~/.agents/skills/ directories. They provide structured workflows (e.g., "tdd", "refactor") that the agent can load on demand. You can create a new skill by writing a markdown file with step-by-step instructions, or use the bundled new-skill skill to have the agent write one for you.
5. How does the goal tool handle complex, multi-step problems? The goal tool uses an iterative approach. It creates a progress markdown file that logs the goal, check command, and outputs from each subagent iteration. Each subsequent subagent reads this file to understand what has been tried and what errors occurred, allowing it to build upon previous work rather than starting from scratch each time, effectively solving complex problems through guided iteration.