What minih Does

minih is a tool for creating and running AI agents that work independently. Instead of chatting with an AI in real-time, you define what an agent should do in a folder (with a prompt, optional instructions, and a schema for what the output should look like), then run it from the command line. The agent executes, produces structured JSON results, and saves everything to a timestamped folder so you can review what happened later.

The main benefit is repeatability and structure. You can run the same agent the same way every time, get consistent JSON output, and build a history of runs over time. This makes it easy to use agents in automation, scripts, or CI pipelines — anywhere you need reliable, unattended work rather than an interactive conversation.

How It Works

When you run an agent, minih assembles a prompt from your files (the main prompt, optional shared instructions, and any input parameters you pass), sends it to Claude via the Copilot SDK, and waits for a response. The agent can use tools — it can read files, write files, run commands, whatever you give it access to. When it's done, minih validates the output against a JSON schema you've defined, stores the raw events and final result in a timestamped folder, and displays a summary.

Every agent must produce a "retrospective" — a structured reflection on what worked, what was confusing, and what the agent wishes were different. This feedback is baked into the output contract, not optional. The idea is that over time, reading these reflections helps you improve both the agent prompts and the tool itself.

Who Uses This

minih is for developers and teams who want to automate tasks that currently require human decision-making or exploration. Real examples: a security scanning agent that runs on every code review, a test validation agent that checks whether a test suite actually works, or a feedback aggregator that collects lessons from multiple other agents. You define the agent once, then trigger it from your terminal, a script, or automation — it runs unattended and produces a report. Because runs are timestamped and stored, you can also look back at the history to see how an agent's behavior has changed over time.

It's deliberately not for interactive work. If you're exploring a problem with an AI and going back-and-forth, use Copilot or Claude directly. minih is for fire-and-forget tasks that should always produce the same kind of output.

A Notable Design Choice

The README emphasizes that minih runs agents with no sandbox and no restrictions — agents have full access to read, write, and execute anything on your machine. There's a warning to use containers or throwaway environments if you're experimenting with untrusted agents. This is a deliberate tradeoff: unrestricted access makes agents powerful and flexible, but also dangerous. The tool is designed for trusted agents that you've written or reviewed yourself.