Pi coding agent explained: a tiny, free, hackable AI coder

The short answer: Pi is a small, free coding agent for the terminal, created by Mario Zechner (GitHub name badlogic), now published under the Earendil organisation on GitHub and released under the MIT licence. It gives an AI model the basics (read files, edit them, run commands) and then stops. It skips sub-agents, plan mode and permission pop-ups on purpose, and expects you to add whatever you want. That makes it flexible and cheap, and it also means safety is your job. It is for developers.

Checked on 11 November 2026 against Pi's own website and its repository. When we opened the older address github.com/badlogic/pi-mono it redirected to the repository above. Note: this is not Pi.ai, the consumer chatbot from Inflection. Different product, different company.

What it is, in plain words

A coding agent is an app that lets an AI model work on a project folder instead of only chatting about it. Pi describes itself on its website as a terminal-based coding agent, and its repository calls it "a minimal, extensible agent harness that you can make your own". A terminal is the text window developers type commands into.

Pi works with many AI providers. Its website lists Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, xAI, OpenRouter, Ollama and more, using either an API key (a password that bills your account with that provider) or a login with a subscription you already have. You type /login inside Pi to connect one.

What makes it different

Most agents compete on features. Pi competes on having fewer. The repository says it "ships with powerful defaults but skips features like sub-agents and plan mode", and its website lists "No sub-agents", "No plan mode" and "No permission popups". The idea is to ask Pi to build what you want, or to install a package that does it your way.

What you do get:

• Extensions, skills, prompt templates and themes. Extensions change how Pi behaves, skills are saved instructions, prompt templates are reusable requests, and themes change its looks.
• Packages. You can bundle those pieces as Pi packages and share them through npm or git.
• Four ways to run it: interactively, in print or JSON mode for scripts, over RPC for other programs to control it, and through a TypeScript SDK for building your own apps on top.

Install it

The repository gives a one-line installer for macOS and Linux, and a PowerShell one for Windows:

curl -fsSL https://pi.dev/install.sh | sh

Pi needs Node.js 22.19 or newer, and the installer can set that up. Then you go to your project folder and type pi.

What it costs

Pi itself is free and MIT-licensed. Your costs are the model provider's charges, or nothing extra if you use a subscription you already pay for. We could not find a paid Pi plan on its website: NOT PUBLISHED. If you see someone selling hosted Pi, check that it is the official project.

The trade-off you must understand: no permission system

This is the part to read twice. The repository states that Pi "does not include a built-in permission system for restricting filesystem, process, network, or credential access. By default, it runs with the permissions of the user and process that launched it."

In plain terms: whatever you could do on that computer, the AI can do. It can delete files, send data over the network and read passwords stored in your folders. For safety the documentation describes three ways to contain it:

• Gondolin extension: routes Pi's actions into a small Linux virtual machine on your computer.
• Plain Docker: runs the whole of Pi inside a container, a sealed box on your machine.
• OpenShell: runs Pi in a sandbox with a policy.

If you use Pi, use one of those, or at least a separate computer or user account with nothing precious on it, and keep backups.

The Oh My Pi connection

Oh My Pi is a separate, community-run fork of Pi that adds many built-in tools. Read our Oh My Pi guide if you want the heavier version. The two projects have different maintainers and different goals.

Who it suits

Developers who like to tinker. If you want a small tool whose behaviour you can reshape, and you are comfortable setting your own safety rails, Pi is a good fit. It also suits people who already pay for an AI subscription and want to use it from the terminal.

Who should skip it

• Anyone without a technical person. Setup and daily use assume you are comfortable in a terminal.
• Anyone who needs approval prompts and built-in guardrails. Pi leaves those to you.
• Businesses that need a company to call. It is an open-source project, with a community and an issues page rather than a support contract.
• Owners whose problem is not code. A coding agent will not fix slow follow-ups. Try the plan finder first.

How it compares

See AI coding agent plans compared for prices, and the harness comparison for model policy and limits. If you want something with more guardrails out of the box, our guides to OpenCode and DeepSeek Harness cover tools with permission settings and access levels.