Skip to content
CodingAgents.md

Pi

Overview

Section titled “Overview”
CompanyOpen source (earendil-works)
TypeTerminal CLI
PricingFree — bring your own API key
ProtocolMCP
Config FormatsAGENTS.md
LanguageTypeScript
Repositorygithub.com/badlogic/pi-mono

What It Does

Section titled “What It Does”

Pi is a monorepo of tools for building and running AI agents. Its centerpiece is a terminal coding agent that reads files, writes code, runs shell commands, and iterates inside your project directory.

The design point is the system prompt. Where most harnesses ship thousands of tokens of instructions, Pi keeps its base prompt deliberately small and lets the model work from the repository instead. It ships as a set of npm packages, so the harness is readable and hackable rather than a black box.

Key Strengths

Section titled “Key Strengths”
  • Minimal system prompt: A small base prompt leaves more of the context window for your actual code
  • Hackable architecture: Plain TypeScript packages you can read end to end and fork
  • Provider-agnostic: Bring your own key across providers
  • Composable: The agent is one package among several, so you can build your own harness on the same parts
  • Open source: No account, no telemetry gate, no seat

Trade-offs

Section titled “Trade-offs”
  • Smaller ecosystem than Claude Code or Cursor — fewer integrations out of the box
  • No managed cloud agent; it runs where you run it
  • A lean prompt means the agent relies more on what your repository tells it, so AGENTS.md and skills matter more here than in a heavier harness

Where to find skills for Pi

Section titled “Where to find skills for Pi”

Pi has no built-in skills registry, so the skills you give it come from wherever you find them. A skill is a SKILL.md file, so a repository path or a downloaded folder is enough — point Pi at the file and it reads it like any other.

The largest open catalogue is atskills.one, with 60,000+ agent skills you can read or copy into your project. See @skills for how referencing works.

This matters more with Pi than with a heavier harness: the lean base prompt means the agent leans on what your repository tells it.

Best For

Section titled “Best For”
  • Developers who want to read and modify their agent harness
  • Token-conscious workflows where prompt overhead is the constraint
  • Building a custom agent on top of an existing, small set of parts