Video summary

Pi Coding Agent is now my absolute favorite...

Main summary

Key takeaways

Technology

Overview

The video introduces Pi (Pi Coding Agent)—or a “Pi agent harness”—as a lightweight, customizable way to run coding agents with open-weight models. It contrasts Pi with Open Code, which is presented as more turnkey and feature-rich.

Key Technological Concepts / Product Features

Purpose of Pi

  • Pi is a coding agent harness that starts almost empty.
  • Its core is essentially a chat interface that can edit files.

Design Philosophy Contrast: Pi vs. Open Code

  • Pi
    • Minimal and DIY, following a Neovim-style approach.
    • No “batteries included.”
    • You install and configure capabilities via extensions.
  • Open Code
    • More out-of-the-box and opinionated.
    • Includes built-in features such as:
      • MCP integration
      • to-dos
      • LSP
      • undo/redo
      • other “full-featured” tooling

Why Pi Now

The speaker argues that open-weight models are rapidly catching up to top proprietary systems, while being:

  • cheaper
  • strong at coding

Examples mentioned:

  • GLM 5.2
  • Kimi K3 (described as state-of-the-art)

Security / Safety Behavior

  • Out of the box, Pi can be dangerous:
    • When asked to “write and run” a Python script, it may execute without explicit approval.
  • Pi’s safety behavior depends on installing a permission system extension.

Setup / Guide Steps (Tutorial-Style)

  1. Install Pi (multiple methods)
    • Examples: curl, Windows PowerShell
    • Package managers: npm / pnpm / bun
    • AUR on Arch
  2. Launch Pi
    • Run pi in a directory to see an “out of the box” interface.
  3. Connect a model provider
    • Requires login via an OAuth flow or an API key.
    • Example shown: Open Code Go using an API key.
  4. Select model and reasoning
    • Switch models using commands like /model
    • Example model: GLM 5.2
    • Set reasoning to maximum via /settings (thinking level → max).
  5. Add missing safety features via extensions
    • Install: pi install npm pi-permission-system
    • Create/configure a permissions JSON policy that:
      • asks before tools
      • asks before bash commands
      • allows read-only commands like ls, git status, git diff
      • requires approval before write/edit actions and other risky operations
  6. Add undo/redo
    • Install: pi install npm pi-undo-redo
    • Notable limitation:
      • Non-git mode undo is limited
      • It can only restore files explicitly touched by Pi
      • Shell-created paths or command-generated changes may not be undo-restorable
    • Demonstrated workflow:
      • create a Python script
      • undo and confirm it’s gone
      • redo and confirm it returns
  7. Add web capabilities
    • Pi cannot do true web search by default; it may need command-based workarounds (e.g., running curl).
    • Install web tool extension: pi install npm pi-web-access
    • After approval, it can perform web search and return summarized results with:
      • multiple sources
      • a web UI showing progress

Reviews / Analysis Conclusions (Speaker’s Recommendation)

  • Choose Pi if you:
    • enjoy tinkering
    • want to build a custom agent from scratch
    • prefer lightweight tooling
  • Choose Open Code (or “Oh My Pi” mentioned) if you want a more ready-to-go experience.

Fun Fact

  • Open Claw is also said to be built on Pi, implying reuse of the Pi harness.

Main Speakers / Sources

  • Primary speaker/source: the video narrator/author (unnamed; the video demonstrates installation, commands, and extensions directly).
  • Referenced tools/projects:
    • Pi Coding Agent
    • Open Code
    • Oh My Pi
    • Open Claw
    • model providers such as Open Code Go
  • Model references mentioned:
    • GLM 5.2
    • Kimi K3

Original video