Pi - One LLM Interface to Rule Them All!

Green Matrix Pi Symbol

"Pi number" by J.Gabás Esteban is licensed under CC BY 2.0 .

Chris Patti - Senior Devops Engineer at MIT Online Learning

Sharpen ONE Ax!

This image depicts 3 axes each labeled with the name of an LLM harness

Work With Practically Any Model or Harness

  • Anthropic Opus, Sonnet through Claude Code or Github Copilot
  • OpenAI GPT X.X via Codex, Github Copilot, etc.
  • Local models through Ollama and many more

Choose any model with /model, or hit Ctrl-p to bop between the 2/3 you use a lot!

It's All About The Context (Window!)

Meme from the movie Aladdin showing Genie and Lamp

Drive By System Prompt Size Comparison

If you look at the Github Copilot CLI System Prompt you'll see it's positively massive!

It contains 2978 lines, 56491 words, and 698106 bytes all crammed into your context window before you even start!

Here's Pi's default system prompt by comparison!

Engineers, Start Your Harnesses!

I did a bit of informal benchmarking and found:

Pi's context footprint is ~5–9× smaller than Claude Code's:

  • Trivial prompt (pong): pi sends ~4.1K cached tokens vs. claude's ~19.6K (4.7×)
  • Tool-using prompt (ls + summarize): pi sends ~4.5K vs. claude's ~39.2K (8.7×)

That means more useful work per round trip to the LLM with Pi than with Claude Code.

Sessions!

Check out all the powerful ways Pi lets you manipulate your conversation path with the LLM!

A Fast Growing Ecosystem

There are so many pi extensions others have built!

Explore liberally, adopt cautiously! :)

Pi's Secret Sauce - Self Extension

A trivial example:

Demo where we watch Pi bolt a Github Issues extension into itself

Now Let's Integrate With An External Service - Sentry

Demo where we install Sentry MCP and use it to search for exceptions

Anthropic Shenanigans

If you want to multi-model with Claude, use pi-claude-bridge.