Codepit The toolbox

Run pi agents inside Claude Code

Adds a pi agent type whose work runs through the pi CLI, on any model pi can reach. It still shows in Claude Code's agent list, streams live, takes follow-ups and can open as a teammate in a tmux pane.

Scan in Codepit View source

Scan checks all of FazalAAli/pi-agent-for-claude on your computer, before anything installs.

Codepit runs on your computer. Open this page there to scan it before installing.

A big honey cog turning a small grey one

See it running

A real recording from the author's repo.

What you get

Ask for pi in plain language and Claude picks the pi-agent-for-claude:pi agent type. Name the agent to get a teammate pane.

  • Any model pi can reach
  • Live streaming
  • Follow-ups
  • Teammate panes

Before you start

  • Claude Code 2.1.275 or later.
  • pi, logged in to at least one provider.
  • Node.js, on macOS or Linux.
  • tmux, for teammate panes.

Try saying

>Have pi run with model kimi k2 and summarize this repo

Claude starts the pi agent with a pi-model: line, pi works with its own tools, and the answer ends with a line naming the provider and model that really answered.

Worth knowing

pi runs its own read, bash, edit and write tools, outside Claude Code's permission prompts, tool policy and safety classifier. A pi agent can change files without asking, so restrict pi in its own config if that matters.

It is experimental: built on Claude Code's early-access function hooks and some undocumented engine behaviour, so a Claude Code update may break it.

With function hooks off, a pi agent runs as Claude Haiku instead. A real pi answer always ends with the answered-by-pi line, so check for it.

The task runs on pi's model, through the provider pi is logged in to. Cost figures for pi agents are priced as Claude tokens, so check your provider's billing.

Read the author's README README.md