Pi

Most skills can be installed with a single command via npx skills add.

Pi does not support the MCP protocol by default and does not provide a built-in MCP implementation.

A common rule of thumb in the agent domain is that clearing context is better than compressing it — after completing a round of tasks, it is best to start a new conversation with /new, because context compression loses detail.

Pi's conversation management unit is the Session, which uses a tree structure; the /tree command can be used to roll back to historical nodes and create branches.

Pi provides a Steering mechanism where appended instructions default to steering and can change the AI's execution direction in real time; the Follow-up mechanism queues instructions to be executed after the AI finishes its current work.

According to Composio's benchmark tests, Pi completes programming tasks 1.5 to 2 times faster than other coding agents, with task costs lower than mainstream frameworks.

A benchmark test conducted by Databricks on their own codebase of one million lines showed that Pi outperformed Claude Code and Codex in most scenarios, with the Pi + Claude Opus combination achieving the highest code quality.

Pi is an open-source coding Agent whose core feature is that the model is fully decoupled and not bound to any single large model company.

Token Pass can be used directly with coding agent tools such as fx, Command Code, and Pi, or you can build a custom personal agent through the AI SDK.

Skep does not wrap or proxy local CLI agents such as Claude Code, Codex or Pi; agents still run as local programs, and Skep is only responsible for session-level message passing.

Mainstream Agent frameworks (such as Claude Code, Codex, and Pi) are open-ended and do not know at startup what requests they will receive or what purpose they will be given, whereas Agents deployed by enterprises almost always have a specific and dedicated purpose.

Atomic Agent's supported import sources include Claude Code, Codex, Hermes, OpenClaw, Pi and Oh-My-Pi, covering dimensions such as skills, memory, MCP servers and session records; among them, Hermes's and OpenClaw's memory data cannot be migrated, because the two have different storage formats and no exchange standard.

This approach uses ChatGPT tasks to organize discrete Pi sessions into a continuously running process, and uses MCP tools and Computer Use to fill capability gaps.

Pi's author responded in a community post that fullscreen is the default in order to fix a Windows terminal issue, and that the old mode will not be removed.

Starting from Pi 1.0, the default interface mode was changed to fullscreen, which takes over the terminal's scroll history; this can be changed to regular in the configuration file to restore the scrollbar.

Connecting to domestically developed models requires setting the corresponding platform's environment variables in the terminal via export; the official documentation provides a reference table of variable names and platforms.

After a successful upgrade, Pi's version number changed from 0.85 to 1.1.

The domestic platforms listed in Pi's official reference table include DeepSeek, Kimi, MiniMax, Alibaba's Tongyi Qianwen, Xiaomi's models, and others.

A small task like FizzBuzz cannot prove that domestically developed models are as capable as the most expensive flagship models, and the cost of large, complex tasks remains relatively high.

Pi requires at least Node 22.19 to run; testing confirmed it runs normally on Node 24.11.