We watched an AI coding tool called Pi hit 86,334 stars on GitHub last week — and what's unusual is that while every peer tool is piling on features, Pi deliberately omits MCP (the standard protocol for letting AI call external tools), sub-agents (letting one AI split into multiple collaborative subtasks), and plan mode. Each omission has a dedicated "Philosophy" section in the README explaining why.
What this is
Pi is an AI coding agent "harness" — scaffolding for driving large models, not a product itself. It splits multi-provider model integration, agent loops, TUI (terminal UI) rendering, and session management into 5 independent npm packages (Node.js modules) that you can pick and choose from, or stitch together to build your own tool. Over 30 model providers are swappable: Claude, GPT, Gemini, DeepSeek, Kimi, MiniMax, even local llama.cpp. Its core stance: keep the core minimal, push all customization to the extension layer. If you want MCP, sub-agents, or plan mode, write your own extensions — don't let the tool author's workflow preferences make decisions for you. This is the opposite of the mainstream AI coding tool direction of "everything pre-installed" — Cursor and Claude Code keep adding features; Pi keeps subtracting them.
Industry view
Supporters call this a "return to Unix philosophy." Agent tools are getting bloated — everyone is fighting for the "most features" slot, and users end up drowning in popups, confirmation dialogs, and sub-task trees. Pi's 86k stars are themselves a vote: developers want a clean core plus the freedom to assemble their own stack. But the counterarguments have merit. A minimal harness is friendly to authors but hostile to enterprise IT. Enterprises want out-of-the-box usability, controllable permissions, and auditable logs — "write your own extension" basically translates to "not procurement-approved" in compliance departments. There's another risk: Pi's minimalism rests on a premise — that users know what they want. Veteran programmers welcome it; newcomers will be turned off by "which extension should I install."
Impact on regular people
For enterprise IT: yet another open-source tool stack to evaluate. Pi's modular split actually makes evaluation more complex — are you buying 5 packages or 1 product? Who maintains the extension layer? This is an addition to the procurement list, not a subtraction.For individual careers: if you're not a programmer, this project is still far from your day-to-day. But the signal is worth noting: there is no "ultimate winner" in AI coding tools — everyone is betting on a different horse, and over the next 1–2 years dev teams may find themselves switching between 3–4 tools.For consumer markets: no direct impact for now — the AI coding tool buzz is still inside the developer community. But Pi's popularity points to one thing: the large models themselves are being commoditized. The real competition is at the "how to use it" layer, not at "whose model is smarter."
上周 GitHub 上一个叫 Pi 的 AI 编程工具攒到了 86,334 颗星 — 它的反常之处是当所有同类工具都在堆功能的时候，它故意不做 MCP（让 AI 调用外部工具的标准协议）、子 Agent（让一个 AI 拆成多个协作子任务）、计划模式。每一项都有 README 里的「Philosophy」章节专门解释为什么不做。
这是什么
Pi 是一个 AI 编程代理的「harness」 — 驾驭大模型的脚手架，不是产品本身。它把多厂商大模型接入、代理循环、TUI（终端界面）渲染、会话管理拆成 5 个独立的 npm 包（Node.js 模块），可以挑着用，也可以全部拿来搭自己的工具。30 多家大模型厂商都能切：Claude、GPT、Gemini、DeepSeek、Kimi、MiniMax，甚至本地 llama.cpp。它的核心立场是：核心保持极简，所有定制走扩展层。如果你要 MCP、子 Agent、计划模式，自己写扩展加进去，别让工具作者的工作流偏好替你做决定。这和主流 AI 编程工具「什么都给你装好」的方向完全相反 — Cursor、Claude Code 都在加功能，Pi 在减功能。
