Pi 1.0評測：登頂HN的開源極簡Coding Agent | Pi 1.0 Review: The Minimalist Coding Agent Topping HN

10月 02, 2026

By Kit 小克 | AI Tool Observer | 2026-10-03

🇹🇼 Pi 1.0評測：登頂HN的開源極簡Coding Agent

Pi 1.0 這幾天衝上 Hacker News 首頁，拿下超過 1000 點讚，是近期最受矚目的開源 Coding Agent（編碼代理）新聞。它由 Flask 作者、Sentry 前工程負責人 Armin Ronacher 的公司 Earendil 發布，核心概念很簡單：做一個比 Claude Code、Codex 更透明、更輕量的終端機編碼代理，而且完全開源、MIT 授權、免費使用。

Pi 1.0 到底是什麼

Pi 最早由開發者 Mario Zechner（libGDX 作者）打造，今年稍早被 Earendil 收購。它是一個裝在終端機裡的 AI 編碼代理，能讀你的程式碼庫、改檔案、跑指令，支援超過 15 家模型供應商，你可以自由切換 Claude、GPT、Gemini 或本地模型。

安裝很簡單：

curl -fsSL https://pi.dev/install.sh | sh
或用 npm install -g @earendil-works/pi-coding-agent

裝好後打 /login 接上你的 API key 或訂閱，就能開始下指令。Pi 本身完全免費，你只需要支付背後模型供應商的費用。

跟 Claude Code 的哲學差異

Zechner 當初做 Pi 的動機，是不滿 Claude Code 的 system prompt 和工具定義每次更新都在變，行為跟著飄移，自己的工作流程常常被打斷。Pi 的設計刻意極簡：一開始連 MCP、子代理（subagent）、背景執行 bash 都故意不做，目的是讓使用者能清楚看到代理到底在想什麼、做什麼，不被隱藏的鷹架（scaffolding）干擾。

Pi 1.0 的大轉彎：加回 MCP

有趣的是，這次 1.0 正式版卻「髮夾彎」加入原生 MCP 支援，還推出實驗性的 Pi Durable 套件，讓長時間執行的代理任務撐過斷線、持續運作。官方同時把全螢幕 TUI 設為預設介面，並把 codemode 的提示詞 token 用量砍了約 40%。換句話說，Pi 從「刻意做減法」走回「該有的功能還是要有」，但步調比 Claude Code 慢、也更克制。

值得從 Claude Code 換過去嗎

老實說，不用急著換。社群實測顯示，Pi 在同樣任務上耗用的 token 比 Claude Code 少（55.9 萬 vs 74.2 萬），但完成度和驗證通過率略輸一截；而且 Pi 沒有 Claude Code 的 Plan Mode、多代理協作和龐大外掛生態，企業團隊要重新摸索工作流程的成本不低。

比較實際的用法是：日常開發、團隊協作繼續用 Claude Code；如果你在意成本、想接本地模型、或單純想要一個「不會每週變臉」的輕量工具，Pi 1.0 值得裝起來試試看。

好不好用，試了才知道。

🇺🇸 Pi 1.0 Review: The Minimalist Coding Agent Topping HN

Pi 1.0 just shot to the top of Hacker News with over 1,000 points, making it the AI dev-tool story of the week. It's an open-source coding agent built by Earendil, the company founded by Flask creator and former Sentry engineering lead Armin Ronacher. The pitch: a terminal coding agent that's more transparent and lighter than Claude Code or Codex — fully open source, MIT-licensed, and free.

What Pi 1.0 Actually Is

Pi was originally built by developer Mario Zechner (creator of libGDX) and acquired by Earendil earlier this year. It's a terminal-based AI agent that reads your repo, edits files, and runs shell commands, connecting to over 15 model providers — Claude, GPT, Gemini, or local models, your choice.

Installing it takes one line:

curl -fsSL https://pi.dev/install.sh | sh
or npm install -g @earendil-works/pi-coding-agent

Run /login to connect your API key or subscription, and you're ready to go. Pi itself is free — you only pay whichever model provider you plug in.

Why It's Different From Claude Code

Zechner built Pi because Claude Code kept shifting under him — system prompts and tool definitions changed with every release, breaking workflows and quietly changing model behavior. Pi's answer was radical minimalism: no MCP, no subagents, no background bash, by design. The goal was to let users see exactly what the agent is thinking and doing, without hidden scaffolding getting in the way.

Pi 1.0's U-Turn: MCP Is Back

Here's the twist: the 1.0 release reverses course and ships native MCP support, plus an experimental "Pi Durable" package that keeps long-running agent tasks alive through disconnects. The full-screen TUI is now the default, and codemode prompt tokens are down roughly 40%. In short, Pi walked back some of its "less is more" purism — but it's still more restrained than Claude Code's growing stack of MCP, skills, swarm mode, and plugins.

Should You Switch From Claude Code

Honestly, no rush. In community benchmarks, Pi used fewer tokens than Claude Code on the same task (558,885 vs. 741,659) but landed slightly behind on verified task completion. Pi also still lacks Claude Code's Plan Mode, multi-agent coordination, and plugin ecosystem — real friction for teams with established workflows.

The practical take: keep Claude Code for daily team work. If you care about cost, want to plug in local models, or just want a tool that doesn't change shape every week, Pi 1.0 is worth installing and trying.

You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士