跳到主要內容

Pi 1.0評測:硬剛MCP一年的AI代理終於妥協 | Pi 1.0 Review: The Coding Agent That Hated MCP Ships It

By Kit 小克 | AI Tool Observer | 2026-10-08

🇹🇼 Pi 1.0評測:硬剛MCP一年的AI代理終於妥協

Pi 1.0 是這週開發者圈最熱的話題:這個由 Flask 作者 Armin Ronacher 創辦的維也納新創 Earendil 推出的開源編碼代理,一上架就衝上 Hacker News 第一名,拿下超過 1,600 個讚、600+ 則留言。但真正讓 Pi 1.0 成為焦點的不是功能清單,而是一個大轉彎:Pi 團隊公開砲轟 MCP(Model Context Protocol)超過一年,1.0 版卻原生支援它了。

從「MCP 很浪費 token」到「我們想通了」

Pi 團隊過去的立場很直白:MCP 工具定義平均吃掉 550 到 1,400 個 token,還沒開始做事就先燒掉一大截上下文。這個立場讓不少受夠臃腫代理框架的開發者買單,也讓 Pi 在「極簡主義」這條路線上建立起口碑。

Codemode:技術上的折衷,不是立場投降

Pi 1.0 的解法不是認輸,而是換個做法:透過名為 Codemode 的功能,模型改成在 WASM 沙盒裡寫程式碼去呼叫 MCP 工具,而不是把整份工具 schema 硬塞進上下文視窗。對照原本「MCP 太浪費 token」的抱怨,這其實是技術層面的繞道,而非立場上的妥協。

還有哪些實際能用的東西

  • 延遲載入工具(Deferred tool loading):工具不會一開始就佔用上下文,用到才載入
  • Anthropic 模型的 cache warming:對話中段也能預熱快取,降低重複成本
  • 對話中途插入 system message:不必重開對話就能換提示詞或工具組合
  • Pi Durable(實驗性):讓長時間執行的代理具備當機復原、重放策略,可以跨終端機、Slack 等介面接續工作

背後團隊:Flask 作者的新創 Earendil

Earendil 由 Armin Ronacher(Flask 創辦人、Sentry 前工程主管)與 Colin Daymond Hanna 共同創立,今年春天收購了 Pi 的原始專案,原作者 Mario Zechner 仍是股東之一。整個 Pi 1.0 專案是 MIT 授權,一行指令就能安裝:curl -fsSL https://pi.dev/install.sh | sh。

對已經在用 Claude Code、Codex CLI 之類終端代理的開發者來說,Pi 1.0 值得裝起來比較看看——它的賣點一直是「等別人證明過了,我們才跟進」,這次對 MCP 的轉彎,剛好是這套哲學最好的示範。

好不好用,試了才知道。


🇺🇸 Pi 1.0 Review: The Coding Agent That Hated MCP Ships It

Pi 1.0 is this week's biggest developer-tool story. The open-source coding agent from Earendil — the Vienna startup founded by Flask creator Armin Ronacher — shot to #1 on Hacker News with over 1,600 points and 600+ comments. But the real headline behind Pi 1.0 isn't the feature list — it's the reversal: Pi publicly criticized MCP (Model Context Protocol) for over a year, and version 1.0 ships native support for it.

From "MCP wastes tokens" to "we found a workaround"

Pi's team made a consistent argument: loading MCP tool schemas into context costs 550 to 1,400 tokens per tool, before the model does anything useful. That stance earned Pi a loyal following among developers tired of bloated agent harnesses.

Codemode: a technical compromise, not a surrender

Pi 1.0 doesn't walk back that critique — it routes around it. A feature called Codemode lets the model write code inside a WASM sandbox to call MCP tools, instead of stuffing full tool schemas into the context window. Measured against the original complaint, this is a technical workaround, not a change of position.

What else actually shipped

  • Deferred tool loading — tools load on demand instead of sitting in context from the start
  • Cache warming for Anthropic models — keeps costs down even mid-conversation
  • Mid-conversation system messages — swap prompts or tool sets without restarting the session
  • Pi Durable (experimental) — long-running agents with crash recovery and replay policies, usable from terminal, Slack, and other surfaces

The team behind it: Flask's creator, again

Earendil was co-founded by Armin Ronacher (Flask creator, former Sentry engineering lead) and Colin Daymond Hanna, and acquired the original Pi project this spring — original creator Mario Zechner remains a shareholder. Pi 1.0 is MIT licensed and installs with one line: curl -fsSL https://pi.dev/install.sh | sh.

If you're already running Claude Code, Codex CLI, or similar terminal agents, Pi 1.0 is worth a side-by-side test. Its whole philosophy has been "wait until something proves itself before adopting it" — and this MCP U-turn is the best demonstration of that philosophy in action.

As always: you won't know if it's good until you try it yourself. (好不好用,試了才知道)

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

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code