跳到主要內容

DeepSeek Harness評測:一週衝21萬星的外掛式代理框架 | DeepSeek Harness Review: 210K Stars, Plugin-Based AI Agent

By Kit 小克 | AI Tool Observer | 2026-09-06

🇹🇼 DeepSeek Harness評測:一週衝21萬星的外掛式代理框架

DeepSeek Harness 是 DeepSeek 在 2026 年 8 月中釋出的開源代理框架(agent harness),主打「萬物皆外掛」(Everything is a Plugin)架構,上線一週就衝上超過 21 萬顆 GitHub 星星、2.5 萬次 Fork,是近期開發者工具圈討論度最高的話題。這篇文章帶你看懂 DeepSeek Harness 解決了什麼問題、跟 Claude Code、Codex 差在哪,以及實測後該注意的地雷。

DeepSeek Harness 是什麼?

DeepSeek Harness(簡稱 dsh)是一套 MIT 授權的開源 AI 編碼代理執行環境,把工具、技能、對話 session,甚至整套代理框架本身都設計成可拆換的外掛模組,開發者能自由組合、替換。

它建立在 DeepSeek 自研的 Cordis 架構上,所有設定都寫在可編輯的 YAML 檔案裡,不藏在隱藏選單中,預設會開一個本機網頁介面(127.0.0.1:3080)方便操作。官方目前定位是「開發者預覽版」,代表功能和行為之後還會有大改動。

DeepSeek Harness 跟 Claude Code、Codex 差在哪?

最大差別在於「誰決定工作流程」。Claude Code 和 Codex 都是廠商鎖定好的固定流程,DeepSeek Harness 反過來把主導權交給開發者,甚至能把 Claude Code、Codex 當成「子代理」塞進 DeepSeek 的外層編排裡跑。

  • 模型彈性:預設接 DeepSeek V4 Flash / Pro,因為是開源 MIT 授權,也能接其他家 API 或自架模型
  • 設定透明:全部走 YAML 檔案,方便版本控管與客製化
  • 定位不同:與其說是 Claude Code 的競品,更像是包在各家代理工具外層的「編排層」

實測缺點:真的比較省嗎?

不一定。早期實測案例顯示,光是做一個簡單的 ISS 追蹤網頁,兩輪對話就燒掉約 2000 萬 tokens,成本並不便宜;目前仍得從原始碼安裝,介面也還很粗糙,隨時可能大改版。想控制成本的話,V4 Flash 目前的性價比評價優於 Pro。

好不好用,試了才知道。


🇺🇸 DeepSeek Harness Review: 210K Stars, Plugin-Based AI Agent

DeepSeek Harness is DeepSeek's open-source agent harness, released mid-August 2026 under an "Everything is a Plugin" architecture. It shot past 210,000 GitHub stars and 25,000 forks within a week, making it one of the hottest developer-tool stories right now. Here's what DeepSeek Harness actually does, how it differs from Claude Code and Codex, and what to watch out for before you try it.

What is DeepSeek Harness?

DeepSeek Harness (dsh) is an MIT-licensed, open-source runtime for AI coding agents that treats tools, skills, sessions, and even entire agent harnesses as swappable plugins developers can mix and match.

It's built on DeepSeek's own Cordis architecture. Every setting lives in editable YAML files instead of hidden menus, and it ships with a local web UI at 127.0.0.1:3080 by default. It's officially a "developer preview," so expect breaking changes ahead.

How does DeepSeek Harness compare to Claude Code and Codex?

The core difference is who controls the workflow. Claude Code and Codex ship fixed, vendor-defined pipelines. DeepSeek Harness flips that — it hands control to the developer, and can even run Claude Code or Codex as sub-agents inside its own orchestration layer.

  • Model flexibility: defaults to DeepSeek V4 Flash/Pro, but being MIT-licensed, it can plug into other providers or self-hosted models
  • Transparent config: everything lives in YAML, making it easy to version and customize
  • Different positioning: less a Claude Code competitor, more an orchestration layer that sits above other coding agents

The catch: does it actually save money?

Not necessarily. Early testers reported burning roughly 20 million tokens across two turns just to build a simple ISS tracker app — not cheap. It still requires installing from source, the interface is rough, and breaking changes are likely. If cost matters, V4 Flash currently gets better cost-to-value reviews than Pro.

好不好用,試了才知道。You won't know until you try it.

Sources / 資料來源

常見問題 FAQ

DeepSeek Harness 是免費的嗎?

是的,DeepSeek Harness 採 MIT 授權完全開源免費,但呼叫模型 API(如 DeepSeek V4)仍需依用量付費。

DeepSeek Harness 可以取代 Claude Code 嗎?

不完全是。它定位比較像「編排層」,甚至可以把 Claude Code 當子代理呼叫,適合想自訂工作流程的進階開發者,而非直接替代品。

DeepSeek Harness 現在能穩定用在正式專案嗎?

目前官方定位是開發者預覽版,介面與 API 都可能大改,建議先在側專案或測試環境試用,暫不建議直接上正式產線。

DeepSeek Harness 支援哪些模型?

預設支援 DeepSeek V4 Flash / Pro,但因為是開源 MIT 授權架構,理論上可以串接其他家 API 或自架的本地模型。

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