一条命令,5 分钟,我把 24 万 star 的 DeepSeek Harness 扒了个底朝天。
先说结论:插件架构是真香,但我暂时不敢装在主力机上。
(先声明:这轮云主机是 Linux,没官方版,真机没装;下面是官方文档 + 三篇独立实测交叉核验的,数字都对得上)
先纠正个事儿:今天英文区有人传它是"Google 开源的 Harness²",错了。
这是 DeepSeek 家的,MIT 开源,8 月 13 号发布,10 天涨了 18.6 万 star,现在 24 万+。
设计理念就一句话:everything is a plugin——模型、记忆、沙盒、agent 循环,全是插件,默认就带 152 个,随便换。
夸点是真夸得出口:
安装就一条命令 npx @deepseek-ai/dsh web,本地起 Web UI;v0.2 还出了桌面版(Mac/Win),免 Node 一键装。
Creator mode 更骚:跟它聊天,现场给你搓个插件装上。
dev.to 有人实打实测了一周:用免费的 Qwen3.8-Flash,搓出一个 2300 行的插件,API 花费 $0。
四种模式 Standard/PTC/Minimal/Creative,想省钱的直接开 Minimal 模式磨。
骂点也一点不含糊:
CVE-2026-82533,CVSS 9.4 分。agent 在沙盒里一条 curl 命令,把自己提权到 danger-full-access,审批全关。
补丁发了,但安全公告没发——你的更新 feed 里根本看不到这事儿。
静态装的插件没有沙盒:能碰你的文件系统、shell,还有你配在设置里的 API key。
Linux 没官方版,"本地"指的是软件跑在你机器上,模型调用全走云端,得要 DeepSeek 账号或 API key。
我的判断:
玩 agent loop 的、抠成本的,扔一次性 VM 里闭眼装;
Claude Code/Codex 老用户,先试用别迁移,人家的循环是身经百战的,这边才几周大;
主力机、放生产代码的机器,先别碰。
做 AI/科技内容的蓝V朋友欢迎互关。agent 的下半场到底是拼模型,还是拼谁的插件架子更开放?用过的来说说,我说错了没 👇
🔵 蓝V互关 | Fellow Premium members, let's connect! 欢迎互关,一起交流 AI 与科技 👥
---
English summary:
Tested DeepSeek Harness (not Google's, despite the rumor) — the 242K-star MIT-licensed agent runtime where everything is a plugin.
Pros: one-command install (npx @deepseek-ai/dsh web), desktop app for Mac/Win, Creator mode builds plugins via chat; one tester built a 2,300-line plugin in a week on a free model, $0 API spend.
Cons: CVE-2026-82533 (CVSS 9.4) let a sandboxed agent escalate itself to danger-full-access via one curl; patch shipped with no advisory. Statically installed plugins run unsandboxed. No official Linux build; model calls still need a DeepSeek account/API key.
Verdict: great for tinkerers in a disposable VM; daily-driver Claude Code/Codex users should trial, not migrate.
@elonmusk If the clouds had spare inference at real scale, they'd be selling it — at a price you'd hate. Which means the answer is more Colossus, not renting.
@WorldInMotionHQ Counting lease payments as "cash burn" does a lot of work in that $600BN number. The neoclouds are the ones to watch though — hyperscalers can absorb the gap, for the smaller guys it's existential.
@WorldInMotionHQ Banks going small and task-specific makes sense — cheaper inference, data stays in-house, and regulators sleep better. The interesting part is whether these models ever get evaluated publicly or stay behind the vault door.
3K Followers 3K Following🤖 SI Engineer building tools for cross-border e-commerce.
🗣️ Teaching Mandarin to the world
📦 Finding hidden gems & top products from China.
DM open
586 Followers 637 FollowingWe often tell others about our dreams, record our dreams with words, and realize our dreams with actions.
If one day, we can give dreams to others...
110 Followers 483 FollowingFounder @ https://t.co/aKJus2A4DM
ex Guardrails AI (acquired by Harvey)
ex Agnostiq (acquired by DataRobot)
ex Aida (acquired by Alida)
586 Followers 637 FollowingWe often tell others about our dreams, record our dreams with words, and realize our dreams with actions.
If one day, we can give dreams to others...
110 Followers 483 FollowingFounder @ https://t.co/aKJus2A4DM
ex Guardrails AI (acquired by Harvey)
ex Agnostiq (acquired by DataRobot)
ex Aida (acquired by Alida)