第5名 DeepSeek-Reasonix：一个可以扔在那自己跑的编程 agentNo. 5 DeepSeek-Reasonix: A Coding Agent You Can Leave Running on Its Own
星数 35,662，兼容工具榜第 5，MIT 协议。
35,662 stars, No. 5 on the compatible tools list, MIT license.
星数 35,662，兼容工具榜第 5，MIT 协议。
它是什么
DeepSeek-Reasonix 是一个开源的终端编程 agent，以单个 Go 二进制文件分发。它的核心卖点写在开头：一个你可以扔在那让它自己跑的编程 agent。围绕这个目标，它把"长时间自主运行时别失控、能随时撤销"这几件事做在了底层。
同一个本地引擎，有四种用法：终端、桌面应用、浏览器、通过 ACP 协议接进编辑器。底层是同一个，不用重复装。
能干什么
- 规划模式：先出计划再动手
- 权限控制：哪些操作能做、哪些要确认
- 工作区沙盒：动作限制在工作区里
- 每轮检查点：随时查看进度，能撤销不想要的改动
- 双模型分工：两个模型分别当规划者和执行者，各自跑在缓存稳定的会话里
- MCP 服务器扩展工具、提示词、资源
- Extension Protocol v1 侧车插件：拦截运行时事件、加自定义提供器和结构化界面
长任务的上下文它也专门管：启动时只注入一小段稳定的环境摘要，过期的工具输出在压缩总结前先剪掉，保证上下文窗口在长时间运行里不被撑爆。
兼容 DSH 到什么程度
它是独立的 agent 框架，预置了 DeepSeek 模型，也能把任何兼容 OpenAI 格式的接口当成一个配置项填进去，不用改代码，同时集成了 DSH 在内的多个编程平台。跟第 1 名类似属于"自己就是完整产品"那一类，区别在它专注终端编码和长任务，不做设计产出。
谁适合用
- 要处理大型、复杂、耗时长的编码任务，又不想一直守在屏幕前逐步批准的开发者
- 喜欢终端工作流，同时偶尔想在编辑器或桌面应用里用同一个引擎的人
- 想拿两个模型分工、做规划与执行分离的团队
分发几乎零摩擦：CGO 关闭后的单文件静态二进制，交叉编译一条命令出六个平台的版本。安装可以走 npm 全局装，macOS 也有 Homebrew。第一次用得先按官方文档把模型端点配好，这一步省不掉。
下面是它的官方标识：
图为 DeepSeek-Reasonix 官方标识，标识里的鲸形图案呼应 DeepSeek 生态。
推荐理由
前五里它最适合"放着不管"这个场景。规划模式、权限、沙盒、每轮检查点这四样凑在一起，让长时间自主运行变成一件你事后还能读、能撤的事，而不是开盲盒。如果你的活儿是那种一开跑就很久的大型编码任务，它比需要人盯着的交互式 agent 更对口。它是第五名，但在"长任务"这条窄赛道上，前五里没有谁比它更专。
数据来源：DSH 插件市场兼容工具榜及 esengine/DeepSeek-Reasonix 仓库，星数为 2026 年 9 月 22 日快照。
35,662 stars, No. 5 on the compatible tools list, MIT license.
What it is
DeepSeek-Reasonix is an open-source terminal coding agent, distributed as a single Go binary. Its core pitch is right there in the opening line: a coding agent you can leave running on its own. Around that goal, it builds the hard parts into the foundation — staying under control during long autonomous runs, and being able to undo things at any time.
The same local engine has four ways to be used: terminal, desktop app, browser, and plugged into an editor via the ACP protocol. The foundation is the same, so there's nothing to install twice.
What it can do
- Plan mode: produce a plan before taking action
- Permission control: which operations are allowed and which need confirmation
- Workspace sandbox: actions are confined to the workspace
- Per-round checkpoints: check progress at any time and undo unwanted changes
- Dual-model division of labor: two models serve as planner and executor respectively, each running in its own session with stable caching
- MCP servers extend tools, prompts, and resources
- Extension Protocol v1 sidecar plug-ins: intercept runtime events, add custom providers and structured interfaces
It also manages context specifically for long tasks: at startup it injects only a short, stable environment summary, and stale tool output is trimmed before compression and summarization, keeping the context window from blowing up during long runs.
How compatible it is with DSH
It is a standalone agent framework, preconfigured with DeepSeek models, and it can also take any OpenAI-compatible endpoint as a configuration entry with no code changes, while integrating multiple coding platforms including DSH. Like No. 1, it falls into the "it is itself a complete product" category; the difference is that it focuses on terminal coding and long tasks, and doesn't produce design work.
Who it's for
- Developers handling large, complex, long-running coding tasks who don't want to sit at the screen approving every step
- People who like a terminal workflow but occasionally want the same engine in an editor or desktop app
- Teams that want two models dividing the work, separating planning from execution
Distribution is nearly frictionless: a single-file static binary with CGO disabled, and one cross-compilation command produces builds for six platforms. Installation can go through a global npm install, and macOS also has Homebrew. The first run requires configuring the model endpoint per the official docs — that step can't be skipped.
Below is its official logo:
The image shows the official DeepSeek-Reasonix logo, whose whale-shaped mark echoes the DeepSeek ecosystem.
Why it's recommended
Among the top five it fits the "leave it alone" scenario best. Plan mode, permissions, sandbox, and per-round checkpoints together turn long autonomous runs into something you can still read and roll back afterward, rather than opening a mystery box. If your work consists of large coding tasks that run for a long time once started, it suits you better than an interactive agent that needs a human watching. It ranks fifth, but on the narrow track of "long tasks," none of the top five is more specialized.
Data sources: the DSH Plugin Hub compatible tools list and the esengine/DeepSeek-Reasonix repository; star counts are a snapshot from September 22, 2026.