dsh-tools-mcp

dsh-tools-mcp

MCP server wrapping DeepSeek's agent harness tools into minimal/standard modes, enabling persistent shell, file editing, search, job management, and more via natural language.

Category
Visit Server

README

dsh-tools-mcp

将 deepseek-harness(DeepSeek AI 开源的 agent harness)中 极简模式(minimal) 与 标准模式(standard) 的 tools 提取并包装为 MCP(Model Context Protocol) 服务器,通过 --mode 参数控制暴露哪套工具集。

工具 schema、描述文案、输出信封(如 read 的 <path>/<type>/<content> 格式、str_replace_editor 的 view/create/str_replace/insert 语义、[exit code: N] 标记)均忠实还原自原仓库 apps/cli/config/agent-presets/{minimal,standard}/agent.cordis.yml 及其对应 packages/*/tool-* 实现。

模式与工具对照

模式 参数 工具
极简模式 --mode minimal 持久 shell(POSIX 为 bash,win32 为 pwsh,状态跨调用保持)+ str_replace_editor
标准模式 --mode standard bash/pwsh(单次执行)、read、write、edit、read_image、glob、grep、job_output/job_list/job_kill、todo_write、ask_user_question、web_search、get_goal/create_goal/update_goal、plan_mode/exit_plan_mode、skill

标准模式还支持 --with-web-fetch 额外暴露 web_fetch(原版 standard 预设默认 fetch: false)。

快速开始

cd dsh-tools-mcp
npm install
npm run build          # 输出到 dist/

# 极简模式
node dist/index.js --mode minimal --cwd D:/develop/project/dsh-tools-mcp

# 标准模式
node dist/index.js --mode standard --cwd D:/develop/project/dsh-tools-mcp

MCP 客户端配置

// mcp.json / WorkBuddy 自定义连接器
{
  "mcpServers": {
    "dsh-minimal": {
      "command": "node",
      "args": ["D:/develop/project/dsh-tools-mcp/dist/index.js", "--mode", "minimal", "--cwd", "D:/develop/project/dsh-tools-mcp"]
    },
    "dsh-standard": {
      "command": "node",
      "args": ["D:/develop/project/dsh-tools-mcp/dist/index.js", "--mode", "standard", "--cwd", "D:/develop/project/dsh-tools-mcp"]
    }
  }
}

两个入口可以同时注册,按需选择极简/标准工具集。

全部参数

--mode <minimal|standard>    Agent 预设模式(默认 minimal)
--cwd <path>                 工作区根目录,文件工具基于它解析相对路径(默认 process.cwd())
--shell <auto|bash|pwsh>     shell 后端;auto = win32 用 pwsh,其余用 bash
--shell-timeout-ms <n>       持久 shell 单条命令超时(默认 300000)
--max-output-chars <n>       输出截断上限(默认 16000)
--skills-dir <path>          启用 skill 工具,扫描该目录下的 SKILL.md 子目录
--search-provider <duckduckgo|tavily|bing>   web 搜索后端(默认 duckduckgo)
--search-api-key <key>       tavily/bing 的 API key
--proxy <url>                web 工具的 HTTP(S) 代理,如 http://127.0.0.1:7897
--with-web-fetch             额外暴露 web_fetch
--allow-outside-root         允许文件工具访问 --cwd 之外的路径(默认禁止)
--tools <a,b,c>              显式指定工具/套件清单覆盖模式(见下)
--help / --version

--tools 覆写

--tools 接受工具名或套件名,按需暴露子集:

# 只要文件相关
node dist/index.js --mode standard --tools read,write,edit,glob,grep

# 只要极简模式的编辑器 + 后台任务
node dist/index.js --mode minimal --tools str_replace_editor,job_output,job_list,job_kill

套件名:shell(持久 shell)、one-shot、fs、search、jobs、todo、web、ask-user、goal、plan、skill。

设计说明

  • 持久 shell:无 node-pty 依赖,用 stdio 管道驱动长驻 bash/powershell 子进程;命令以随机 UUID 标记包裹以可靠捕获完成状态与退出码;cd/export 等状态跨调用保持;超时或 shell 退出时自动重置(与原始 tool-bash-persistent 语义一致)。
  • 路径安全:默认文件工具仅允许访问 --cwd 工作区内的路径,越界报错提示加 --allow-outside-root。
  • web_search 默认 DuckDuckGo(免 key);国内网络建议 --proxy http://127.0.0.1:7897,或换 --search-provider tavily --search-api-key <key>。
  • 状态均为进程内(session 级):todo/goal/plan/后台任务状态随服务器进程生命周期,重启即清空——与原版"会话级状态"等价。

与原版 dsh 的差异(明确说明)

项 原版 dsh 本 MCP 包装
持久 shell node-pty PTY 终端 stdio 管道长驻进程(无交互提示符,状态保持等价)
ask_user_question dsh UI 提问桥 MCP stdio 下无法同步交互,返回问题文本并指示模型转述给用户
plan_mode/exit_plan_mode UI 审批流 服务端标志位 + 计划文本回传
subagent_*/workflow/ralph 依赖 dsh 运行时(LLM 子代理、工作流引擎) 未包装——需完整 dsh 运行时,详见下方
grep 引擎 内置 ripgrep 二进制 node 正则实现(常见语法等价,部分 ripgrep 特性不支持)

未包装的工具(原版 standard preset 中的运行时耦合工具,需完整 dsh runtime):subagent_control、list_agents、subagent、subagent_fork、workflow、ralph。如需这些能力,请直接使用 deepseek-harness 本体。

测试

npm run build
npm test          # 冒烟测试:两种模式工具暴露 + 核心工具行为 + --tools 覆写
node scripts/state-test.mjs   # 持久 shell 状态保持 / 退出码 / 优雅重置(PROXY=... 可测 web_search)
node scripts/bash-test.mjs    # bash 后端验证

源码参考

提取自 deepseek-ai/deepseek-harness(MIT License,master 分支):

  • 模式定义:apps/cli/config/agent-presets/minimal/agent.cordis.yml、.../standard/agent.cordis.yml
  • 工具实现:packages/shell/tool-bash-persistent、packages/fs/tool-str-replace-editor、packages/fs/tool-fs、packages/fs/tool-fs-search、packages/jobs/tool-jobs、packages/todo/tool-todo、packages/web/tool-web、packages/interaction/tool-ask-user、packages/goal/tool-goal、packages/plan/plan-mode、packages/skill/tool-skill 等

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured