appbridge MCP server
Commit Nov 7, 2026Provides AI agents with a structured toolkit to interact with local files, office documents, system commands, and Obsidian vaults.
Also known as stevenwang15/appbridge
- Stars
- 1
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 7, 2026
- License
- MIT
What is appbridge MCP?
appbridge provides AI agents with a structured toolkit to interact with local files, office documents, system commands, and Obsidian vaults.
Summary generated from the project's README
About appbridge MCP server
Developers can provide AI agents with direct control over local computing environments by using appbridge to transform files, system commands, and knowledge bases into a structured set of actionable tools. Instead of forcing an agent to manually construct complex shell strings, this server provides a predefined verb list of 48 tools that include file system operations, git commands, and Obsidian vault management. It works as a neutral MCP server compatible with Claude Desktop, Cursor, and other MCP clients, or as a specialized plugin for DeepSeek Harness.
Users can manage local office documents like docx, xlsx, and pptx without needing Microsoft Office installed, as the server uses Node.js built-in modules for parsing. The system also supports running complex, multi-step workflows called recipes that can execute tasks in parallel or in the background. Security is managed through a policy configuration that defines specific allowed paths, executable commands, and network hosts, ensuring the agent operates within defined boundaries.
What you can do with it
- 1Execute complex multi-step workflowsA developer can hand an agent a recipe to search a vault and save the results to a new file automatically.
Written from the project's README. Every use case is checked against a quote from it.
Configuration
APPBRIDGE_PLATFORM
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablestevenwang15/appbridge is public and reachablegithub.com/stevenwang15/appbridge ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-07github.com/stevenwang15/appbridge ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubappbridge
给 agent 装上手脚:把电脑上的东西变成它能直接调用的工具,把干活的步骤存成能反复跑的配方。
电脑上的能力本来是散的:文件在硬盘里、命令在终端里、知识库在 Obsidian 里、文档是 docx。 appbridge 把它们统一成一张动词表——每个工具一个名字、一组参数、一个确定的返回。
- 零外部依赖:只用 Node 内置模块,一个
node就能跑 - 两种外壳:中立 MCP server(任何宿主都能接)+ DeepSeek Harness 插件
- 内核与插件分开:插件是薄壳,逻辑全在内核,换宿主不用重写
48 个工具 · 10 种工具类型 · 7 个配方 · 32 个测试 · 176 KB
它解决三件事
一、让 AI 真的能动手。 它本来只能说话。接上这 48 个工具,它能读笔记库、跑 git、解析 Word 和 PPT、 调本机模型、截屏、发通知——这些原来只有人在终端里敲得出来的动作,现在 AI 能直接调。
二、让重复的活儿只做一次。 一次跑通,存成配方;下次一条命令,不重新推理,也不会每次做得不一样。
三、接一个新应用 = 加一段 JSON。 不用写代码。有命令行填 cli,有本地端口填 http,有数据目录填 fs。
为什么是「动词表」而不是让 AI 现拼命令
不是不能跑 shell,是跑 shell 会一直在低水平上出错。
| 让 AI 现拼 shell 命令 | 给它一张动词表 | |
|---|---|---|
| 每次执行 | 现场拼字符串,引号、空格、转义都可能错 | 名词化调用,参数是结构化的 |
| 出错时 | 命令写错,整条废掉,得重试 | 缺参数在调用前就报清楚,改一个参数重跑 |
| 返回结果 | 一段文本,还要再解析 | 结构化的值,下一步能直接引用 |
| 复现 | 每次重新拼,可能不一样 | 写进配方,顺序固定,可重放 |
| 加新能力 | 教它一套新命令的写法 | 加一段 JSON 声明 |
一句话:动词表让 AI 把注意力花在「干什么」上,而不是「这条命令怎么写」。
快速开始
git clone https://github.com/stevenwang15/appbridge.git
cd appbridge
node bin/cli.js doctor # 体检:这台机器上哪些工具能用
node bin/cli.js tools # 48 个工具
node bin/cli.js call sys.info # 直接调一个
node bin/cli.js run vault-search --set "q=适当性" # 跑配方
doctor 检查四件事:平台对不对、命令装没装、端口通不通、路径存不存在。
cp config.json config.local.json # 然后改里面的 vault / model
三种用法
1. 命令行
node bin/cli.js doctor
node bin/cli.js call docx.read --set path=报告.docx
node bin/cli.js run parallel-demo --async # 后台跑
node bin/cli.js jobs / wait <id> / kill <id>
2. MCP server(中立外壳,推荐)
任何支持 MCP 的宿主都能接,不依赖宿主版本:
{
"mcpServers": {
"appbridge": {
"command": "node",
"args": ["/absolute/path/to/appbridge/bin/mcp.js"]
}
}
}
DeepSeek Harness / Claude Desktop / Cursor / 任何 MCP 客户端通用。接上就是 48 个工具。
3. DeepSeek Harness 插件
npm install appbridge-dsh-plugin
插件把内核工具注册成 DSH 原生工具(ctx.tools.register + defineTool),实现见
packages/dsh-plugin。
插件壳依赖宿主提供的
@deepseek-ai/dsh-tools。如果你的 DSH 版本不暴露它, 用上面的 MCP 方式接入,功能完全一样。
48 个工具
Office(不装 Office 也能用) — docx/xlsx/pptx 本质是 zip + XML,用 Node 内置 zlib 自己解析
docx.read docx.write xlsx.read pptx.read
压缩包 / 表格 zip.list zip.read zip.extract csv.read
文件 fs.read fs.write fs.append fs.list fs.find fs.search
Obsidian obs.list obs.search obs.read obs.write obs.append obs.tags obs.backlinks
命令 / 版本库 cli.run git.log git.status git.diff git.branch git.show
系统 sys.info sys.disk sys.process sys.screenshot sys.notify sys.say clip.read clip.write app.open
图片 img.info img.resize textutil.convert
本机模型 / 网络 ollama.models ollama.chat http.get
数据加工 text.join json.get echo delay
加一个工具 = 往 tools.json 加一段 JSON,不用写代码:
"git.status": {
"desc": "git 工作区改动",
"kind": "cli",
"command": "git",
"args": ["status", "--short"],
"cwd": "{{repo}}",
"params": { "repo": { "type": "string", "default": "." } }
}
配方:步骤的顺序
{
"name": "vault-search",
"when": "在 Obsidian 库里搜关键词",
"args": { "q": { "type": "string", "required": true } },
"steps": [
{ "id": "hits", "tool": "obs.search", "args": { "pattern": "{{q}}" } },
{ "id": "body", "tool": "text.join", "args": { "items": "{{$1}}", "sep": "\n" } },
{ "id": "save", "tool": "fs.write", "args": { "path": "out/search.md", "content": "{{$2}}" } }
]
}
变量:{{参数}}、{{$1}}(第 1 步输出)、{{date}} {{time}} {{stamp}} {{cwd}}
配方是数据不是脚本:agent 能填参数生成,人能直接改,每一步都能单测。
异步执行
配方内并行
{ "id": "slow", "parallel": [
{ "id": "a", "tool": "delay", "args": { "ms": 800 } },
{ "id": "b", "tool": "delay", "args": { "ms": 800 } } ] }
实测 3×800ms:顺序 2.44s,并行 0.84s。
整条配方后台跑 — --async 31ms 返回任务号,状态落盘 jobs/<id>/,可跨进程查、等、杀。
调用范围
policy.json 决定这些工具能碰到哪里。它是配置,不是目的——
想让 AI 多干点,往范围里加;想收窄,删掉。
{
"base": ".",
"allowExec": ["git", "node", "osascript", "screencapture", "sips", "ps", "textutil", "tesseract", "blender"],
"allowRead": ["./**", "{{config.vault}}/**"],
"allowWrite": ["./**", "{{config.vault}}/**"],
"allowNet": ["127.0.0.1", "localhost"]
}
- 路径:
path按读/写判定,src只读,dest只写 - 命令:必须在
allowExec里 - 网络:host 必须在
allowNet里
base 固定相对程序目录,不跟 cwd 走。
说明:
cli类工具的范围是「跑哪个命令」,命令自己往哪写由那个命令决定。这一层管的是能力边界,不是系统沙箱。
可移植性
| 文件 | 放什么 | 进版本库 |
|---|---|---|
config.json |
可移植默认值 | 是 |
config.local.json |
本机差异(知识库路径、模型名) | 否 |
环境变量:APPBRIDGE_VAULT APPBRIDGE_MODEL APPBRIDGE_PLATFORM APPBRIDGE_POLICY
平台差异自动降级:macOS 专属工具标了 "platform": ["darwin"]。
APPBRIDGE_PLATFORM=linux node bin/cli.js doctor # 38/48
APPBRIDGE_PLATFORM=win32 node bin/cli.js doctor # 36/48
仓库结构
appbridge/ 内核(npm: appbridge-core)
├── bin/cli.js 命令行入口
├── bin/mcp.js MCP server 入口(中立外壳)
├── src/invoke.js 工具调用器
├── src/recipe.js 配方执行器
├── src/jobs.js 后台任务
├── src/doctor.js 体检
├── src/zip.js ZIP 读写
├── src/office.js docx / xlsx / pptx
├── src/mcp.js MCP 工具描述
├── src/config.js 配置合并
├── tools.json 48 个工具
├── policy.json 调用范围
├── recipes/ 配方
└── packages/dsh-plugin/ DSH 插件壳(薄)
测试:32 个(内核 25 + 插件契约 7)
开发者
node --test test/ # 内核测试
node --test packages/dsh-plugin/test/ # 插件契约测试
License
MIT
appbridge MCP questions
What kind of tools does appbridge provide?
It offers 48 tools across 10 types, including Office document parsing (docx, xlsx, pptx), file system operations, Obsidian integration, git commands, and system utilities like screenshots and notifications.