Microsoft MCP
Model Context Protocol servers for Microsoft 365, enabling AI assistants to interact with Calendar, Contacts, OneDrive, Outlook, and SharePoint via the Microsoft Graph API.
README
<div align="center">
<img src=".github/assets/icons/outlook.png" width="54" alt="Outlook" /> <img src=".github/assets/icons/calendar.png" width="54" alt="Calendar" /> <img src=".github/assets/icons/onedrive.png" width="54" alt="OneDrive" /> <img src=".github/assets/icons/sharepoint.png" width="54" alt="SharePoint" /> <img src=".github/assets/icons/contacts.png" width="54" alt="Contacts" />
<h1>Microsoft MCP</h1>
<p> <b>Model Context Protocol servers for Microsoft 365.</b><br/> Calendar · Contacts · OneDrive · Outlook · SharePoint — on the official <a href="https://www.npmjs.com/package/@modelcontextprotocol/sdk"><code>@modelcontextprotocol/sdk</code></a>, over stdio or Streamable HTTP. </p>
</div>
Each server speaks the real MCP protocol and runs over either transport:
- stdio — for local MCP clients that launch the server as a subprocess (Claude Desktop, IDEs, the MCP Inspector).
- Streamable HTTP — for remote/hosted use, with the Microsoft Graph access token supplied per request via
Authorization: Bearer.
Servers
| Server | npm package (and binary) | Tools | |
|---|---|---|---|
| <img src=".github/assets/icons/calendar.png" width="22" alt=""/> | Calendar | ms-calendar-mcp |
9 |
| <img src=".github/assets/icons/contacts.png" width="22" alt=""/> | Contacts | ms-contacts-mcp |
7 |
| <img src=".github/assets/icons/onedrive.png" width="22" alt=""/> | OneDrive | ms-onedrive-mcp |
9 |
| <img src=".github/assets/icons/outlook.png" width="22" alt=""/> | Outlook | microsoft-outlook-mcp |
14 |
| <img src=".github/assets/icons/sharepoint.png" width="22" alt=""/> | SharePoint | ms-sharepoint-mcp |
23 |
All tools are thin wrappers over the Microsoft Graph v1.0 API.
Layout
microsoft-mcp/
├── apps/ # one MCP server per Microsoft 365 product
│ ├── calendar/
│ ├── contacts/
│ ├── onedrive/
│ ├── outlook/
│ └── sharepoint/
│ └── src/
│ ├── tools.ts # declarative tool definitions (schema + handler)
│ └── index.ts # run({ name, version }, tools)
└── packages/ # shared building blocks
├── core/ # MCP server bootstrap + dual transport (stdio / HTTP)
├── graph/ # Microsoft Graph HTTP client
├── validation/ # id / path / query sanitizers
└── logger/ # structured JSON logging (stderr-only — stdio-safe)
A server is just a list of tools handed to run():
// apps/calendar/src/index.ts
import { run } from "@microsoft-mcp/core";
import { tools } from "./tools.js";
void run({ name: "microsoft-calendar", version: "1.0.0", title: "Microsoft Calendar" }, tools);
// a single tool
defineTool({
name: "get_event",
description: "Get a single calendar event by ID.",
inputSchema: { event_id: z.string().describe("Event ID") },
confirmationPolicy: "never",
handler: ({ graph }, { event_id }) => {
validateId(event_id, "event_id");
return graph.request("GET", `/me/events/${event_id}`);
},
});
confirmationPolicy ("always" for mutating/destructive tools, "never" for read-only) is surfaced to clients as MCP readOnlyHint / destructiveHint annotations.
Requirements
- Node.js >= 20
- pnpm 10 (
corepack enable)
Setup
pnpm install
pnpm build # build all servers (turbo) -> apps/*/dist/index.js
pnpm check-types # typecheck everything
Tests & CI
pnpm test # run the vitest suite once
pnpm test:watch # watch mode
pnpm test:coverage # run with a v8 coverage report (-> coverage/)
Tests live next to the code as *.test.ts and run on TypeScript source directly (no build step). The shared packages/* are covered by unit and integration tests — including a full Streamable-HTTP round-trip against a live server — and CI enforces a coverage floor on them. Each apps/* server ships an invariant suite that locks its tool surface (unique snake_case names, valid schemas and confirmation policies).
Every push and pull request to master runs CI: typecheck → build → tests with coverage. The coverage badge is regenerated from the run.
Authentication
You sign in once with your Microsoft account; the server then caches a refresh token and acquires access tokens silently from then on — no pasting, no 1-hour expiry. Sign-in uses your own Microsoft Entra ID app registration (free) so the servers act on your behalf.
1. Register an Entra ID app (one time)
-
Azure Portal → Microsoft Entra ID → App registrations → New registration. Name it anything; pick the Supported account types that fit (single-tenant, multi-tenant, and/or personal accounts).
-
Authentication → Add a platform → Mobile and desktop applications → add redirect URI
http://localhost, and set Allow public client flows to Yes (enables the--device-codefallback). -
API permissions → Add a permission → Microsoft Graph → Delegated permissions → add the scopes for the servers you use (then Grant admin consent if your tenant requires it):
Server Delegated scopes Calendar Calendars.ReadWriteContacts Contacts.ReadWriteOneDrive Files.ReadWrite.AllOutlook Mail.ReadWrite,Mail.SendSharePoint Sites.ReadWrite.AllAll servers also use
User.Read. (offline_accessis requested automatically for refresh.) -
Copy the Application (client) ID.
2. Sign in (one time per machine)
Set MICROSOFT_CLIENT_ID, then run the server's login command. A browser opens; after you consent, the token is cached under ~/.config/microsoft-mcp/:
export MICROSOFT_CLIENT_ID=<your-client-id>
npx -y ms-calendar-mcp login # opens the browser
npx -y ms-calendar-mcp login --device-code # headless: shows a code to enter
From then on the server refreshes tokens automatically. Use a non-default tenant with MICROSOFT_TENANT_ID (default common).
Advanced: supply your own token
To bypass the built-in flow, supply a pre-acquired Graph token directly:
- stdio: set
MICROSOFT_ACCESS_TOKEN(takes precedence over the cached sign-in). Good for quick tests — mint one withaz account get-access-token --resource https://graph.microsoft.com --query accessToken -o tsv. - HTTP: send
Authorization: Bearer <token>on eachPOST /mcprequest. Each request is stateless with its own token, so callers never share credentials — this is the model for hosted/remote deployments, which handle their own auth.
Running
stdio (e.g. Claude Desktop)
Each server is published to npm and runnable with npx — no clone or build. Sign in once first (npx -y ms-calendar-mcp login, see Authentication), then:
// claude_desktop_config.json
{
"mcpServers": {
"microsoft-calendar": {
"command": "npx",
"args": ["-y", "ms-calendar-mcp"],
"env": { "MICROSOFT_CLIENT_ID": "<your-client-id>" }
}
}
}
Or point at a local build instead of npm:
{
"command": "node",
"args": ["/abs/path/microsoft-mcp/apps/calendar/dist/index.js"],
"env": { "MICROSOFT_CLIENT_ID": "<your-client-id>" }
}
During development you can skip the build and run the TypeScript directly:
MICROSOFT_ACCESS_TOKEN=<token> pnpm --filter ms-calendar-mcp dev
Streamable HTTP
# build first, then:
PORT=3000 node apps/calendar/dist/index.js --http
# or, in dev:
pnpm --filter ms-calendar-mcp dev -- --http --port 3000
The server exposes POST /mcp (the MCP endpoint) and GET /healthz. Point any Streamable-HTTP MCP client at http://localhost:3000/mcp with an Authorization: Bearer header.
Transport selection
Resolved in this order: --stdio / --http flag → MCP_TRANSPORT=stdio|http → default stdio.
HTTP port: --port <n> → PORT → 3000.
Environment variables
| Variable | Used by | Description |
|---|---|---|
MICROSOFT_CLIENT_ID |
stdio | Entra ID app (client) ID for sign-in. Required for the login flow. |
MICROSOFT_TENANT_ID |
stdio | Tenant for sign-in: common (default), organizations, consumers, or a tenant ID. |
MICROSOFT_ACCESS_TOKEN |
stdio | Pre-acquired Graph token; overrides the cached sign-in when set. |
MICROSOFT_MCP_CACHE_DIR |
stdio | Override the token-cache directory (default ~/.config/microsoft-mcp). |
MCP_TRANSPORT |
both | stdio (default) or http. |
PORT |
http | Listen port (default 3000). |
MCP_HTTP_BODY_LIMIT |
http | Max request body size (default 50mb) for base64 uploads. |
MCP_DEBUG |
both | Any non-empty value enables debug logs (to stderr). |
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。