outlook-mcp

An MCP server that gives AI agents structured access to a Microsoft Outlook mailbox through the Microsoft Graph API — with one deliberate constraint at its core:

Agents read and draft. Only humans send.

There is no send tool in this server, and the OAuth token it holds does not include the Mail.Send scope. Even a fully compromised or badly prompted agent cannot send email through it — the capability doesn't exist at the token level. Everything outbound lands in your Drafts folder for human review.

I built this to power an autonomous operations pipeline: scheduled headless agent sessions sweep the mailbox twice a day, reconcile every open commitment and follow-up into a dashboard, and queue reply drafts for review. It runs identically under Claude Code and OpenAI Codex — one integration layer, two vendors' agents.

Permission model

Auth is MSAL Authorization Code Flow with PKCE through a localhost-only callback. Tokens cache to ~/.outlook-mcp/ with 0600 permissions.

Tools (10)

Setup

1. Register a (free) app in Entra ID

Microsoft Entra admin center → App registrations → New registration:

• Supported account types: your tenant only, or multi-tenant + personal accounts (then use OUTLOOK_MCP_TENANT_ID=common)

• Redirect URI: platform Mobile and desktop applications → add http://localhost:3847/auth/callback

• Authentication → enable Allow public client flows

• API permissions → Microsoft Graph → Delegated → Mail.ReadWrite, MailboxSettings.Read

No client secret is needed (public client + PKCE).

2. Build and authenticate

npm install
npm run build

export OUTLOOK_MCP_CLIENT_ID=<your app client id>
export OUTLOOK_MCP_TENANT_ID=<your tenant id or "common">

npm run auth   # opens browser once; token cache persists at ~/.outlook-mcp/

3. Connect an agent

Claude Code — .mcp.json in your project (or claude mcp add):

{
  "mcpServers": {
    "outlook": {
      "command": "node",
      "args": ["/path/to/outlook-mcp/dist/index.js"],
      "env": {
        "OUTLOOK_MCP_CLIENT_ID": "<client id>",
        "OUTLOOK_MCP_TENANT_ID": "<tenant id>"
      }
    }
  }
}

OpenAI Codex — ~/.codex/config.toml:

[mcp_servers.outlook]
command = "node"
args = ["/path/to/outlook-mcp/dist/index.js"]
env = { OUTLOOK_MCP_CLIENT_ID = "<client id>", OUTLOOK_MCP_TENANT_ID = "<tenant id>" }

Same server, same tools, either agent. That portability is the point of MCP.

Architecture

┌─────────────┐     stdio      ┌──────────────┐     HTTPS     ┌─────────────────┐
│ Claude Code │◄──────────────►│              │◄─────────────►│ Microsoft Graph │
├─────────────┤   JSON-RPC     │  outlook-mcp │   REST v1.0   │  /me/messages   │
│ OpenAI Codex│◄──────────────►│  (Node/TS)   │               │  /me/mailFolders│
└─────────────┘                └──────┬───────┘               └─────────────────┘
                                      │
                               ┌──────┴───────┐
                               │ MSAL + PKCE  │  Mail.ReadWrite only —
                               │ token cache  │  no Mail.Send on the token
                               └──────────────┘

• Transport: stdio (JSON-RPC), one process per agent session

• Auth: MSAL Auth Code Flow + PKCE, localhost callback, silent refresh from disk cache

• Bodies: HTML mail converted to clean plain text before it reaches the model (token efficiency)

• Validation: every tool input validated with zod before any Graph call

Development

npm run dev       # tsc --watch
npm run inspect   # MCP Inspector against the built server

License

MIT © Don Fournier