iris
Enables AI agents to create, list, update, and delete Microsoft 365 email drafts, with a hard guarantee that no message can be sent automatically, ensuring human review before any email goes out.
README
iris
A Microsoft 365 mail server for AI agents that cannot send email.
Not "will not". Cannot. iris requests the delegated Graph scope Mail.ReadWrite
and never Mail.Send. The access token it holds has no capability to transmit a
message, so no prompt, no jailbreak and no bug in this code can make one go out.
It writes drafts into a folder in your mailbox. You open Outlook and press Send.
That is the whole design. Everything else is detail.
Why this shape
The usual worry about giving an agent your mailbox is that it will send something you did not sanction — to the wrong person, with the wrong tone, or because someone talked it into doing so. The common answer is a confirmation prompt, which is a guardrail: code that asks permission, and code can be bypassed.
iris removes the capability instead. Microsoft Graph will reject a send attempt made with this token, because the consent screen you approved never included that permission. The security boundary is Microsoft's, not this program's, and it holds even if this program is wrong.
The trade is real: a human is in the loop on every message, by construction. If you want autonomous sending, iris is the wrong tool.
Install
uvx iris-mcp # run without installing
pip install iris-mcp
Python 3.10+.
Setup
You must register your own Entra application. There is no shared app registration and no hosted service — iris talks directly from your machine to your tenant. This is deliberate: a shared app would mean trusting someone else's client ID with access to your mail.
- Entra admin centre → App registrations → New registration. Single tenant is fine. No redirect URI needed.
- Authentication → Settings → enable Allow public client flows. Device code sign-in needs this. No client secret is used anywhere.
- API permissions → Microsoft Graph → Delegated → add
Mail.ReadWrite. Add nothing else. Do not addMail.Send— if it is present, the guarantee above is void. - Copy the Application (client) ID and Directory (tenant) ID. Neither is a secret.
Then add iris to your MCP client:
{
"mcpServers": {
"iris": {
"command": "uvx",
"args": ["iris-mcp"],
"env": {
"IRIS_CLIENT_ID": "<application client id>",
"IRIS_TENANT_ID": "<directory tenant id>"
}
}
}
}
Sign in once: call iris_login, open the URL, enter the code, then call
iris_login_finish. The token cache is written next to the server, mode 600.
Tools
| Tool | What it does |
|---|---|
iris_login |
Starts device-code sign-in, returns a URL and a code |
iris_login_finish |
Completes sign-in; safe to call repeatedly while you type the code |
iris_auth_status |
Who is signed in, which scopes, and whether Graph is reachable |
iris_create_draft |
Writes a draft (to/cc/bcc, subject, body or HTML, optional reply-to) |
iris_list_drafts |
Lists what is waiting in the draft folder |
iris_update_draft |
Revises a draft in place |
iris_delete_draft |
Deletes a draft; requires confirm=true |
Where drafts go
Into a dedicated top-level mail folder, AI Drafts by default
(IRIS_DRAFT_FOLDER). It is created on first use. Set the variable to an empty
string to use the normal Drafts folder instead.
These are real drafts and Outlook sends them normally — but because they live in their own folder, they do not appear in the Drafts view. That is the point: agent-written mail sits somewhere you have to go and look, rather than mixed in with your own half-finished messages.
One wrinkle worth knowing: Graph's createReply always lands a reply in Drafts
first, so iris moves it afterwards, and a move assigns a new message id.
Other controls
- Recipient allowlist —
recipients.allow, one address or domain per line. Absent or empty means all recipients are permitted. PointIRIS_ALLOWLISTelsewhere if you prefer. - Kill switch — create a
DISABLEDfile beside the server, or setIRIS_DISABLED=1, and every tool refuses. - Audit log — every call is appended to
audit.log(IRIS_AUDIT_LOG).
Limits
No attachments. No shared or delegated mailboxes — /me only. No folder nesting
via parentFolderId. Sign-in is delegated device-code as a public client, so
the blast radius is exactly one mailbox: yours.
Security
The no-send guarantee, how to verify it yourself, and — just as important — what
iris can reach with Mail.ReadWrite: see SECURITY.md.
License
MIT — see LICENSE.
<!-- mcp-name: io.github.SuperAngryMonkey/iris -->
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。