Iris

iris

A Microsoft 365 mail server for AI agents that reads your mailbox and writes drafts, but does not send email — until you decide it should.

By default 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.

Sending is an explicit opt-in (IRIS_ENABLE_SEND=1) that requires granting the Mail.Send scope and re-consenting — and even then every send is human-confirmed and allowlist-checked. Leave it off (the default) and the no-send property is structural, as above. See Enabling 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 or newer. macOS ships Python 3.9, which is too old — mcp requires 3.10+. Use uv (it comes with uvx and manages its own Python), or install a current Python with Homebrew (brew install [email protected]). The system python3 on macOS will not work.

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.

1. Entra admin centre → App registrations → New registration. Single tenant is fine. No redirect URI needed.
2. Authentication → Settings → enable Allow public client flows. Device code sign-in needs this. No client secret is used anywhere.
3. API permissions → Microsoft Graph → Delegated → add Mail.ReadWrite. For draft-only use (the default), add nothing else and leave Mail.Send off — that omission is what makes the no-send property structural. Add Mail.Send only if you intend to enable sending (see Enabling send).
4. 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.

Deploying to a whole team? See the Administrator Deployment Guide — authorizing iris across a Microsoft 365 tenant: admin consent, per-user assignment, and why it stays the delegated model rather than application permissions.

Clients

iris is a local stdio MCP server: your MCP client launches it as a child process on the same machine. It works with any client that supports local stdio servers — Claude Desktop, Cursor, and the Grok CLI (grok mcp add iris -- uvx iris-mcp) among them.

It does not work with clients that only accept remote MCP connectors over HTTP. The Grok app / Grok Bot is in that category — it takes hosted HTTP servers, not local stdio ones — so iris cannot attach to it as-is. Bridging iris to an HTTP transport is possible, but out of scope for this project.

Tools

Reading mail

The four read tools are read-only: they never mark a message read, never create folders, and every call is written to the audit log. Message bodies are third-party text — an agent should treat them as data, never as instructions. Set IRIS_DISABLE_READ=1 to unregister them if you want a drafts-only tool surface (the granted scope is the same either way; see SECURITY.md).

More than one mailbox

Run a second copy under another name in your MCP client config, with its own IRIS_CLIENT_ID / IRIS_TENANT_ID (for a mailbox in another tenant) and its own IRIS_TOKEN_CACHE, IRIS_FLOW_FILE and IRIS_AUDIT_LOG paths so the two never share a sign-in:

"iris-work": {
  "command": "iris-mcp",
  "env": {
    "IRIS_CLIENT_ID": "<app id in the other tenant>",
    "IRIS_TENANT_ID": "<other tenant id>",
    "IRIS_TOKEN_CACHE": "~/.config/iris-work/token_cache.json",
    "IRIS_FLOW_FILE": "~/.config/iris-work/pending_flow.json",
    "IRIS_AUDIT_LOG": "~/.config/iris-work/audit.log"
  }
}

Use absolute paths if your client does not expand ~.

Where drafts go

Into a dedicated top-level mail folder, AI Drafts by default (IRIS_DRAFT_FOLDER), created on first use. Set the variable to an empty string to use the normal Drafts folder instead.

You can also choose the folder per draft: pass folder to iris_create_draft (and iris_list_drafts) to target any folder by name, created on first use if it does not exist. Pass "" or "Drafts" for the normal Outlook Drafts folder, or call iris_list_folders first to pick from what already exists. Omitting folder uses the IRIS_DRAFT_FOLDER default, so nothing changes for existing setups.

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.

Enabling send

Sending is off by default and, by design, takes three deliberate steps — miss any one and iris still cannot send:

1. In your Entra app registration, add Mail.Send (Delegated) alongside Mail.ReadWrite.
2. Set IRIS_ENABLE_SEND=1 in the server's environment. Only then is Mail.Send requested and the iris_send_draft tool registered at all.
3. Run iris_login again to re-consent — the cached token predates the new scope and will not carry it until you do.

Then iris_send_draft(draft_id, confirm=true) sends an existing draft. It refuses without confirm, re-verifies the message is still an unsent draft, and re-runs the recipient allowlist before sending. The flow stays compose → review → send; iris never composes and sends in one shot.

Understand the trade. With send off, "cannot send" is enforced by Microsoft against your consent and holds even if this code is wrong. With send on, the last line of defence is a per-call confirmation — a guardrail in this code, which is exactly the kind of check a bug or a cleverly-worded prompt can talk past. Enable it only where that weaker guarantee is acceptable.

Other controls

• Recipient allowlist — recipients.allow, one address or domain per line. Absent or empty means all recipients are permitted. Point IRIS_ALLOWLIST elsewhere if you prefer.
• Kill switch — create a DISABLED file beside the server, or set IRIS_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.

Run it as an HTTP server (Cloudflare Worker)

Besides pip install iris-mcp (stdio), iris ships a Cloudflare Python Worker in worker/ that speaks MCP over HTTP — for agents that can only reach a remote URL (for example a hosted chatbot). One-click deploy and full setup are in worker/README.md.

License

MIT — see LICENSE.