M365 Inbox Agent for Azure Functions (Python)

An AI agent that triages your Microsoft 365 inbox: it escalates urgent mail to Teams, drafts replies, and emails you a daily briefing - all following rules you write in plain markdown.

Powered by Azure Functions and Microsoft 365 connectors. You bring the markdown logic and a little Python for your rules.

Run it locally in minutes against sample data, point it at your real inbox, then deploy it so it runs on its own.

What it does for you

• 📋 Daily Briefing: a daily summary of what matters lands in your inbox.
• 🚨 Escalate: VIP / urgent mail is posted to your Teams channel, with an @mention.
• ✉️ Reply: action-required mail gets a grounded draft reply.
• 💬 Chat: ask read-only questions about your recent mail.

→ Full walkthroughs in docs/use-cases.md.

Prerequisites

• Python 3.13+. Easiest install: uv, then uv python install 3.13.
• Azure Functions Core Tools v4 ≥ 4.12.0 (the v5 preview is not yet compatible).
• Azure Developer CLI (azd).
• An Azure subscription. azd provision (Quickstart step 2) creates the Microsoft Foundry model deployment the agents need - required even for the offline path.
• For real M365 (or azd up): permission to authorize Microsoft 365 connectors, plus the connector-namespace CLI extension:

macOS / Linux:

curl -fsSL https://aka.ms/connector-namespace-cli-install | sh

Windows:

powershell -Command "Invoke-WebRequest -Uri 'https://aka.ms/connector-namespace-cli-install' -OutFile 'install.sh'; wsl bash install.sh"

Quickstart

Five steps: install, get resources, run locally, try it, deploy.

1. Install the tools

macOS:

brew tap azure/functions
brew install azure-functions-core-tools@4
brew install azure-dev
npm install -g azurite
curl -LsSf https://astral.sh/uv/install.sh | sh

Linux / Windows / WSL: Use the Core Tools v4 install guide, azd install guide, Azurite install guide, and uv install guide.

2. Get the resources you need

azd provision creates the Foundry model deployment the agents require (needed even offline); hydrate copies the settings into local.settings.json. No API keys - managed identity throughout.

macOS / Linux:

azd provision
./infra/scripts/hydrate-local-settings.sh

Windows:

azd provision
.\infra\scripts\hydrate-local-settings.ps1

Yes - even the offline, sample-data run needs this one-time azd provision, because the agents call a Foundry model. It provisions the model deployment only; it never touches your inbox.

3. Run the local client

azurite --silent --skipApiVersionCheck --location .azurite   # terminal A
uv run func start                      # terminal B
uv run python chat.py                  # terminal C

(These commands work identically on macOS, Linux, and Windows.)

🪟 Windows local dev: If uv run func start fails with ModuleNotFoundError: No module named 'azure_functions_agents', the Microsoft Store python.exe alias on your PATH is shadowing the venv Python. See Troubleshooting: Windows local dev for the fix - you'll need to remove the Store Python and disable App execution aliases in Windows Settings.

4. Try it (offline, safe)

Pick 1, 2, or 3. The client shows a 🟡 Offline banner and runs every agent in DRY RUN against sample-data/inbox/*.json: it produces the full deliverable as text and calls no connector, so nothing is ever sent. Pick 5 to chat with the sample inbox.

Want it to act on your real inbox while still local? See Go live with real M365.

5. Go live (set your recipient and Teams target)

Set who real mail and Teams posts go to, before you deploy. Without these, the deployed agents stay in DRY RUN (the chat client's doctor banner will tell you).

azd env set MAILBOX_OWNER_EMAIL you@your-tenant.com     # required for LIVE mail
azd env set TEAMS_TEAM_ID        <team-id>              # optional, enables Teams alerts
azd env set TEAMS_CHANNEL_ID     <channel-id>           # optional, enables Teams alerts

Get the Teams ids by opening the target channel in Teams → ⋯ → Get link to channel (the URL contains both ids), or via docs/configuration.md.

Already deployed with placeholders? Set them now and re-run azd up (or just azd provision) to push the new values to the Function App.

6. Deploy to Azure, then try it again

azd up

Known issue (point-in-time): azd up may currently fail at the deploy step on Python 3.13 - the Flex remote build uses Python 3.11.8 (Azure/azure-dev#8538). Simple workaround: docs/deploy-python-313.md. This note can be removed once the bug is fixed.

Now inbox-triage fires automatically on every new email - no client, no waiting. Send yourself a message, then watch your Teams channel (VIP / incident) or your inbox (replies). Tail the live trace with azd monitor --logs.

7. Make it yours

• ✍️ Edit skills/vip-rules.md to set your VIPs, what to skip, and what escalates to Teams.
• 🔁 The weekly-rule-suggestions agent proposes tuning that you approve by hand.
• 🔒 Use this repo as a private template before adding real rules or tenant data.

→ Full guide: docs/customize.md.

8. Clean up

azd down --purge

Hitting an error? See docs/troubleshooting.md.

How it works (the short version)

This is an Azure Functions app on the serverless agents runtime. Each agent is a markdown file (*.agent.md) that reasons over your rules in skills/*.md; a small tools/match_rule.py adds deterministic classification. Microsoft 365 actions go through Entra-authorized MCP connectors - no app secrets, managed identity end to end.

→ Deeper dives: How it works · Configuration & deployment · Customize · Troubleshooting & reference

Learn more

• Serverless agents runtime in Azure Functions
• Office 365 Outlook connector · Microsoft Teams connector
• Python vs Markdown variants