Belvedere: An AI Assistant for Email and Admin
Running an AI assistant to handle email, drafts, and administrative tasks via OpenClaw and Telegram.
I've recently brought on a new assistant to handle email and administrative work. His name is Belvedere — he's an AI agent running on OpenClaw, a platform that lets me orchestrate autonomous tasks without building custom infrastructure from scratch.
What is OpenClaw?
OpenClaw is a runtime for AI agents. Think of it as a more structured alternative to prompting Claude in a chat window. It lets me define agents (like Belvedere or my trading bot, Gordon) that run on a schedule, connect to external services, and execute defined workflows. Each agent has its own session, context, and permissions. In practice, it means I can have multiple specialized assistants running concurrently, each handling a specific domain.
What does Belvedere do?
Belvedere checks my email inbox every 15 minutes between 7 AM and 10 PM UTC. When new messages arrive, he reads them, drafts replies in his voice (deferential, understated, efficient), and sends them to me via Telegram for approval before actually sending. This gives me control without requiring me to manage the inbox directly. Each morning, Belvedere also sends me a summary of my running to-do list via Telegram, so I start the day with a clear sense of what's pending.
How it works: The Telegram loop
OpenClaw integrates with Telegram, so Belvedere and I communicate through direct messages. I can ask him to check something, approve a draft, or adjust a policy. He sends me draft emails, I reply with a thumbs up or a correction, and he handles the send and filing. The workflow is lightweight and asynchronous — I'm not required to respond immediately.
What actually happens
The mechanics are straightforward. Belvedere connects to my iCloud mailbox via IMAP, retrieves recent messages, identifies which ones he's seen before (so he doesn't repeat himself), and for new emails, he crafts a response. He signs off as "Belvedere, Andy's assistant" so senders know they're hearing from an AI, not from me directly. Once approved, the email is sent via SMTP and automatically filed in my Sent folder.
Why bother?
Honestly, I'm a techie and I can, so I did. Email is straightforward to automate — it's just IMAP and SMTP — and OpenClaw made it trivial to wire up. The inbox still requires judgment calls from me on occasion, but the routine stuff (acknowledgments, confirmations, scheduling) doesn't need my attention. Belvedere handles it cleanly and I approve the sends in batches via Telegram. It's the sort of thing that would have taken a weekend to bolt together five years ago; now it's an afternoon project.
What's next?
The system is still early. Right now it handles standard transactional email. Over time, I might extend it to handle calendar conflicts, extract action items from longer threads, or integrate with other services. OpenClaw is purpose-built to let me add these features without rewriting infrastructure each time.
If you're using OpenClaw or building something similar, I'm happy to discuss the setup.