I spent last weekend doing something I'd been putting off for months: actually setting up an AI assistant that works for me, not just one that impresses people at dinner parties.

I have a busy life — a wife, a dog, a couple of kids, a home project or two in flight, and an inbox that never stops. Sound familiar?

I'd played with ChatGPT and Claude plenty. But using them like a search engine is different from having something that actually knows your life and works in the background while you're busy doing other things.

Here's what I built, what surprised me, and what I'd do differently.

What I Actually Built

I set up OpenClaw — an open-source AI agent framework that runs on your own machine. The agent (I named him Servo — a sarcastic horse with an old soul, if you're wondering) connects to my Gmail, Outlook, Google Calendar, and Obsidian vault. He monitors my email in real-time, triages it, and only surfaces the stuff that actually matters. He sends me a morning briefing every day at 7:30 AM. He adds tasks to my Obsidian project notes. He manages my calendar.

All over Signal, which means it works exactly like texting a really competent assistant.

What Works Better Than Expected

Email triage is genuinely transformative. I told Servo to drop marketing emails, newsletters, streaming promos, and charity solicitations I've already donated to. He surfaces shipping updates, financial alerts, and anything actionable. I went from dreading my inbox to actually staying on top of it.

The signal-to-noise improvement is hard to overstate. Today I got flagged about a blocked Google sign-in attempt, an AWS infrastructure deadline, and an oil price spike tied to strikes on Iran. I did not get 40 marketing emails I'll never read.

Tasks go directly to the right place. Instead of dumping everything into one inbox, I have Servo route tasks to specific Obsidian projects — family stuff goes to Bork Family, work projects have their own notes, home projects have their own project. It sounds simple but it's the difference between a system you actually use and one you abandon.

The morning briefing changes your day. Having a concise summary of weather, calendar blocks, and open tasks waiting for me when I wake up means I start the day oriented instead of scrambling. It's a small thing with outsized impact.

What Was Harder Than Expected

Gmail setup has more steps than you'd expect. OAuth scopes, token caching, getting the CLI working correctly — it's not plug-and-play. Budget an hour and follow the docs carefully. It's worth it, but don't expect five minutes.

Outlook was even more involved. Microsoft's Graph API requires registering an Azure app, configuring the right permissions, and authenticating via device code flow. I also had to write a custom polling daemon to bridge Outlook into the OpenClaw webhook system since there's no native hook. Totally doable — but definitely an afternoon project, not a quick setup.

Signal took some patience too. Setting up Signal as the delivery channel meant running signal-cli, linking a device, and getting the message routing configured correctly. Once it works it's seamless — but the initial setup has some rough edges. Totally worth it for the "just text your assistant" experience.

Cron jobs are finicky. I set up an automated morning briefing and it failed silently for two days before I figured out the Signal delivery target wasn't configured correctly. Worth spending time upfront getting the plumbing right.

Obsidian integration requires some setup. The agent writes directly to Markdown files, which is powerful — but you need to be consistent about your vault structure for it to work well. I spent time this weekend setting up proper project files, and it was worth it.

Knowing what to delegate. There's a learning curve to figuring out what's worth asking your AI assistant versus just doing yourself. Anything with context — "add this task to the right project based on what we've talked about" — works great. Simple lookups are faster to just Google.

What Surprised Me Most

The personality matters more than I expected.

I gave Servo a persona — a horse with an old soul, warm but sarcastic, no corporate filler phrases. And it actually changes how I interact with the system. I look forward to the updates. I have a sense of who I'm talking to. It sounds silly, but "assistant with no personality is just a search engine with extra steps" turned out to be true.

What's Next

I'm planning to write more about this as I go. Next up: how I'm using it for Obsidian knowledge management, and what it looks like after a month of real use.

If you're curious about OpenClaw or want to compare notes on AI assistant setups, reach out — I'm always happy to geek out about this stuff.