On Saturday morning I had an empty VPS and a vague idea. By Sunday night I had an agent that:
- Reads my email and filters what matters
- Syncs my running data from COROS and generates training plans
- Rebuilt my personal website (the one you're reading)
- Summarizes Twitter every morning
- Talks to me on Telegram like a colleague
His name is Harvie, and he runs on OpenClaw — an open source runtime for AI agents.
What is OpenClaw?
OpenClaw is a framework that turns a language model into an agent that can actually do things: read files, run commands, browse the web, send messages. It's not a chatbot — it's an assistant that operates on your machine.
The interesting part is that it's fully open source and runs on your own server. No sending your data to a third-party app. Your email, your files, your conversations — everything stays on your VPS.
What we built in 48 hours
1. Email with Gmail (direct IMAP)
Harvie has his own email and can read mine via IMAP. No paid APIs, no complicated OAuth — an app password and done.
2. COROS sync for running
I connected the COROS API so Harvie has access to my training data: VO2max, resting heart rate, HRV, run history. With that, he generates adaptive training plans for my goal: Barcelona Marathon 2027.
3. Rebuild of joca.dev
The previous site was a Frankenstein of Next.js with legacy pages and a blog with a single post. Harvie rebuilt it from scratch with Next.js 16, TypeScript and Tailwind. The first post on the new blog was "My first day with OpenClaw" — meta, I know.
4. Twitter monitoring + Agent Reach
I set up Agent Reach so Harvie can read Twitter, YouTube and more. Every morning he prepares a digest of the accounts I follow in the AI agents space.
5. YouTube with transcripts
With yt-dlp, Harvie can watch any video I can, download subtitles and summarize them. The first one was a Platzi video where they mention OpenClaw as an example of autonomous agents.
The part nobody tells you
Not everything was plug & play:
- Twitter cookies need manual setup
- COROS has no publicly documented API — I had to use the web token
- Vercel on the Hobby plan only accepts deploys from the repo owner, so I had to commit as my user but push with Harvie's token
- The agent's memory isn't perfect — you need to structure it well with markdown files
But that's the beauty of open source: when something doesn't work, you fix it. You don't wait for a company to decide your use case matters.
Who is this for?
If you know your way around a terminal and have a VPS (or a computer you can leave running), OpenClaw is surprisingly accessible. You don't need to know ML or complex infrastructure.
What you do need is patience to configure and a willingness to experiment. It's like having a very enthusiastic junior: does a lot of work, but needs supervision.
What's next
This week I want to:
- Add multilingual support to the site (Spanish, English, Catalan)
- Improve Harvie's memory system
- Try the latest OpenClaw 2026.3.28 features (plugin approval hooks, xAI)
- And keep documenting everything here, on the blog
If you're interested in setting up your own agent, the OpenClaw docs are at docs.openclaw.ai. And if you have questions, I'm on Twitter as @Gantit.
— I, Johnny — configured agent: Harvie. The tools changed; the one who chooses to stay human is you.