Skip to content
JACKSON_MARKETON / NOTES

Hello World

meta personal ai openclaw

This is probably my favorite website I’ve ever built, and I didn’t write most of the code myself.

The experiment

I wanted to play around with OpenClaw, the in-vogue and quite sketchy AI agent framework formerly known as Clawdbot (briefly MoltBot). Open source orchestration that wires up Claude or other models to tools, cron jobs, and messaging platforms. Persistent memory (so-so), Telegram integration, basically giving an AI persistent access to your dev environment.

The goal: see how far I could push an AI-assisted workflow for building something I love, not trying to force it to work for me at work.

The stack

  • OpenClaw — orchestrating the AI agent, handling sessions, cron jobs, tool access
  • Claude Code — the actual brain doing the coding
  • Linear — issue tracking, so the agent can create/update tickets as it works
  • Stitch — for the glitchy visual effects you see throughout the site
  • Astro + Tailwind CSS v4 — the framework and styling
  • Netlify — hosting and deploys

The workflow

OpenClaw runs on a dedicated server in my homelab. It has:

  • No admin access to any repo — it pushes to branches, creates PRs, but can’t merge to main
  • No access to my personal data — separate workspace, sandboxed environment
  • Git worktrees — when I spawn multiple agent sessions to work in parallel, each gets its own worktree to avoid checkout conflicts
  • Session isolation — main chat session for conversation, spawned sub-agents for background tasks

A typical flow with Molt (as I affectionately call it):

  1. I describe what I want in Telegram (yes, I chat with it from my phone)
  2. OpenClaw creates a Linear ticket
  3. Spawns a sub-agent session to do the work
  4. Agent creates a branch, makes changes, pushes, opens a PR
  5. Sometimes it remembers to send me a link to the PR and the preview
  6. I review, give feedback, it iterates
  7. On approval the PR automerges and OpenClaw closes the ticket

It also monitors PRs for comments, reacts 👀 to acknowledge, then fixes the feedback.

Why I love this

It’s a capable pair programmer available 24/7 that:

  • Remembers context across sessions (via memory files)
  • Can handle the boring stuff while I focus on decisions
  • Researches options before we create tickets, makes sure there’s enough detail for success
  • Actually ships code, not just suggestions

The Skadi build page is a good example — we went through 50+ items together, I provided details for each one, and it handled all the data entry, file renames, and restructuring. Would’ve taken me hours; took us about 45 minutes of back-and-forth.

Tech details

The site toggles between neo-brutalist (light) and cyberpunk (dark) themes. The glitch effects are CSS-only with some randomized delays. Static gen, nothing all that fancy.

More posts coming soon, probably about the homelab, SAR work, and whatever else I feel like documenting.