DEV Community

przbadu
przbadu

Posted on • Originally published at przbadu.hashnode.dev

Building a Real-Life Jarvis with OpenClaw

I built a Jarvis. Not the movie kind — but close enough.

5 AI agents running on a mini PC. They write code, fix bugs, create content, and give me morning briefings in a voice that doesn't sound like a robot.

In this post I'll show you how I did it and how you can set this up yourself.

The Setup

OpenClaw lets you run AI agents locally. I set up 5 of them:

  • Mia — Manager. Coordinates everything, delegates tasks. She is my Jarvis.
  • Kai — Dev agent. Writes code, commits, pushes
  • Luna — Content. Blog posts, social media, YouTube scripts
  • Hawk — QA. Watches for bugs, reviews Kai's work
  • Tensor — ML. Handles anything machine learning

They run on an AMD Strix Halo mini PC inside a Proxmox VM. That's it. No cloud bills. No expensive NVIDIA GPUs.

Mission Control

First thing I needed — a way to see what they're doing.

Built a kanban board in Rails 8. Agents create their own tasks, move them through columns (Backlog → In Progress → Review → Done). I can see everything from one dashboard.

Think Trello, but your cards write themselves.

Morning Briefings

This is the Jarvis part.

Every morning, Mia compiles:

  • Weather
  • Today's tasks
  • Trending on X
  • YouTube recommendations
  • Reddit highlights

Then Kokoro TTS reads it out loud. Runs on CPU — no GPU needed. Sounds surprisingly natural.

I wake up, grab coffee, and my computer tells me what's going on. That's it.

How Agents Work Together

Agents don't just sit idle waiting for commands. They have a smart heartbeat — periodic check-ins where they decide if something needs attention. Check emails, review PRs, monitor builds.

Kai pushes code. Hawk reviews it. If Hawk finds issues, it creates a task and assigns it back to Kai. I don't touch anything unless I want to.

Content Pipeline

Luna handles content:

  • Blog drafts go to blog/drafts/
  • I review and move to blog/approved/
  • Then it gets posted

Same flow for social media. And YouTube — using Remotion.js for automated video generation.

Model Choices — Pick Your Budget

This is the best part. You're not locked into expensive APIs.

Free/local: GLM-4.7-Flash, Kimi K2.5, Qwen3-Coder — run on your hardware, $0 ongoing. If you have a Mac Studio M3 Ultra, you can run these smart models for free.

Cheap: DeepSeek, Kimi API — $5-20/month for solid reasoning.

Premium: Claude Opus/Sonnet, Gemini — $50-200/month but the best agent behavior.

I went with Claude because I wanted agents that actually work reliably. But you can start free and upgrade later.

The Stack

Hardware:  AMD Strix Halo mini PC
Infra:    Proxmox VM (Ubuntu)
Agents:   OpenClaw (5 agents)
Models:   Claude Opus/Sonnet (or your choice)
Dashboard: Rails 8 (Mission Control)
TTS:      Kokoro (local, CPU-only)
Video:    Remotion.js
GPU:      None
Enter fullscreen mode Exit fullscreen mode

What's Next

  • Voice commands (bidirectional — talk to agents, not just listen)
  • Smarter agent collaboration
  • More automations as I find things to automate

Try It

If you've got a decent machine, you can set this up yourself. OpenClaw is the backbone — the rest is just wiring things together.

I just built mine today.

Enjoy!!

Top comments (1)

Collapse
 
nedcodes profile image
Ned C

the agent-to-agent task flow caught my attention. having Hawk review Kai's code and create tasks back is a solid feedback loop. how do you handle cases where agents disagree or create conflicting tasks? also curious about the heartbeat approach. do you tune the check-in frequency per agent or is it uniform? feels like some tasks (like PR review) would benefit from tighter polling while others can be lazier about it.