# My 5090 rig sits at home all day, so I built a tunnel that lets my AI agents use it

> Source: <https://dev.to/shootingsyh/my-5090-rig-sits-at-home-all-day-so-i-built-a-tunnel-that-lets-my-ai-agents-use-it-3hfj>
> Published: 2026-10-28 04:07:31+00:00

Ok, this is about how I work with my desktop. I bought a 5090 + AMD 9950X3D last year and was super excited about it. You can imagine I would do a lot of projects with it. The only problem is I cannot stay with it the whole day. But it didn't feel like something I couldn't mitigate.

I was using opencode at first, so I explored using SSH to remote into a tmux session together with Tailscale. It worked for a while, then failed when I went on a cruise and the ship's internet didn't support VPN.

Next I tried tunnels. I tried VS Code tunnel, which doesn't really work for me since I use my mobile phone and the UI is awful on mobile. Then I used Cloudflare Tunnel, which finally worked stably. However, the experience with opencode was still not so good — the UI is still a bit dumb, and opencode was not really good at managing the computer, especially when I need to install things or configure the desktop for new models.

So I installed OpenClaw back and tried to use its WhatsApp/Discord channels. It worked really well for a while — especially when the Muse API was launched and the contributor tier was super cheap, I used Muse to do the normal management and a local model to do coding. Initially I used WhatsApp to talk to my opencode to do projects, but I dropped opencode soon for my in-house coding agent. The problem now is that the maintenance of OpenClaw is not trivial. The Windows companion is taking a lot of CPU, which kept my fan humming, and the browser control never worked with it.

So I turned to public agents. I have a ChatGPT subscription, the $20 tier. I used to use it to do some coding directly on GitHub, but it can't run long jobs. And I also have a Muse subscription, which was good for many browser tasks and long jobs, though always blocked by websites.

So it came to an idea: why cannot I let these public agents use my own computer?

Shortly after I had the idea, ChatGPT actually started to support it, using an MCP called DesktopCommander. It works, though I had concerns about exposing my computer to a public service that is not open sourced. Also it always got blocked by itself due to unknown reasons. And Muse kept not supporting anything like that.

So I decided to build my own MCP tunnel, which resulted in this project: HomeMCP. It's open source, relay included — your machine gets a name like yourname.tunnel.homemcp.si, TLS ends on your own machine, and anything sensitive needs a passkey approval. Hosted relay is $0.99/mo, or bring your own tunnel: [https://github.com/shootingsyh/homemcp](https://github.com/shootingsyh/homemcp)