All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
MATRIX - AI agent in our Manager that admins VPS over SSH
Hi all! We recently built MATRIX, an AI agent directly into our control panel that can connect to a VPS over SSH, diagnose problems, and install or configure software.
https://ginernet.com/en/agentic-vps/
The part we spent the most time on wasn’t the SSH integration or the LLM, it was figuring out how to let an agent operate on a real server without giving it unchecked control.
Before anything runs, MATRIX has to propose an intervention plan: commands/actions, risks, and backups. Execution is then blocked until the user clicks Authorize plan in the control panel.
That authorization happens outside the conversation. The button creates a record in our database that the execution layer checks before allowing the agent to proceed, so the model can’t simply claim that authorization was granted or bypass it through chat.
During execution there’s a live SSH transcript, with token usage and cost updating in real time. We also added an emergency stop, hourly spending limits, and intentionally did not expose a server deletion tool to the agent. Destroying a VPS can only be done manually by the human from the control panel.
Users can choose between AI providers in the US, EU, or Asia (useful when data residency matters) or connect their own agent through our MCP server instead.
The landing page includes an interactive mock of the whole intervention/authorization loop, so you can see how it works without creating an account.
Using it against a real VPS requires an account. The platform is prepaid and billed hourly; the smallest VPS works out to roughly $5/month.
I’d be especially interested in feedback on the authorization/safety model, and happy to answer technical questions about how we built it.
Comments
Sounds nice, looks good, but I prefer the lowendWAY.
Someone coded (I forked it) a 10 line python agent with shell as tool access.
https://github.com/Ne00n/smol/blob/main/smol.py
You just setup a llama.cpp server, connect the agent and you good to go.
If you have a decent local model, any such provider is cookedTM.
Especially if you have a spare 3090/4090 with 32GB VRAM which can run Qwen 3.8 at decent speeds.
Yeah that doesnt sound the lowendway
You can run a MoE without any GPU at decent speeds.
Neither do I have a 3090 spare, cries.
He did a clean up, which I dislike though, now it looks structured and its actually readable.
https://github.com/smol-env/smol/blob/main/smol.py
edit: he actually Improved it, nevermind, I have to update my fork, yeesh
Have you thought about spoon?
Kindest best.
Noted. Avoid ginernet. Cause: using ai, even in sensitive areas.