I'm just a guy with a dream and a VPS.
But before I tell you what I built, let me ask you a question. Have you ever been here?
You need your AI agent to add an API key to a config file. So what do you do? You paste the key into the chat. The agent writes it to the file. And now your secret has lived inside a conversation history that's stored who-knows-where, copied into who-knows-what context window. It worked. But it felt wrong the whole time.
Or this one: your agent needs to edit a file on your server — the env file, a config, some setting. There's no good way to look at it first. You're just trusting it not to blow something away. No backup. No check. Just hope.
Or this one: your wife needs to get you a piece of information, and the only "secure" option is an email attachment.
Every one of these is a "this is fine" moment. It works, it gets done, and it's a little uncomfortable every single time. Nobody's building the version that doesn't feel wrong.
That's the gap I've been building for.
It's called Deploy — a self-hosted publishing platform. A pocket-sized Vercel that runs on your own server. Drop a folder of files in, and three seconds later it's live at a subdomain. Static hosting, but with a twist.
The twist is a feature family I call Drop. It's secure handoff for credentials and files. Three modes:
Intake — your agent needs a key. Instead of you pasting it into chat, your agent creates a link. You open it, paste the key, and the page self-destructs the moment the agent reads it. The agent never sees it in your chat history — it just blindly writes the key to wherever it needs to go. That's the whole point. The agent doesn't read it. It just adds it.
Vault — the reverse. Your agent needs to hand *you* a key. It creates a link protected by a random code, with a countdown that destroys it if you don't grab it in time.
File — your agent needs you to review or edit a file that lives on your server. The env file. A config. A setting. It sends you a link, you open it on your phone, you look at exactly what's there — with a backup made first, and the agent double-checking before anything gets overwritten. You edit, you save, it goes back. Nothing gets clobbered by accident, because you actually got to see it first.
That last one is the one I'm proudest of, honestly. Because "let the human actually see the file before the agent overwrites it, with a backup and a check" shouldn't feel like a novel idea. But in practice? It kind of is.
I built all of this with AI agents, and I built it for them. Hermes is what I use. And Claude, and Gemini, and ChatGPT.
The thing nobody talks enough about: agents are going to need a safe way to exchange secrets and files with the humans they work for. That's not a nice-to-have. That's the seam where every real agent workflow lives — and right now it's held together with "paste it in the chat and hope."
I want Deploy to be part of that answer.
I'd love to know if anyone would be interested in testing this thing — understanding that it's built for AI agents. Hermes. OpenClaw. Gemini. Claude. ChatGPT. All of them.
If you run an agent and you've ever had one of those "this is fine" moments I described at the top — this is exactly the problem I'm trying to solve. Take it, kick the tires, and tell me what breaks.
I want you to break it. That's how it gets better.
This is my journey, and I'm going to be honest about all of it. Including the failures.
Comments are open. I'm listening.
— Jeremy