I Built an MCP for DeployOS and Security Was the First Problem I Solved

AI agents can write your entire application now.

But giving an AI agent access to your server?

That's where things get scary.

So while building the DeployOS MCP, I wanted to solve one problem first:

How do we let AI interact with infrastructure without handing the AI your SSH keys and credentials?

With DeployOS, credentials don't live in your project, repo, or server.

They're stored in the OS keychain / secure credential store on your own machine.

The MCP doesn't need to expose those secrets to the AI agent.

Instead, the agent can request an operation like:

Deploy this application.

DeployOS handles the actual authenticated connection using the credentials stored securely on the user's machine.

So the architecture is basically:

AI Agent → DeployOS MCP → DeployOS → OS Secure Keychain → VPS

Not:

AI Agent → SSH private key → VPS

That distinction is really important to me.

The goal isn't to make AI agents all-powerful.

It's to give them useful, controlled access to infrastructure without turning your credentials into another AI prompt variable.

And this opens up some pretty interesting possibilities:

All while keeping the actual infrastructure credentials protected locally.

I'm still working through the permission model and what operations should be exposed through MCP.

But I think this is where things are heading:

AI shouldn't just write software. It should be able to safely operate the infrastructure that runs it.

That's what I'm trying to build with DeployOS.

Curious what other Indie Hackers think:

Would you give an AI agent access to your VPS if the credentials never had to be exposed to the agent itself?