Zachary Gray

The agent I text. It runs my other agents across three machines, asks me before anything risky, and never sees a password.

What it does

I gave my AI a phone number, and now I text it like anyone else.

Rocky is the one front door to everything else I run, from my course brain to my three machines to the Claude Code sessions that do the actual work. I don't need a terminal, or even a computer. I text it from wherever I am, and the answer comes back in the same thread.

Probably the most useful thing it does is turn my sloppy texts into good prompts. I send it lazy half-sentences, and it rewrites them into a proper prompt for each session. When I want the wheel, it hands me the session link and I take over.

What I actually ask it for

  • Answers grounded in my real lectures and Harvey
  • Study guides
  • A morning brief across school, projects and life
  • Code in any repo
  • Moving files between machines
  • Running agents in parallel
  • Turning my sloppy texts into good prompts
  • Handing me the session link when I want the wheel
  1. Me:

    make me a study guide for the math test

  2. Rocky:

    Got it handled. Your study guide is ready. Here it is:

    exam-1-study-guide.mdDocument
  3. Me:

    actually make it cover chapter 5 too

  4. Rocky:

    On it. If you want to steer this one yourself, here's the session:

    Study guideClaude Code, running

A text becomes work

Three machines, each with a specialty, all working at the same time.

  1. An always-on Mac mini

    The texts land here, and Rocky lives here. It also runs the builds and serves the previews.

  2. A PC with a GPU

    Local models. When a text needs one, the work goes here, and the answer comes back to wherever I asked.

  3. My laptop

    Where I open the result, like the preview of a website change I asked for from my phone.

Everything runs in parallel, and every reply comes back to the thread I texted from. Somehow it works on TUwifi.

Your concerns are valid

An agent with a phone number needs rules. These are its six.

  1. It never sees a password or a key. 1Password handles the logins.
  2. Only I can text it.
  3. Anything risky asks me first.
  4. A second model audits every command it wants to run.
  5. One word shuts it all down.
  6. Locked down by me, and by the Hermes team.

What it runs on

Hermes, which is open source, and not mine.

Rocky runs on Hermes, open-source software from Nous Research that turns a chat model into an always-on agent that can act, use tools and remember. I didn't write Hermes. What I built is everything around it: hooking it up to iMessage on an always-on Mac mini, the plumbing that hands work to Claude Code sessions on whichever machine fits the job, and the guardrails.

OpenClaw is the same idea from a different project, if you want to compare.

Everything it depends on

  • Claude Code
  • ChatGPT
  • Hermes
  • 1Password
  • Granola
  • Todoist
  • GitHub
  • Google Drive
  • Google Calendar
  • Vercel
  • Tailscale
  • BlueBubbles
  • iMessage
  • llama.cpp
  • Docker

None of it is software I wrote from scratch. Agents wired together the tools I already used every day, so it fit itself to me. Nobody else can run this exact thing, and that's the point. Yours would fit you, and every piece is swappable, models included.

The code is private, but the talk explains how it works.

It used to be NEPTR

The talk still goes by that name.

I Gave My AI a Phone Number

TU's AI and ML Club, September 2026

I gave this talk to TU's AI and ML Club, where I'm the vice president. It opened with me texting the agent from the stage, back when it was still NEPTR, and asking it to build the last slide of the talk and text me when it was done. Then I gave the talk. At the end I refreshed the deck to see what it had made.

NEPTR is the pie-throwing robot from Adventure Time, the Never-Ending Pie-Throwing Robot. It was made as a prank and became a being. In my house it stood for Needs Excessive Permissions To Run.

One slide near the end of the talk took the common questions.

Am I ok?
No.
Am I entering AI psychosis?
Maybe.
Is this awesome?
Absolutely.

Keep going

There are 40 more projects. Here's the next one.