Herdr PocketVibecoding usually means sitting with an agent: you describe what you want, it writes and runs code, you look at the result and steer. Most of that time, though, you are not typing. You are waiting while the agent reads files, runs tests and edits code. Remote vibecoding moves that waiting off your chair. The agent keeps working on your own computer, and you step in from your phone when it actually needs you.
This post describes a workflow that works with Claude Code and Codex running in herdr on a Mac or Linux machine, with Herdr Pocket on the phone. It also lists what the phone is not good at, because knowing the limits is most of what makes this pleasant.
What runs where
Nothing runs on the phone. Your agents run in herdr on your computer, in real terminals, with your repositories, tools and credentials where they already are. herdr tracks each agent's state: working, blocked when it needs input or approval, done when it finished and you have not looked yet, and idle once you have.
Herdr Pocket connects to that computer over SSH, directly or through a private network such as Tailscale, and shows those agents. Lock the phone, lose signal in a tunnel, and the agents carry on. When you come back, the app reconnects and shows where things stand.
If you don't have that setup yet, start with setting up a Mac with herdr and Tailscale.
The loop
1. Start the work at your desk
Remote steering works best when the hard thinking happens before you leave. Give each task its own herdr workspace, ideally in its own git branch or worktree, and write a prompt with a clear finish line: "add retry with backoff to the payment client, keep the public API, run the unit tests, and stop when they pass." A vague prompt is fine at a keyboard. On a phone it becomes a long back-and-forth with your thumbs.
Claude Code's plan mode is useful here. Let it propose a plan while you are still at the desk, approve it, then leave it to execute.
2. Let the list tell you where to look
Herdr Pocket shows the agents from all your computers in one list, sorted Needs you, Done, Working, Idle. You don't open each agent to check on it. If nothing is at the top, nothing is waiting.
3. Answer the question it is waiting on
When an agent asks a question or wants permission, the Needs you sheet shows it with the same choices the terminal offers, as buttons. Pick one, add a note to an approval, or type your own answer. If an agent has gone the wrong way, you can stop it there too.
4. Read what it did
For Claude Code, Codex and several other agents on computers running herdr, Chat shows the conversation the way you read messages: Markdown, code blocks and the tools the agent ran. Changes shows the branch, the files the agent touched and a line-by-line diff. And every agent has its real terminal, streamed live, when you need to see exactly what is on screen.
5. Steer, then go back to your day
The Composer gives you the full iOS keyboard with autocorrect and dictation, so a follow-up like "good, now add a test for the timeout case" takes a few seconds to say. Snippets keep the prompts you send often one tap away. You can attach a screenshot or a file, up to 64 MB; it is copied to your computer and its path goes into your message. On a computer running herdr you can also start a new agent in a fresh git worktree from the phone.
Habits that help
- Keep the approval prompts. It is tempting to switch an agent to a mode that never asks, so it never blocks. A blocked agent costs you a tap; an agent that pushed or deleted something you did not want costs you an evening. Claude Code and Codex both let you choose how much the agent may do without asking; leave the risky actions behind a prompt.
- One task per workspace. Parallel agents only stay readable when each has its own workspace and branch.
- Ask for summaries. End prompts with "when you're done, summarise what changed and what you didn't do". A summary reads well on a phone; a scrolled terminal does not.
- Say it, don't type it. Dictation in the Composer is faster than thumbs for anything longer than a sentence.
- Glance at diffs, review at the desk. Changes is good for checking that the agent touched the right files. A real review of a large change still belongs on a big screen.
The honest limits
- Notifications only while the app is open. In the current beta, Herdr Pocket tells you with a sound, a buzz and a banner when an agent starts needing you, but only while it is open. There are no remote push notifications in this version, so a blocked agent waits until you look.
- The connection pauses in the background. Shortly after you leave the app, iOS suspends it and Herdr Pocket disconnects; it is not an always-on VPN. It reconnects when you return, and the agents never notice.
- Your computer has to be on. A sleeping or offline Mac can't run agents or answer the phone. Keep it awake and connected.
- Small screen, small changes. The phone is great for approvals, short replies and status checks. It is not the place for a large refactor review or debugging a tricky failure.
- Chat has a list. Chat works for Claude Code, Codex, Pi, Oh My Pi, Grok Build and OpenCode on macOS and Linux computers running herdr. Other agents, and Windows or Herdr Core computers, use the live terminal.
What it adds up to
Remote vibecoding is less about coding on a phone and more about not being the bottleneck. The agents keep the keyboard work. You keep the decisions, and you can make them from the kitchen, the train or a walk, a few taps at a time.
Want to try the loop? Herdr Pocket is in a free public beta for iPhone and iPad: join it on TestFlight, connect the computer your agents run on, and see who needs you.