How to run any terminal AI agent from your phone
Every few weeks a new AI agent launches that runs on your own machine instead of someone else's server. Every single one has the same gap: no phone app.
Look at the last six months. OpenClaw shipped as an open-source, local-first agent that lives inside your chat apps and became one of the fastest-growing repos on GitHub. Hermes Agent from Nous Research followed with persistent memory and passed 135,000 stars in under three months. Paperclip landed in March as a self-hosted Node server with its own dashboard and a heartbeat scheduler. Claude Code and Codex, the two coding agents most people already run daily, are the same shape again.
Different authors, different feature sets, same architecture underneath: a process that runs in a terminal, on a machine you own, doing real work - editing files, running shell commands, checking a calendar, sending an email. Local-first is the whole point. It is also exactly why none of them have a phone app. A phone app means a server in the middle, and the entire pitch of this generation of agents is that there is no server in the middle.
The workaround everyone reaches for, and why it breaks
Install any of these and the first time you leave your desk, you hit the same wall. The standard fix is to find your Mac's local IP address and type something like http://192.168.1.23:PORT into your phone's browser. It works for exactly as long as your phone is on the same Wi-Fi as your Mac, which stops being true the moment you leave the house, walk out to the driveway, or your router hands out a new address.
- It only works on the same network, so it is useless from anywhere that matters - the coffee shop, the train, the other room with a closed door.
- A new IP breaks the bookmark, with no warning that it broke.
- A raw local port with no auth in front of it is not something you want reachable at all, let alone something you want to expose further to make it work remotely.
- And it only solves it for the one agent you just set up. The next one you install, you do it again.
That last point is the one that actually matters here. Solving "how do I reach OpenClaw from my phone" and solving "how do I reach Hermes Agent from my phone" look like two separate problems, and if you solve them one product at a time, you will keep solving them - because a new agent with this shape ships almost every month now.
The question that actually generalizes
They are not really separate problems. Every one of these agents runs on a machine, not in the cloud. If you can reach the machine, from anywhere, securely, you can reach whatever is running on it - the agent you set up last week, and the one that has not been released yet.
That is a smaller problem than "build a mobile client for every agent," and it is the one worth actually solving. It splits into three things a phone needs to do, and only one of them is agent-specific:
- Reach the Mac from anywhere, not just the same Wi-Fi network. This is pure networking and has nothing to do with which agent is running.
- See and control it when something needs a real look, not just a status line - a browser window the agent opened, a file mid-edit, a dialog it is waiting on.
- Get told the moment it needs you, for the agents that speak up when they are stuck, so you are not the one polling a chat thread to check.
Only the third one cares what you installed. The first two do not know or care whether the terminal in front of them is running OpenClaw, Hermes Agent, Paperclip, Claude Code, or something that ships next month.
How MightyMouse fits this
To be precise about what this is: MightyMouse turns your iPhone into a screen, trackpad, keyboard and microphone for your Mac. It is not an integration with OpenClaw, Hermes Agent, Paperclip, or any other agent, and it does not claim to be. It is the access layer for the Mac those agents happen to be running on, which means it does not need updating every time a new one ships.
Once it is installed on the Mac and paired to your phone with a QR code, three things are true without any agent-specific setup: your phone can see the actual Mac screen live over a secure tunnel, not just a local network, so a closed-door office or a different Wi-Fi does not matter; the phone acts as a real trackpad and keyboard, so a "yes, go ahead" or a one-line correction typed into a terminal actually lands there; and dictation runs through Whisper on the Mac itself, so talking to it does not mean your voice goes to a transcription server.
For the subset of agents that speak MCP - which today means most coding agents, including Claude Code and Codex - there is one more piece: the agent can call a tool that pushes a notification straight to your phone the moment it needs an answer, and you can reply from the lock screen without opening the phone at all. That part is agent-aware by necessity, because it hooks into a protocol the agent speaks. The screen, trackpad and voice pieces are not - they work the same whether the terminal is running an MCP-speaking coding agent or a chat-first agent that has never heard of MCP.
The honest tradeoff
This does not turn your phone into a second machine. The Mac still has to be on and awake - if the lid is closed and it is asleep, there is nothing on the other end to reach, agent or otherwise. MightyMouse gets you to the Mac; it does not replace it.
And it is a general-purpose remote, not a purpose-built mobile interface for any specific agent. Mirroring a real desktop onto a 6-inch screen is genuinely fine for checking in, approving a step, or nudging something back on track. It is not a substitute for sitting at the actual keyboard for a long editing session. If what you want is a phone-native chat UI for one particular agent, some of them - OpenClaw among them - already give you that for the conversational half, and this is for the other half: the part that happens on the actual machine.
Why this stays true after the next one launches
The reason to solve it at the machine level rather than the agent level is survivorship. Of the agents that were the "one to watch" eighteen months ago, most are not the ones anyone is running today. The pattern - local-first agent, terminal-based, no phone client - has outlived every individual product built on top of it, and there is no sign that changes. An access layer for the Mac does not care which agent wins that race, because it was never betting on one.
Reach whatever is running on your Mac, from your phone, from anywhere.
Your iPhone as a screen, trackpad and mic for your Mac. Free to try.
Common questions
How do I access my self-hosted AI agent from my phone?
The agent itself almost never ships a phone app, because it is designed to run on a machine you own rather than a server. The fix is to make the machine reachable from your phone - securely, from anywhere, not just your home Wi-Fi - rather than waiting for each agent to build its own mobile client.
Does this work over just my home Wi-Fi, or from anywhere?
MightyMouse defaults to a secure tunnel for signed-in users, so your phone reaches your Mac whether you are on the same network or not. It falls back to the local network only if you are not signed in.
Is this specific to one AI agent?
No, and that is the point. It works the same whether the terminal on your Mac is running OpenClaw, Hermes Agent, Paperclip, Claude Code, Codex, or an agent that has not been released yet, because it operates at the level of the Mac, not the agent.
Do I need to leave my laptop open and awake?
Yes. The agent runs on your Mac, so the Mac has to be on for there to be anything to reach. This solves being away from your desk, not being away from the machine entirely.
Whatever terminal agent you run today, or install next month: MightyMouse turns your iPhone into a screen, trackpad and mic for your Mac, with free local voice transcription, so reaching it never depends on which one you picked.