What “controlling your PC from your phone” actually means here
There are two completely different things people mean by this phrase, and picking the wrong one wastes an afternoon.
The first is screen mirroring — TeamViewer, AnyDesk, Chrome Remote Desktop, RustDesk, VNC. Your desktop is streamed to the phone as video and you poke at it with a finger. It is the right tool when you need to see a graphical application, click a button in an installer, or fix something visual.
The second is remote task execution. You describe what you want in a sentence, the computer does it, and you get the answer back as text or a file. No screen, no cursor, no pinch-to-zoom. This is what the OpenClaw Desktop App does, and for the majority of the things people actually want from their computer while away from it — find a file, check a log, run a build, summarise a folder, restart a stuck service — it is dramatically faster than squinting at a mirrored 27-inch display on a 6-inch screen.
This guide covers the second kind. If you specifically need to see your desktop, use RustDesk or Chrome Remote Desktop instead — both are free and better suited to it.
What you can actually ask for
These are the request types that work reliably, drawn from what people use day to day:
- Find and fetch files. “Send me the invoice PDF I saved last Tuesday.” The agent searches, finds it, and sends the file back through the chat.
- Check on long-running work. “Is the render finished?” “What is the last line of the deploy log?” Far quicker than opening a remote desktop session to read one line of text.
- Run scripts and commands. “Run the backup script.” “Restart the local database.” “Pull the latest changes on the work repo.”
- Summarise and search. “What changed in the project folder since Friday?” “Summarise the notes file from this morning’s call.”
- Housekeeping. “Move everything older than 90 days out of Downloads into an archive folder.” “How much disk space is left?”
- Scheduled jobs. Ask once and have it run every morning — a daily digest, a backup check, a disk-space warning pushed to your phone.
Where it struggles: anything that needs a graphical interface. “Click the third button in the Photoshop toolbar” is not a good fit. Anything expressible as a file operation, a shell command, or a question about text works well; anything that fundamentally needs eyes on a screen does not.
How the connection works
The mechanism is worth understanding, because it explains why there is no network configuration to do.
The desktop app runs a local agent on your computer. That agent opens an outbound connection to the messaging platform you connected — Telegram’s bot API, WhatsApp’s web protocol, Slack, or Discord. When you send a message from your phone, it travels to the platform’s servers as normal, and your computer picks it up over the connection it already holds open.
Because every connection is outbound, nothing on your machine listens for inbound traffic. There is no port to forward, no dynamic DNS to maintain, no static IP to buy, and no open port for anyone to scan. It works from behind a home router, a corporate firewall, or a phone hotspot without touching any of them.
The AI model is a separate choice. Point the agent at Claude, GPT or Gemini through an API key, or at a local model through Ollama. With a local model, your messages and files never leave your computer at all — the only thing crossing the internet is the chat message itself.
Setup, step by step
1 Install on the computer you want to control
Download the free OpenClaw Desktop App for macOS or Windows and run the installer on the target machine — the one that will do the work. Everything is bundled, so there is no terminal, no Node install and no WSL2 on Windows.
This machine needs to stay powered on and awake. On a laptop, that means disabling sleep or leaving it plugged in.
2 Choose an AI model
In the app, open the provider screen and pick one:
- Local (free). Install Ollama and select a local model. Nothing leaves your machine and there is no per-token cost. Best for privacy and for people who do not want a bill.
- Cloud (paid, smarter). Paste an API key for Claude, GPT or Gemini. Noticeably better at multi-step tasks and ambiguous instructions. You pay the model provider directly for what you use.
- OpenRouter. One key, many models, including free tiers. A good middle ground while you are experimenting.
You can change this later without redoing anything else, so do not agonise over it now.
3 Connect the chat channel
Open Channels and add the app you want to message from:
- Telegram — create a bot with
@BotFather, copy the token into the app. Most reliable option for this use case. - WhatsApp — scan a QR code with your phone, the same flow as WhatsApp Web.
- Slack / Discord — install the app into your workspace or server. Best when a team shares the machine.
4 Send a test message
From your phone, message the connected account:
What files did I add to Downloads this week?The agent runs the search on your desktop and replies in the thread. A reply means the whole chain works. If nothing comes back, the usual cause is that the channel was added with restricted messaging — check the channel’s access settings in the app.
5 Lock down the permissions
Before you rely on this, open the permissions screen and narrow the blast radius:
- Scope file access to the specific folders you care about rather than your entire home directory.
- Keep approval prompts on for destructive operations — deletes, overwrites, anything touching system paths.
- Make sure the channel only accepts messages from you, not from anyone who finds the bot.
Which chat app to use
The channel you pick changes the experience more than people expect.
Telegram — the default recommendation
Telegram was designed with bots in mind. Automated accounts are a supported first-class feature rather than a tolerated edge case, file transfers are clean and generous, and the bot API is free. If you have no strong preference, start here.
WhatsApp — the most convenient, with a caveat
For most people WhatsApp is simply where they already are, and messaging your own computer in the same app as everyone else is genuinely pleasant. The caveat is that WhatsApp is a consumer messenger and automated use is not officially blessed on personal accounts. Volume and pattern matter — a handful of messages a day to your own machine is a very different profile from blasting hundreds of automated messages. We covered the real risk factors in this breakdown of what actually gets accounts flagged.
Slack and Discord — for shared machines
If more than one person needs to reach the computer — a build server, a shared render box — a channel in Slack or Discord gives you a visible, threaded log of who asked for what. That audit trail is worth a lot when something goes wrong.
How this compares to remote desktop tools
An honest comparison, because these tools are not competing for the same job:
- Bandwidth. Screen streaming needs a steady connection and degrades badly on patchy mobile data. Text instructions need almost nothing and work fine on one bar.
- Speed for small tasks. Checking one log line takes seconds by message and a minute or more by remote desktop — connect, authenticate, wait for the stream, find the window, zoom in.
- Phone ergonomics. Driving a desktop UI through a phone screen is genuinely unpleasant. Typing a sentence is not.
- Visual work. Remote desktop wins outright. If you need to see the screen, you need to see the screen.
- Setup and exposure. Most remote desktop tools need an account with a third party that can, in principle, broker access to your machine. The chat approach keeps the agent local and the connection outbound.
Plenty of people run both: a remote desktop tool for the rare occasion they need eyes on the screen, and a chat channel for the twenty small things a week that do not.
Security: the parts worth taking seriously
You are giving an AI agent the ability to run commands on your computer. That is a real capability and it deserves real limits.
Scope its access deliberately
The agent has exactly the access you grant it. Restrict file access to the folders that matter instead of pointing it at your home directory. If you never want it touching system paths, say so in the permission settings rather than trusting it to decline.
Keep approval prompts on
Approval prompts on destructive actions are the single highest-value setting here. An agent that asks before deleting is an agent whose worst mistake is a wasted question.
Control who can message the channel
This is the failure mode people actually hit. A Telegram bot with open messaging can be found and messaged by strangers. Restrict the channel to your own account, and treat the bot token like a password — anyone holding it can talk to your computer.
Consider a local model
With a cloud model, the content of your requests and the file excerpts the agent reads are sent to that provider. With a local model, none of that leaves your machine. If the computer holds client work, source code, or anything under a confidentiality obligation, this is the setting to change first.
Common problems and fixes
No reply at all
Almost always the channel’s access policy. Newly added channels can default to a restricted mode that silently drops direct messages from unpaired users. Open the channel settings in the app and confirm your account is allowed.
It replies but refuses to do things
A permissions boundary, working as designed. Check what the agent is scoped to reach — the folder or command you asked about is probably outside it.
It stops working when the laptop closes
Expected. The agent runs on your machine; a sleeping machine runs nothing. Disable sleep on the target computer or use an always-on desktop as the target.
Replies are slow or the agent misunderstands
Usually the model. Small local models are fast but literal-minded, and struggle with multi-step or ambiguous requests. Switching to a stronger cloud model for a week is the quickest way to tell whether the problem is the model or the instruction.
Getting started
The whole setup is one installer and one channel connection. If it turns out not to suit how you work, uninstalling removes it cleanly — there is no account to close and no subscription to cancel.
Frequently asked questions
Can I really control my PC from my phone for free?
Yes. The OpenClaw Desktop App is free and open source, and it runs entirely on your own computer. The only thing that can cost money is the AI model you choose: cloud models such as Claude, GPT or Gemini are billed per use by their provider, while a local model run through Ollama costs nothing at all. There is no subscription, no per-task fee and no seat licence for the app itself.
Is this the same as remote desktop software like TeamViewer or AnyDesk?
No, and the difference matters. Remote desktop software streams your screen to your phone so you can click on a tiny cursor. This approach sends an instruction in plain language instead, and the AI agent carries the task out on the desktop and replies with the result. It uses far less bandwidth, works fine on a slow mobile connection, and is much better for tasks like searching files or running scripts. It is worse than remote desktop when you genuinely need to see and click a graphical interface.
Do I need to open ports on my router or set up a static IP?
No. The computer connects outbound to the messaging platform, so there is no inbound connection to accept and nothing to forward. This is what makes it work behind home routers, corporate firewalls, and mobile hotspots without any network configuration. It also means you never expose a listening port to the public internet.
Which chat app should I use to control my computer?
Telegram is the most reliable choice for remote PC control because it is built around bots, handles file transfers cleanly, and does not restrict automated accounts. WhatsApp is the most convenient if it is where you already live, but it is a consumer messenger and automated use always carries some account risk. Slack and Discord are best when a team shares the machine.
Does my computer have to stay switched on?
Yes. The agent runs on your desktop, so the machine must be powered on and awake to receive and execute commands. On a laptop, disable sleep on the target machine or connect it to power. Some users leave a spare mini PC or an always-on desktop as the target and control it from anywhere.
Is it safe to let an AI agent run commands on my machine?
It is as safe as the limits you put on it. The agent has whatever access you grant, so treat it like a user account rather than a sandbox. Keep approval prompts on for destructive actions, scope its file access to specific folders instead of your whole home directory, use a dedicated account for the messaging channel, and never connect a channel that strangers can message. Because the model can run locally, you can also keep every command and file on your own machine.
Can I control a Windows PC and a Mac the same way?
Yes. The OpenClaw Desktop App ships for both macOS and Windows and the chat side works identically. The difference is only in what you ask for — shell commands and file paths follow each platform's conventions, so a script written for one will not always run on the other.