You work on a desktop at home and a laptop everywhere else. The project folder lives in OneDrive or Dropbox so both machines have it. Then an AI agent edits forty files on the desktop while the laptop is asleep, the laptop wakes up, and you get report (Laptop's conflicted copy).docx, a half-synced node_modules, and a Git repo that says every file changed.
Sync services were built for people saving a document now and then. Agents write many files fast, delete and recreate them, and keep lock files open. That's the worst case for sync.
Fix 1: for code, Git is the sync
If it's code, take the folder out of OneDrive and use Git as the only way the two machines share it. Commit and push from one, pull on the other. Exclude build output and dependencies in .gitignore. It's boring and it works, and the AI is good at writing the commit message.
The catch: you have to remember to push before you leave the desk. When you forget, the laptop has yesterday's code.
Fix 2: one home for the work
The other pattern is to stop copying the work at all. Pick one machine, usually the desktop that's always on, as the home of the project. Everything happens there. Other devices don't hold a copy; they reach in.
You can do that with remote desktop, which is fine for an hour of real work and painful on a phone. Or you can do it from the AI chat you already use, which is what I do.
I build Local Commander, so I'm biased. An agent runs on the desktop; ChatGPT or Claude on any device can read, edit and run things there. From the laptop at a café:
On Home PC, open the Q3 report folder, update the totals in summary.xlsx from the three CSVs that came in today, and tell me what changed.
From the phone, later:
On Home PC, did the build of the site pass? If it did, show me the git log for today.
Nothing moves between devices except the conversation. The files stay on the desktop, the build runs on the desktop, and there's nothing to resolve when you get home.
What changes in practice
- No conflicted copies. There's one copy.
- No "which machine has the latest version". The home machine does, always.
- Heavy work stays on the strong machine. The laptop doesn't need the 10 GB of dependencies or the GPU.
- The phone becomes useful. You can't edit a spreadsheet on a phone, but you can ask for the edit.
Where it stops
- The home machine has to be on and online. If it's a desktop, turn sleep off; if it's a laptop that travels, this pattern doesn't fit.
- It's not a file-sync service. If you need the files on the plane, carry a copy (Git again).
- It works through files and commands, not the screen. For layout work in a GUI app, use remote desktop.
- It isn't a sandbox. An approved computer lets the AI do what your user account can do; choose only the folders it needs when you set it up.
Setting it up
- On the home machine, run
npx local-commander-agent@latest(Node.js 20+) and approve it in the browser. - Choose access: the whole user account, or only the project folders.
- Add
https://localcmd.app/mcpas a connector in ChatGPT or Claude (and Cursor, if you use it on the laptop) and sign in. - Take the project folder out of OneDrive. Keep backups the normal way.
You can cut the machine off from your account at any time.