Desktop
An AI agent for your local workspace.
Many real workflows live in files, folders, repositories, terminals, and desktop apps. ZeroTwo brings the worker to that context with approval gates and clear run history.
Local files and folders
Read, write, organize, summarize, and transform workspace files.
Repositories and CLI
Inspect diffs, run checks, and produce code deliverables with review points.
Approvals and safety
Pause before high-impact actions like writing, deleting, sending, or running sensitive commands.
What the desktop agent can reach
In the browser, ZeroTwo works with what you upload and the apps you connect. On the desktop it also works with the machine itself: the directory you point it at, the files inside it, the repositories you keep locally, and the commands you would otherwise run in a terminal yourself.
Working inside a directory
You give ZeroTwo a folder, and that folder is the boundary. Inside it the agent can list, search, read, edit, and create files — renaming a directory of exports to one consistent scheme, pulling figures out of a stack of PDFs into a single spreadsheet, applying the same change across a project, or reading a long document that never has to leave the machine to be summarized.
Repositories, diffs, and the command line
For code the useful unit is usually the diff rather than the file. ZeroTwo can read what changed on a branch, run the project’s own test and lint commands, and report what broke — then produce the fix as a reviewable edit instead of a wall of text to paste back in. It uses the toolchain already installed, so the results match what you would get running those commands yourself.
Why the approval gate matters here
Reading is free; changing is not. Writing a file, deleting one, running a command, or sending something outward pauses for confirmation and shows what is about to happen. That gate is the reason it is reasonable to point an agent at a real working directory instead of a throwaway copy of one.
Choosing desktop over the web app
Desktop is the better choice when the context is local and bulky: a repository, a folder of documents, a dataset that is awkward to upload, or anything you would rather not move off the machine at all. The web app stays simpler for work that already lives in connected services. Both sign in to the same account and share the same history, so a thread started in one is visible in the other.
What it does not do
The desktop agent is not a background process watching your machine. It acts when you hand it a task, within the directory you granted it, and stops when the task is finished. It also does not replace your editor or your shell. The point is to hand off the mechanical passes — the rename across two hundred files, the first read of an unfamiliar repository, the check that every export actually parses — while the decisions stay with you.