A local digital coworker, living on your own machine.
Reads your files, edits code and runs commands on your own computer. Every task gets its own Git worktree, and every change waits for your review.
Not only engineers. Most of this arrives as plugins — mention one, authorize it once, and it works on the files on your disk. You watch it run and check the result before it counts.
Editing, diagramming, spreadsheets, site publishing and scripting arrive in the same window.
Contracts, footage and customer data are read locally, which keeps procurement and legal calm.
Every thread shows what it spent, so a seat's cost is a number, not a guess.
Each new thread gets its own Git worktree, so a refactor, a bug fix and a test run never overwrite each other. Switch between threads at any time to watch progress.
Plan mode lays out the steps before touching a file. When the work is done you get a readable diff: stage it hunk by hunk, commit, push, open a pull request, ask for a revision, or throw it away.
Issue triage, alert sweeps, dependency bumps, flaky-build hunting: set a time and hand it over. Each run happens in its own background worktree and lands in the inbox for review.
Open a browser panel beside the thread: it loads the page, clicks through the flow, reads the DOM and takes over the desktop when a task needs it. You watch both halves at once, no screenshot ping-pong.
Point it at the footage: it trims the dead air, orders the takes, writes and burns in subtitles, drops in the soundtrack and exports at the size each channel wants.
Some jobs only exist inside an app: importing footage into an editor, filing an entry in a legacy system, walking a checklist through three windows. Grant access and it takes the mouse and keyboard, step by step, in full view.
Type @ to mention one. The first call asks for authorization and names exactly what it may touch. Every call after that is written into the thread: the tool, the arguments, how long it took, and the file it produced.
MCP-based, so your own internal tools plug in the same way.
Three levels per thread: read-only, auto inside the worktree, or full access. Switch mid-conversation.
Write your team's conventions once — commit style, review checklist, release steps — and it applies them on every task.
Deski builds, debugs and ships. Deski Work creates, learns and explores. Same window, one click apart.
One pool of credits covers every model on the Model marketplace and every AI app — chat, images, video, digital humans, the infinite canvas and Deski Agent all draw from the same balance. Pick the tier that fits and move up any time.
Free compute every day, so everyone can actually use AI.
For the price of a coffee, your AI productivity stays on every day.
Built for heavy users — turn AI into your digital production team.
Sign in, pick a project folder, and send the first task.
Deski is a local AI digital coworker from XiaoZhan AIhub: a desktop app for macOS and Windows. It reads files, edits code, runs commands and writes documents on your own computer, for hours at a stretch. Every task runs in its own Git worktree, starts with a plan, and its changes wait for your review before they are merged.
Yes, it is free to download and use. The free plan gives 100 credits every day you sign in, usable on every model. Heavier use can move to the Plus or Pro plan for more compute and member benefits; prices are at https://deski.dflop.top/en#pricing .
macOS 12 Monterey or newer (separate builds for Apple Silicon and Intel) and Windows 10 / 11 (64-bit). It keeps itself up to date and needs no administrator rights.
The repository stays on your computer: files are read and written on your own disk and commands run in your own shell, with nothing uploaded to a remote container. While working on a task, the parts Deski needs the model to see are sent as conversation to the AI model you picked.
A browser chat only sees the code you paste into it, one thread at a time. Deski works on your machine: it reads and writes files on your disk, runs several tasks in parallel worktrees, has a built-in diff and review queue, runs scheduled tasks, and reaches your tools through plugins (MCP) and an in-app browser.
Yes. Weekly updates, meeting notes, specs and slide outlines, cleaning up spreadsheets into reports, editing video, walking a flow through another app: authorize it once, watch it work and check the result.
Deski runs on XiaoZhan AIhub's models; you can switch between the GPT, DeepSeek, GLM, Kimi and Qwen families, and every thread shows the credits it used.