A disposable cloud terminal for your agent: run shell commands and long-running processes, search, read and edit files — all inside an isolated sandbox.
No configuration required
Deploy and start querying — nothing to connect.
Standard
Sessions run up to 60 minutes before the sandbox is reclaimed.
60+ AI models
Chat with 60+ AI models on the same workflow — switch to a different model mid-conversation and re-run the same prompt, or use Compare mode to put several side-by-side.
How models use it and what it is built for.
Desktop Commander is one of the most-installed MCP servers: it gives a model a terminal, a process manager and a file editor. Run locally, that means an AI with shell access to your own laptop. Run here, it means the same tools pointed at a throwaway Linux sandbox instead.
That is the point of hosting it. The agent can install packages, start a dev server, tail a log or rewrite a config file, and none of it reaches your machine, your credentials or your files. When the deployment stops, the sandbox and everything in it is destroyed.
Desktop Commander sends usage telemetry by default. Hosted deployments run with the package's own telemetry kill switch set, so nothing about your session is reported to a third party.
Concrete operations exposed as tools.
In order. Each step assumes the previous one worked.
There is nothing to fill in. The sandbox boots, installs Desktop Commander and hands you a live MCP URL in under a minute.
Open the deployment in Agent Studio, or paste the URL into any MCP client that speaks Streamable HTTP.
Ask the model to run `uname -a` and list the home directory. You will see it is a fresh Ubuntu machine with Node and Python available.
Paste any of these into Agent Studio once the server is connected.
Clone https://github.com/expressjs/express, install dependencies and run the test suite. Summarise any failures.
Write a Python script that fetches the Hacker News front page and prints the top 10 titles, then run it.
Start a static file server on port 8080 in the background, then curl it to confirm it responds.
Find every TODO comment in the cloned repo and group them by file.
Known constraints, stated plainly.
Deploy the server hosted, then watch which tools a model actually reaches for — with full JSON input and output on every call. Switch models mid-conversation to compare how each one uses the same server.
About the Desktop Commander MCP server.
Safer than giving it your own. Every deployment is an isolated virtual machine with nothing of yours on it, and it is destroyed when you stop it. The worst a bad command can do is break a sandbox you were going to throw away.
No. Locally, Desktop Commander controls the machine it runs on. Hosted, the machine it runs on is the sandbox — your computer is never involved.
No. Hosted deployments set DESKTOP_COMMANDER_DISABLE_TELEMETRY, the package's built-in kill switch.
No. Each deployment starts from a clean machine. Copy anything you want to keep out before the deployment stops — for example by having the agent push to a Git remote.