Compute
Cloud code execution that actually runs.
Sandboxed Python, Node, and shell: agents and you, side-by-side.
Three languages, ready.
Anything you'd normally install just works.
for data, charts, and most everything
for web and API work
for files and quick scripts
Every run starts clean. Nothing leftover from before, nothing from anyone else.
Runs in its own little box. Only reaches outside when you gave the agent permission (like web search).
Charts, files, anything it makes lands in your Drive as a real file you can keep, share, or roll back.
ChatGPT's code runner is one chat at a time and the outputs vanish with it. Here, code runs against the files in your Drive and the results land back as real files you can keep using.
What you get
Every run starts in a clean environment. No leftover state, no surprises from prior sessions.
Three runtimes ship by default. Pip / npm install whatever you need on the fly.
Mount idapt folders into the sandbox so generated scripts work against actual data, not toy examples.
Files written during execution land back in your workspace, versioned, with a trace of which run produced them.
When an agent needs to compute, it executes here. You see the script, the output, and can re-run it yourself.
Who it's for
Drop in a CSV, ask for the pivot you need, get a chart and the pandas snippet that produced it.
Prototype a function, run a test against staging data, refine without touching your local env.
Reproduce a paper's numerical result: the sandbox makes it copy-and-rerun, not download-and-pray.
How it works
- 1Ask
Describe the computation in chat or give an agent the goal.
- 2Run
The model writes code; the sandbox executes it; you see the output inline.
- 3Iterate
Tweak the request: the run is fast and disposable, so failed attempts cost nothing.
- 4Keep what works
Save the script to a folder, schedule it, or hand it to an agent for repeat runs.
Frequently asked
Drop a spreadsheet. Ask a question.
The AI runs the numbers. You get the chart and the script behind it.