Codeinterpreter MCP integration

Because the code runs remotely, plt.show() has no display to render into. Scripts write the image to /home/user and Get File returns it instead. The same rule applies to any output worth keeping: write it to a file rather than relying on a rendered cell.

In /home/user. Upload File puts an attachment there, Execute Code and Run Terminal Command read and write there, and Get File reads from there, so a path outside that folder is the usual reason a script cannot find its input.

Execute Code returns stderr and the error alongside stdout, so a failure comes back with its traceback rather than a blank result. That lets the script be corrected and rerun in the same sandbox, and anything still unresolved goes to a human agent with the full conversation history attached.

The sandbox is a separate cloud environment created on demand by Create Sandbox, not your own infrastructure, so nothing executes on your servers. Anything sensitive only gets there because Upload File deliberately put it there, so treat what you upload with the same care as data sent to any third-party processor.