Can I import pandas?
Yes — and here's the exact, complete list of everything else pre-installed in the sandbox container, straight from src/boxkite/preset_packages.py. This is the whole list; nothing here is abbreviated.
Python packages, by category
Data processing
Visualization
Scientific computing
Excel & spreadsheet
Document processing
HTML & XML parsing
Data serialization
Image processing
Date & time
Output formatting
Templating
Text processing
HTTP clients
Node.js packages
Command-line tools available on PATH
This list is closed by design — not an oversight you can work around.
Agents cannot install additional packages at runtime — the sandbox has no network access, and pip/npm are removed from the container entirely. Everything usable has to be in the preset list above; a pip install attempt fails, and a missing-module error points back at this same list.
If you need something that isn't listed, the fix is to rebuild the sandbox image with it added — this is a deliberate tradeoff of the isolation model, not a gap to route around from inside a session.
"Rebuild the image" doesn't have to mean hand-editing a Dockerfile — there's a self-serve API for it.
The declarative builder lets you request a pinned-version python_packages/apt_packages/npm_packages list on top of one of four pre-approved bases, and get back a digest-pinned image_idto create sandboxes from — no raw Dockerfile access, no arbitrary base image. It's opt-in and off by default pending a live-cluster security review (see the custom images guide for exact scope and current status), but the mechanism itself is real, shipped code today, not a roadmap item.