Sandbox Environments

Bake a base image, packages and internal registries into a reusable image so agents, contexts and apps start with everything they need installed

Every agent run, context build, and app runs its code in an isolated DataGen sandbox. A Sandbox Environment is a reusable recipe for that sandbox: a base image plus apt / pip / uv / npm packages, optional internal package registries, and CPU / memory / disk sizing.

DataGen bakes the recipe into an image in the background. Anything bound to the environment then starts from the baked image, so nothing is installed at run time and every run sees the same package versions.

Open Sandbox Environments

Open the More menu on the left rail and choose Sandbox Environments. Use Pin to sidebar to keep it on the rail.

If Sandbox Environments is not in the menu, the feature is not enabled for your account. Agents then run on the standard image. Contact DataGen to enable it.

Admins and Builders can create, edit, and delete environments. Every member can view them.

---

Create an environment

1. Start the dialog

   Click New environment.

2. Optionally start from a template

   Pick a template under Start from template (optional). It pre-fills the packages; you can edit anything before creating.

3. Name it and list packages

   Enter a Name, then fill in the package fields you need.

4. Create

   Click Create. The build starts in the background and the card shows its progress.

Built-in templates

Template	pip packages
Data Science	pandas, numpy, scipy, scikit-learn, matplotlib, seaborn, polars, pyarrow, openpyxl, statsmodels
Data Engineering	duckdb, polars, pyarrow, sqlalchemy, psycopg2-binary, pandas
Web Scraping & HTTP	requests, httpx, beautifulsoup4, lxml
Documents & Files	openpyxl, python-docx, pypdf, pillow, markdown, pandas

A template is copied into the new environment. Editing the environment later never changes the template, and vice versa.

Fields

Field	What it does
Name	Required. Shown in every environment picker.
Description	Optional. Shown on the card; templates fill it with the package list.
Base image (optional)	A custom base image such as python:3.11-slim. Applies on Daytona only; Modal uses a built-in Debian base where apt / pip / uv / npm all work.
apt packages	System packages, e.g. ffmpeg libpq-dev.
pip packages	Python packages installed with pip.
uv packages	Python packages installed with uv (faster than pip).
npm packages	Node packages.
Internal registries (optional)	pip / uv index URL, npm registry URL, and apt sources.list for installing from a private mirror or artifact repository. URLs must be http:// or https://.
Resources (optional)	CPU cores (1 or 2), Memory (1 GB or 2 GB), Disk (2 GB or 5 GB, Daytona only). Default uses the deployment default shown in the dropdown.
Workspace default	Marks this environment as the workspace default. It is listed first and labeled (default) in pickers.

Package lists accept spaces, commas, or new lines as separators. Pin versions without spaces, for example pandas==2.2.2, psycopg[binary], or @scope/tool@1.4.0.

Modal vs Daytona. DataGen Cloud runs sandboxes on Modal; on-prem deployments use Daytona. On Modal the Base image and Disk settings are ignored. On Daytona, resources are part of the image itself, so changing them rebuilds the image.

---

Build status

DataGen builds the image whenever the build inputs change — base image, packages, registries, or resources. Each card shows an agent status badge:

Status	Meaning
pending	Queued; the build is about to start.
building	The image is being built. A progress bar shows elapsed time.
ready	The image is built. Runs can start from it.
failed	The build failed. The card shows build failed: <error> — usually a misspelled package or a version that does not exist. Edit the environment to fix it; saving starts a new build.

Notebook apps use a separate notebook image built from the same recipe. Its badge appears once it has been requested; use Build notebook image (or Retry notebook build) on the card. Notebook sandboxes always get at least 10 GB of disk.

The App images table below the list shows dependency images each custom app bakes when you save it. They are managed by the app; you can delete one that no live version boots from.

---

How agents, contexts and apps use an environment

Where	Setting	If the image is not ready
Agent	Sandbox environment card -> Environment, then Save environment	The run refuses to start with Sandbox environment image is not ready (build status: ...).
Context builder	Context settings -> Sandbox environment	The builder starts from the standard image until the build finishes. Changes apply the next time the builder session starts cold.
Notebook app	App Builder -> Sandbox	Binding requires the environment’s notebook image to be ready.
Custom app	App Builder -> Sandbox	Changing it rebuilds the app’s dependency image on the new environment, then recreates the sandbox.
Data builder	Data -> Builder configuration -> Sandbox environment	—

Leaving an agent on Default (standard image) runs it on DataGen’s standard image.

---

Gotchas

Agent run fails right after editing an environment

Every change to packages, registries, base image, or resources starts a rebuild, and agents bound to the environment refuse to run until it is ready again. Make edits when no scheduled run is due, or wait for the badge to turn green.

Deleting an environment

Agents that use it fall back to the standard image — without its packages. Rebind them to another environment first if their code imports anything from it.

Installing at run time instead

Code can still pip install inside a sandbox when it has outbound network, but it pays the install on every run and can drift between runs. Bake anything used more than once.

Alpine / musl base images

Many binary wheels, such as psycopg2-binary, ship only for glibc. On a musl base, choose packages that publish musl wheels (for example psycopg[binary]).

Next

Harness

Choose the agent runtime, pin its version, and route it to your own LLM gateway.

Query a Database from Code

A worked example using the Data Engineering template.