Runner environment

Every run step in a workflow executes inside a sandbox container based on Wolfi Linux. This page documents what's available inside the runner, what's auto-injected, and how to install extra dependencies.

Auto-injected environment variables

The runner injects the following variables into every step automatically — you do not need to declare them in env:.

Variable	Purpose
H2O_CLOUD_CLIENT_PLATFORM_TOKEN	Platform access token for authenticating against H2O Cloud APIs (AI Engine Manager, MLOps, Feature Store, Drive, …).

Use the token via the standard login() helper of each H2O Python client:

- name: List DAI engines
  run: |
    python3 -c "
    import os, h2o_engine_manager
    clients = h2o_engine_manager.login(
        platform_token=os.environ['H2O_CLOUD_CLIENT_PLATFORM_TOKEN'],
    )
    print(clients.dai_engine_client.list_engines())
    "

Pre-installed tools

The runner image is intentionally minimal. Only the following are pre-installed.

Package	Notes
python-3.13	Python runtime
uv	Fast Python package installer — use with sudo
bash	Default shell
curl	HTTP client
wget	HTTP client
jq	JSON processor
zip / unzip	Archive tools
aws-cli	AWS CLI
ca-certificates	TLS root certificates
sudo	Passwordless sudo for the sandbox user

Runner shell

• User: sandbox (uid 1000, gid 1000)
• Shell: /bin/bash
• Working directory: /home/sandbox
• Sudo: passwordless
• Architecture: amd64
• Each step starts a fresh shell. Env vars set inline with export, the current working directory, and any background processes do not persist into the next step. Use the step-level env: field for variables; use H2O Drive to share files across jobs.

Installing Python packages

Use uv with sudo — the system Python is owned by root.

- name: Install client
  run: sudo uv pip install --system h2o-engine-manager driverlessai

pip is not pre-installed. Do not use pip install directly.

H2O Python clients

Component	Install command
AI Engine Manager	sudo uv pip install --system h2o-engine-manager
MLOps	sudo uv pip install --system h2o-mlops
Driverless AI	sudo uv pip install --system driverlessai
H2O-3	sudo uv pip install --system h2o
Feature Store	sudo uv pip install --system h2o-featurestore

What is NOT available

• No pip — use uv.
• No git.
• No docker.
• No make, gcc, or build toolchain.
• No sed, awk, or full GNU coreutils — Wolfi uses busybox equivalents.
• No GPU drivers or CUDA.
• No persistent filesystem between jobs — use H2O Drive for cross-job data sharing.
• No pre-installed Python packages beyond the standard library.

Related

• Steps — run, upload, download, working_dir, env.
• Environment variables — declaring your own variables at workflow, job, and step level.
• Secrets — referencing values from H2O Secure Store.
• Storage — moving files in and out of H2O Drive.
• Examples — working workflows that use this runtime to integrate with H2O products.

---

Feedback

• Submit and view feedback for this page
• Send feedback about H2O Workflows to cloud-feedback@h2o.ai