Databricks Sandbox

Databricks Sandbox is a Databricks compute / bi capability, introduced 2026.

An isolated, on-demand serverless environment for running untrusted code - an agent harness, AI-written code, background workers, or an IDE/SSH session - driven through the Databricks REST API, SDK, or CLI, with a persistent home directory.

• The home directory is /home/sandbox-agent, which is a fairly direct statement about who the docs think the primary user is. 🔗
• The first time you SSH in, Databricks quietly rewires coding harnesses such as Claude and Codex to send their inference through AI Gateway, so the agent is governed before it types anything. 🔗
• Stopping a sandbox does not stop the storage bill - the home directory is billed for as long as the sandbox exists, so the only way to stop paying for storage is to delete it. 🔗

⚠ Limitations: Fixed at 4 vCPUs and 8 GB of RAM, with up to 100 GB in the home directory and 10 GB outside it (deleted on stop); default quotas of 40 active sandboxes per user (400 total) and 100 active per workspace (1,000 total); network egress is unrestricted and not configurable, ingress is closed; public IP addresses can change at any time; it is not a Spark runtime; the startup environment can't be customized; and Databricks might purge sandbox data at the end of the Beta. 🔗

Category

Compute / BI

Introduced

2026🔗

Also known as

sandbox, databricks sandbox

Verified

2026-09-27