Databricks CLI for coding agents

If a coding agent can use the shell in a Databricks project, it can use the Databricks CLI to check project configuration and review proposed changes. The CLI sends requests to the Databricks platform through its REST API. Databricks provides this interface for reading and changing workspace and account resources.

A coding agent can review a Databricks Asset Bundle before deployment. A bundle is a set of local configuration files that describes Databricks resources, such as jobs. The agent can check the bundle, create a JSON plan of proposed deployment changes, and stop before changing the workspace. The commands below are based on Databricks documentation. We have not installed or tested them for this article.

Install and authenticate

On macOS, install the CLI with Homebrew:

brew tap databricks/tap
brew install databricks
databricks version

Use the official CLI installation instructions for other operating systems and current installation details. Databricks classifies CLI versions 1.0.0 and later as generally available. Versions 0.18 and earlier are the legacy CLI and no longer receive support or new features, so check the version before giving an agent CLI work.

Use browser-based login to authenticate to the workspace during interactive development:

databricks auth login --host https://<your-workspace-host>
databricks auth profiles

Replace <your-workspace-host> with your workspace URL. The login command opens a browser and stores the token in the operating system’s secure storage when available. The auth profiles command lists saved profiles. Keep credentials out of agent prompts; the CLI uses authentication configured in the environment where it runs. For CI/CD, Databricks recommends OAuth client credentials or token-based authentication instead of browser login. See the CLI authentication instructions for setup details.

Validate and plan a bundle

Run these commands from a repository that already contains a bundle configuration, typically databricks.yml. The dev target below is an example: use a target defined in that file.

databricks bundle validate -t dev
databricks bundle plan -t dev --output json > plan.json

Validation checks the bundle configuration and reports problems. It does not deploy resources. Planning compares the resources listed in the bundle with those in the workspace, then writes proposed changes as JSON to plan.json. The plan command contacts the workspace to inspect its current state, so the active CLI profile must have access to that workspace. Databricks documents these commands and plan options in its CLI command reference.

The agent can summarize which resources the bundle proposes to create, update, or delete. Review the JSON before approving deployment. The plan does not apply the proposed changes.

To apply an approved plan, run the separate write command:

databricks bundle deploy -t dev --plan=plan.json

The CLI sends changes to the workspace during deployment. The identity used by the CLI must have the permissions those changes require. If a person needs to approve proposed changes first, assign planning and deployment to separate agent tasks.

Give the agent Databricks instructions

The CLI’s aitools command group can install and manage Databricks skills for supported coding agents. Skills are instruction files that give an agent Databricks-specific guidance, such as patterns for working with bundles, jobs, and SQL. See the Databricks agent skills documentation for supported agents and installation details.

Skills give the agent task guidance, but they do not configure a Databricks Model Context Protocol (MCP) server. An MCP server exposes tools that the agent can call, and you must configure it separately. If the agent only needs to run CLI commands in a repository, a skill is optional. It can still give the agent useful Databricks conventions.

Example agent request

In a repository with a configured dev target, ask the agent:

Inspect the existing Databricks bundle. Run validation and generate a JSON plan for the dev target. Summarize the proposed resource changes and any validation errors. Do not deploy or modify workspace resources.

This request tells the agent what to inspect and review, and makes clear that it does not have approval to make changes. The agent still needs an installed CLI and an authenticated profile that can access the workspace.

For more command groups and options, use databricks -h or consult the CLI command reference. Developers exploring other shell tools for coding agents can browse the directory.