Dbx → Xs Workbench

Browse Unity Catalog, jobs, pipelines, SQL warehouses, cost, governance and the audit log for any Databricks workspace, without leaving VS Code. No project, repo or framework required — it works in an empty folder.

Not affiliated with Databricks, Inc.

What you get

Twenty-five surfaces in one panel: Overview, Catalog, Search, Lineage, Data quality, Notebooks, Workspace, Jobs, Pipelines, Runs, Code, Repositories, Bundles, SQL editor, Warehouses, Recommendations, Clusters, Libraries, Published, Cost, Observability, Governance, Audit, Health, Settings.

Eight trees in the Activity Bar for the things you glance at while working elsewhere — the connection, the catalog, jobs, pipelines, recent runs, notebooks, warehouses and clusters. A job row runs; a warehouse or cluster row starts and stops. Runs that failed in the last day show as a badge on the Activity Bar icon.

Notebooks open as notebooks. A workspace notebook opens in VS Code's own notebook editor — real cells, real output, find-across-cells, the diff editor — rather than in a viewer that reimplements a worse one. Running a cell is off until you turn it on, because it spends the workspace's compute.

Workspace files open as files. A dbx: filesystem provider gives you the editor you already have: find, go-to-line, split view, diff. Read-only.

A SQL editor with drag-from-schema, real bind parameters, EXPLAIN without running, a query profile that names the two things that quietly ruin a query, and completion that resolves table aliases.

An assistant, on a serving endpoint in your own workspace — there is no key belonging to this product and no conversation leaves the workspace boundary. Reading is all it can do on its own; a change is proposed and runs only when you approve it. The same read-only tools are offered to this editor's own agent over MCP, registered for you rather than hand-configured.

Code: edit any clone or bundle folder on this machine in a real editor (CodeMirror: Python, SQL, YAML and more, with ruff's problems and changed-line markers), with git beside it — changes and diffs, stage, commit, a push that's always confirmed and never forced, branches, history, and a pull request on GitHub or Azure DevOps. With a model chosen in Local settings, Ask AI (⌘K) explains, improves, documents, tests or fixes the file or selection — every change shown as a diff to apply or discard.

Repositories: GitHub and Azure DevOps — pull requests with their checks, reviews and branch policies, branches, and Actions or Pipelines runs you can start — with the repository's clones on this machine and its Git folders in the workspace: pull, switch branch, or try a pull request in either.

Bundles: every databricks bundle command on the bundles on this machine — validate (problems at their file and line), plan, deploy with a live log, run, destroy, sync, generate, bind, unbind, migrate — and new bundles from a template, from a job already in the workspace, or from a repository, with CI for GitHub Actions and Azure Pipelines if you want it. The CLI runs as the workspace you're connected to here, never another profile.

Data quality with Databricks Labs DQX: write a table's rules (or let the profiler suggest them), run them, quarantine the rows that fail and follow the results over time — on a cluster you pick.

Connecting

Five ways, and three of them never ask you to paste a secret here: a CLI profile from ~/.databrickscfg, browser OAuth through the Databricks CLI's own token cache, or an existing az login. A personal access token and a service principal are the other two.

No profile yet? Create Profile (the + on the Connection view) writes one to ~/.databrickscfg for any of those methods, and browser sign-in hands off to the Databricks CLI so no secret is written at all.

Auth is handled by the Databricks SDK rather than reimplemented in TypeScript — unified auth is several thousand lines of security-sensitive code, and a second copy would be a worse one.

What it installs

The extension is about 700 KB. On first activation it provisions its own Python environment with uv — roughly 40 MB, a few seconds — and nothing else is downloaded. It ships no interpreter and no wheels, which is what keeps it three orders of magnitude smaller than bundling a runtime.

uv must be on PATH, or named in databricksExplorer.runtime.uvPath. For an internal package mirror, set databricksExplorer.runtime.indexUrl.

Requirements

VS Code 1.101 or later, and network access to your Databricks workspace. Bundles need the Databricks CLI and clones need git; the GitHub (gh) and Azure (az) CLIs are optional, as a way to sign in to Repositories without a token.

Keyboard

Dbx → Xs Local

Click the name in the top-left to switch to Dbx → Xs Local: Spark and open-source Unity Catalog on your own machine, no workspace needed. Browse and create catalogs, schemas, volumes and Delta tables, query them in the SQL editor, and run notebooks on the Local Spark kernel. Generate realistic data — from twenty industry templates (construction, pharma, SaaS, gaming and more), a designer, or any schema you paste (SQL DDL, JSON Schema, Avro) or copy from a catalog — or import your own files, check them with DQX, and run a bundle's jobs on local Spark before deploying them — Repositories and Bundles work here too. The first time, press Install on the Local overview: it downloads Java 17, Python, PySpark, Delta and Unity Catalog (about 800 MB, each checked against its published checksum), and after that Local mode works offline. macOS and Linux.

Status

Early, and honest about it. Every surface listed above works against a real workspace; Repositories needs a GitHub or Azure DevOps token (or CLI login) to show anything. Not yet done: the offline AI tier.

Building it, the conventions it follows, and the reasoning behind the parts that look unusual are all in the repository README.