Use Workbench AI apps

Prior reading: Overview of cloud app types

Purpose: This document provides an overview of creating and using Workbench AI cloud apps.

Introduction

Workbench gives you the option to create a Workbench AI cloud app.

This app type is a custom JupyterLab Workbench cloud app that uses generative AI to help researchers query data, build dashboards, and manage workflows. It runs on a Model Context Protocol (MCP) server that lets Claude Code perform various actions inside a Workbench workspace, including listing resources, querying BigQuery (on GCP-backed workspaces), exploring Data Explorer underlays, and executing wb CLI commands.

Create your Workbench AI cloud app

• Using the UI
• Using the CLI

Set up your cloud app

Launch your app in JupyterLab by selecting the name of your Workbench AI app. Note that the app must be running.

Run Claude

Open a Terminal in your app and run claude to start a Claude Code session.

The first time you run Claude, you'll be prompted to select a login method. You can log in via your Claude account that has a paid subscription or via the Anthropic Console account.

Explore your workspace and data

Once you're logged in, you can ask Claude questions about your workspace's resources and data.

Example prompts include:

• "What data collections are available in my workspace?"
• "Show me the schema for my primary BigQuery dataset."

This example shows how a Workbench AI app can list resources:

You can also summarize an AWS Aurora resource:

Build dashboards and custom Workbench apps

Claude can also generate a complete, deployable Flask or Streamlit dashboard from a description or a BigQuery table. It can handle proxy URL configuration, relative path requirements, and server settings automatically. It provides an app URL when done.

Example prompts include:

• "Build a dashboard that visualizes patient cohort counts from <my-table>."
• "Create a Streamlit app to explore and filter my genomics dataset."
• "Build a REST API that exposes my BigQuery data."

This video shows how Claude can visualize GTEx sample data as a dashboard:

Troubleshoot failed workflows

Claude can diagnose failed WDL or Nextflow jobs by pulling logs, identifying the failed task, checking resource allocation, and recommending a specific fix — with no manual log hunting required.

Example prompts include:

• "My workflow job failed, can you help me figure out why?"
• "Troubleshoot the latest failed job in my workspace"
• "Fix my workflow — it's running out of memory."

Write analysis code

Claude Skills are included for common research domains, allowing Claude to write production-ready analysis code with the necessary libraries and patterns. The skills are taken from Scientific Agent Skills.

Domains and capabilities include:

• Bioinformatics: Single-cell analysis (scanpy/anndata), differential expression (PyDESeq2), RNA velocity, sequence analysis
• Drug Discovery: Cheminformatics (RDKit), molecular ML (DeepChem), bioactivity databases (ChEMBL), target-disease associations (Open Targets)
• Genomics: Gene annotations (Ensembl), protein data (UniProt), variant interpretation (ClinVar), 3D structures (PDB)
• Data Analysis: ML pipelines (scikit-learn), statistical tests (statsmodels), interactive visualization (Plotly, Seaborn)
• Clinical: Clinical trial search (ClinicalTrials.gov), literature search (PubMed), survival analysis (lifelines)