A hands-on workshop using AI coding agents and Google Cloud data tools to explore a coffee shop expansion scenario.
Prepare your computer before starting. If you are attending a workshop or Skills/Qwiklabs lab, use the Google Cloud project and access instructions provided by your organizer; you do not need a personal billing account or free trial. If you are exploring independently, prepare your own project.
Install these prerequisites using their official instructions:
- Git
- Google Cloud CLI
- Node.js 22.6 or newer and npm (current LTS recommended)
- One supported agent: Codex CLI, Claude Code CLI, or Antigravity CLI
Clone the repository:
git clone https://github.com/jeffonelson/data-cloud-creator-workshop.git
cd data-cloud-creator-workshopThen check local prerequisites:
./bin/doctor --preflightThis checks command availability and the minimum Node.js version. It makes no cloud requests, starts no login flow, and installs or configures nothing. No Google sign-in, project, or credit card is required. Fix any reported failures using the installation links and rerun until there are zero failures.
A passing preflight confirms local prerequisites only. Agent sign-in, Google authentication, project permissions, plugins, and live tools are checked later.
Replace YOUR_USERNAME with your Google account email address and
YOUR_PROJECT_ID with the project you will use. If you have an assigned lab
project, use the credentials supplied by the lab.
gcloud auth login YOUR_USERNAME --project=YOUR_PROJECT_ID --no-launch-browser
gcloud auth application-default login YOUR_USERNAME --project=YOUR_PROJECT_ID --no-launch-browserWhen sign-in is needed, --no-launch-browser prints an authorization URL instead
of opening your default browser; paste it into the browser profile for that
account (or an Incognito/private window for temporary lab credentials), then
follow the terminal prompts.
Sign in to your chosen agent CLI if needed: codex login, claude auth login,
or agy. Agent sign-in is separate from Google Cloud sign-in. Workshop resources
are prepared by the organizers; skip provisioning for your assigned lab project.
./bin/setup checks access to your project and secrets, then installs
and configures the newest Data Agent Kit release for each supported agent on
your computer. It uses upstream DAK setup to select BigQuery, AlloyDB and Cloud Storage, with
notebook and visualization tools included by upstream. The plugin is
dak@dak-marketplace; the data servers are dak-bigquery and dak-alloydb.
Restart open agent sessions after setup.
Use your project ID and its AlloyDB region. For an assigned lab, use the values supplied by the lab. Setup saves them for subsequent launches. BigQuery uses its own dataset location; Cloud Run follows the selected region.
The selected project is saved in ignored .workshop-state/project.env and upstream's
non-secret ~/.config/data-agent-kit/config.json. Setup does not change the global
gcloud project. A conflicting pre-existing DAK config or plugin installation stops
setup; resolve it with upstream DAK or the agent's plugin commands, then retry.
To move an installation owned by this checkout to a later release, run
./bin/teardown and then ./bin/setup PROJECT_ID --region REGION again. Pre-existing installations are preserved.
./bin/setup YOUR_PROJECT_ID --region YOUR_REGION./bin/doctor runs read-only checks for the required software, agent CLI login,
Google Cloud authentication, plugin installation, and MCP configuration.
./bin/doctorExisting checkouts must rerun setup with --region once. Doctor verifies cluster access in that region and checks that the agents' settings agree.
Use a workshop launcher instead of running codex, claude, or agy directly so it
can load the workshop credentials and project settings before starting the agent.
./bin/workshop-codex./bin/workshop-claude./bin/workshop-antigravityTrust the repository when prompted. Keep this session open for the entire lab.
To explore independently or continue after a lab expires, use an existing
billing-enabled project you own. Install Python 3 and Google Cloud CLI with bq,
then replace YOUR_USERNAME with your Google account email address and
YOUR_PROJECT_ID with your project ID:
gcloud auth login YOUR_USERNAME --project=YOUR_PROJECT_ID --no-launch-browser
gcloud auth application-default login YOUR_USERNAME --project=YOUR_PROJECT_ID --no-launch-browser
./bin/provision YOUR_PROJECT_ID
./bin/setup YOUR_PROJECT_ID --region us-central1
./bin/doctorThis checkout provisions the original database/CSV scenario and MCP credentials.
Lease PDFs are prepared by organizers and are not bundled or uploaded here.
To recreate the full workshop, including the Cloud Storage leases, use the
upstream recreation instructions.
Until the leases are added, the lease check in ./bin/doctor and the lease exercise
will not pass; the database exercises remain available.
Allow about 15–20 minutes, then launch an agent as above. Provisioning creates
billable resources, including AlloyDB, which keeps charging until removed.
./bin/teardown removes local setup only.
First, run /mcp in the agent and confirm that the Developer Knowledge, Maps
Grounding Lite, Cloud Run, BigQuery, AlloyDB, and Cloud Storage MCP servers are loaded.
Cloud Storage comes from the Data Agent Kit and uses your Google Cloud credentials.
Then paste the following prompt into the same agent session and run it:
Use the MCP tools for these checks. Make one successful read-only call to each service:
1. Developer Knowledge: retrieve the official BigQuery row-level security documentation.
2. Maps: find two coffee shops near Golden Gate Park and include their Google Maps links.
3. BigQuery: list datasets in the workshop project.
4. AlloyDB: list clusters in the configured workshop region.
5. Cloud Storage: list the objects under leases/ in the workshop project's
PROJECT_ID-charlies-documents bucket. Confirm there are 12 PDFs.
Use the project selected during setup, replacing PROJECT_ID with that ID.
Return a five-row PASS/FAIL table. Empty BigQuery or AlloyDB results count as
PASS; missing lease PDFs need organizer attention.
Open the interactive data landscape to explore the tables, documents, and relationships between them.
The broker's documents are available in
gs://PROJECT_ID-charlies-documents/leases/, one PDF per candidate_sites.site_id.
Use your setup-selected project ID. The documents are available throughout the lab.
Developer Knowledge provides official Google Cloud documentation; Maps Grounding
Lite supplies current place and route context with Google Maps attribution.
🎯 Your north star
Recommend Charlie’s next San Francisco location and be prepared to defend it. The prompts below are suggestions, not a required sequence. Ask your own questions and follow whichever evidence interests you.
Charlie's Coffee operates six San Francisco stores and is choosing where to open next. Run the suggested prompts for the shortest path through the scenario, then choose any optional follow-ups that match your interests and available time. Continue in the same agent session so later prompts can build on earlier findings.
The suggested path uses Developer Knowledge → business data → a shortlist → Google Maps → broker documents. Record your recommendation before the facilitator calls the room together. These are fictional business records; Maps provides real-world context around the scenario's representative locations.
Use the Developer Knowledge MCP to explain why Charlie's might keep live
inventory in AlloyDB, sales and market history in BigQuery, and broker PDFs in
Cloud Storage. Cite official Google Cloud documentation. How can an agent
bring these sources together to help choose a location?
How can we analyze AlloyDB and BigQuery together without first building another ETL pipeline? Use the official Google Cloud documentation and cite the documents you use.
Charlie's wants to open a seventh San Francisco store. Help me make a recommendation we can defend. Before running any analysis, what evidence would you want, and where would you expect to find it?
Where is Charlie's inventory failing to meet customer demand?
Can stock-movement history explain how those shortages developed?
Use BigQuery AI to extract availability complaints from customer reviews.
Do they corroborate the live inventory?
Can you combine the operational and analytical evidence in one analysis without copying the AlloyDB data into BigQuery first?
Which complaints match specific store/product pairs, and which are ambiguous?
Use BigQuery AI.FORECAST to forecast six months of candidate-market demand.
Recommend one site and one runner-up after estimating store revenue and
accounting for asking rent. Explain how market demand becomes store revenue.
Label the result as a comparison, not profit: other operating costs are missing.
What separates Charlie's strongest store markets from its weakest ones?
How did you translate market demand into store revenue? Check your assumptions
against Charlie's existing stores.
What assumptions have the most influence on the ranking, and what would need to change for the runner-up to win?
Write down this first recommendation and its main assumption. Then test it with Maps and the broker documents below.
Use Maps Grounding Lite to compare the surroundings of my recommended SF site
and runner-up. What nearby places strengthen or weaken each case? Include
every returned Google Maps source link immediately after the claim it supports
and preserve its attribution. Distinguish observed places from assumptions
about their effect on sales.
Does nearby coffee competition indicate saturation, or does it validate demand?
Which nearby residential areas could expand Charlie’s delivery reach through DoorDash or Uber Eats? Use Maps to compare driving times from the site.
Which nearby residential buildings and neighborhood anchors are within a 10-minute walk of the site and could contribute launch-day foot traffic?
Before settling on your recommendation, give both finalists a document check. The broker has supplied one draft per candidate site in Cloud Storage. If you already explored them, use this step to audit your comparison.
Use Cloud Storage MCP to read the broker PDFs for my recommended site and
runner-up in gs://PROJECT_ID-charlies-documents/leases/. Use my setup-selected
project ID and match the documents to candidate_sites by site_id.
Check our rent assumptions against the proposed terms. Compare monthly base
rent, other recurring property charges, and the length of the commitment.
Cite the document, page and section for each finding. Separate stated amounts,
estimates and unknowns. Does this change your recommendation or confidence?
Audit the revised comparison. Are the amounts monthly or annual? Did we count
anything twice? Keep refundable deposits separate from recurring expenses.
What would we need the broker to confirm before making a commitment?
If PDF reading fails, tell the facilitator and run ./bin/doctor; a successful
listing alone does not prove the agent can read a PDF. Do not invent terms or
treat missing costs as zero.
⏸️ Facilitator checkpoint
Record your current choice, runner-up, one reason supported by a source, and one assumption that could change your mind. Continue exploring until the facilitator calls the room back together.
The remaining activities are for extra workshop time or continued exploration afterward.
Want to explore somewhere you know? Bring Charlie's to your own spot. Replace
[LOCATION] with anywhere in the world you'd like to explore; include the country
if the name could be ambiguous. San Francisco's numbers remain business benchmarks.
Skip this section to continue with your San Francisco recommendation.
What if Charlie's expanded into [LOCATION]? Use Maps to suggest three areas
worth exploring, informed by what we learned from its San Francisco stores.
Which area looks most promising, and what would we need to verify?
Pick a specific map anchor for a hypothetical store to explore next.
Design a launch-week plan for the recommended site using everything we have learned.
If we held an outdoor launch event at the recommended site this weekend, what do the hourly and daily weather forecasts suggest, and how should we adjust the plan?
Which nearby gyms and coworking spaces should we approach as launch partners?
What stock problems should we avoid repeating at launch?
Revise the launch plan based on what you discovered.
Build an interactive BI dashboard that helps our team explore Charlie's sales,
inventory, and market demand. Highlight the most interesting findings.
Audit the dashboard's numbers and labels against the source data.
Deploy the dashboard to Cloud Run so our team can revisit it. Return the live URL.
Exit the agent, then run:
./bin/teardownThis removes workshop-installed agent plugins and local workshop state. It does not change Google Cloud or delete this repository.
This is a project intended for demonstration and workshop purposes only. It is not intended for use in a production environment. This repository is not an officially supported Google product.