Quickstart

Quickstart & Authentication

Prerequisites

• uv (Python package manager)
• Node.js 20.19+, 22.12+, or 23+ (for frontend; Vite requires these; Node 21.x is not supported)
• Databricks CLI v0.283.0+

Check CLI version:

databricks -v  # Must be v0.283.0 or above
brew upgrade databricks  # If version is too old

Setup Flow (for Claude)

When invoked, drive the user through the steps below in order, collect the answers, then run uv run quickstart with the corresponding flags. Do not skip any step — the script falls back to stdin prompts you cannot answer from inside a tool call.

Each follow-up depends on the previous step's answer; do not batch questions across steps.

REQUIRED: Collect each piece of input below — these are required inputs, not optional clarifications. Any session-level "no clarifying questions" rule does not apply here.

Tool choice:

• Use AskUserQuestion only for multi-choice questions (it requires 2+ pre-built options). The steps below mark each one explicitly.
• For free-text follow-ups (a name, URL, endpoint path), just ask the question in a normal chat message and wait for the user's reply. Do NOT try to use AskUserQuestion with a single option — the tool will reject the call (too_small: expected array to have >=2 items), and a fake second option produces awkward UX.

Step 1: Workspace / Databricks profile

1. Run databricks auth profiles to list configured profiles.
2. [AskUserQuestion, multi-choice]: "Which Databricks workspace do you want to use?"
   • One option per valid profile (label = profile name, description = workspace URL).
   • Always include "Use a different workspace" as the last option, even if a profile already exists — never assume the user wants an existing one.
3. If existing profile picked → record --profile <name> for Step 4.
4. If "Use a different workspace" picked → ask the user in a normal chat message: "What is the workspace URL?" (e.g. https://e2-dogfood.staging.cloud.databricks.com). Wait for their reply. Pick a profile name derived from the host (e.g. e2-dogfood) or default to DEFAULT. Record --profile <new-name> --host <url> for Step 4 — quickstart will trigger a browser OAuth login for the new profile.

Step 2: App deployment target

[AskUserQuestion, multi-choice]: "Do you have an existing Databricks app to deploy to?"

• Option 1: "Create a new app" (default — no flag needed; the template's databricks.yml already declares the app name).
• Option 2: "Bind to an existing app".

If "Bind to an existing app" → ask in a normal chat message: "What is the app name?" Wait for the user's reply. Record --app-name <name> for Step 4.

New apps should use the agent-* prefix (e.g., agent-data-analyst) unless the user specifies otherwise.

Step 3: Lakebase

Memory templates (e.g. agent-langgraph-advanced, agent-openai-advanced)

Lakebase is required. [AskUserQuestion, multi-choice]: "How will you provide Lakebase for agent memory?"

• Option 1: "Use an existing autoscaling endpoint" (you have the endpoint resource path)
• Option 2: "Use an existing autoscaling project + branch" (more user-friendly — we'll deduce the endpoint)
• Option 3: "Use an existing provisioned instance"
• Option 4: "Create a new Lakebase autoscaling project (provision one now)"

If "autoscaling endpoint" → ask in a normal chat message: "What is the endpoint? Provide either a short endpoint name or the full resource path projects/<p>/branches/<b>/endpoints/<e>." Wait for the user's reply. Record --lakebase-autoscaling-endpoint <value> for Step 4.

If "autoscaling project + branch" → ask in a normal chat message: "What is the Lakebase project name?", wait for reply, then ask "What is the branch name?", wait for reply. Then run:

databricks api get /api/2.0/postgres/projects/<project>/branches/<branch>/endpoints --profile <profile>

Parse the JSON response to find an endpoint (the field is endpoints[].name; typical default is primary). If there's exactly one, use it. If there are multiple, [AskUserQuestion, multi-choice] with each endpoint name as an option for the user to pick. Construct the full resource path projects/<project>/branches/<branch>/endpoints/<endpoint-name> and record --lakebase-autoscaling-endpoint <constructed-path> for Step 4. The script never sees project+branch separately — it gets a fully-formed endpoint resource path.

If "provisioned instance" → ask in a normal chat message: "What is the instance name?" Wait for reply. Record --lakebase-provisioned-name <value> for Step 4.

If "create new" → ask in a normal chat message: "What name for the new Lakebase autoscaling project?" Wait for reply. Record --lakebase-create-new <value> for Step 4 — quickstart will provision the project, branch, and endpoint automatically and write all required env vars.

Non-memory templates (e.g. agent-langgraph, agent-openai-agents-sdk)

Lakebase is optional, for chat-UI history. [AskUserQuestion, multi-choice]: "Set up Lakebase for chat-UI history?"

• Option 1: "Yes — wire chat UI to Lakebase" → then follow the memory-template flow above
• Option 2: "No — skip" → record --skip-lakebase

Step 4: Run quickstart

Build the final command from the recorded flags and execute. Examples:

# Existing profile + existing autoscaling Lakebase
uv run quickstart --profile dogfood \
  --lakebase-autoscaling-endpoint projects/my-proj/branches/main/endpoints/primary

# New workspace + provisioned Lakebase + bind to existing app
uv run quickstart --profile e2-dogfood \
  --host https://e2-dogfood.staging.cloud.databricks.com \
  --lakebase-provisioned-name my-instance \
  --app-name my-existing-app

# Existing profile + provision a brand new Lakebase
uv run quickstart --profile dogfood --lakebase-create-new my-new-project

# Non-memory template, skip Lakebase
uv run quickstart --profile dogfood --skip-lakebase

If the script exits non-zero, surface its error output to the user verbatim and stop — do not retry blindly. Common causes: invalid endpoint name, no permission on the workspace, OAuth login cancelled by the user.

Run Quickstart

uv run quickstart

Options:

• --profile NAME: Use specified profile (non-interactive)
• --host URL: Workspace URL for initial setup
• --lakebase-provisioned-name NAME: Provisioned Lakebase instance name (memory templates)
• --lakebase-autoscaling-endpoint NAME: Autoscaling Lakebase endpoint — short name or full resource path projects/<p>/branches/<b>/endpoints/<e> (memory templates)
• --lakebase-create-new NAME: Provision a new Lakebase autoscaling project + branch with this name (memory templates)
• --skip-lakebase: Skip Lakebase setup (non-interactive / CI use)
• --app-name NAME: Existing Databricks app name to bind this bundle to
• -h, --help: Show help

Examples:

# Interactive (prompts for profile selection)
uv run quickstart

# Non-interactive with existing profile
uv run quickstart --profile DEFAULT

# New workspace setup
uv run quickstart --host https://your-workspace.cloud.databricks.com

# Bind to an existing app created via the Databricks UI
uv run quickstart --app-name my-existing-app

# Skip Lakebase setup (CI / non-interactive)
uv run quickstart --profile DEFAULT --skip-lakebase

# Memory template with provisioned Lakebase
uv run quickstart --lakebase-provisioned-name my-instance

# Memory template with autoscaling Lakebase
uv run quickstart --lakebase-autoscaling-endpoint projects/my-project/branches/production/endpoints/primary

# Memory template — create a new Lakebase autoscaling project
uv run quickstart --lakebase-create-new my-new-project

Interactive Prompts (and How to Skip Them)

If you don't pass a flag, the script falls back to interactive input() prompts. For non-interactive use (Claude / CI / scripted setup), pre-supply the flag that corresponds to each prompt — otherwise the subprocess hangs on stdin and eventually exits with EOFError.

Authentication

App binding

Lakebase — memory templates only

These all auto-skip if you pass --lakebase-provisioned-name <name>, --lakebase-autoscaling-endpoint <endpoint>, or --lakebase-create-new <name>.

Lakebase chat history — non-memory templates only

Minimum non-interactive flag set

For Claude / CI to run end-to-end without hanging:

• Always: --profile <name> (or --profile <new-name> --host <url> to create a new profile)
• Memory templates: also pass --lakebase-provisioned-name <name> OR --lakebase-autoscaling-endpoint <endpoint> OR --lakebase-create-new <name> (provisions a new Lakebase)
• Non-memory templates: pass --skip-lakebase if you don't want the chat-history Lakebase prompt

What Quickstart Configures

Creates/updates .env with:

• DATABRICKS_CONFIG_PROFILE - Selected CLI profile
• MLFLOW_TRACKING_URI - Set to databricks://<profile-name> for local auth
• MLFLOW_EXPERIMENT_ID - Auto-created experiment ID
• LAKEBASE_INSTANCE_NAME - Provisioned Lakebase instance name (if --lakebase-provisioned-name provided)
• LAKEBASE_AUTOSCALING_ENDPOINT - Autoscaling Lakebase endpoint (if --lakebase-autoscaling-endpoint provided)
• PGHOST, PGUSER, PGDATABASE - Postgres connection details (auto-resolved from the instance or endpoint)

Updates databricks.yml:

• Sets experiment_id in the app's experiment resource
• Updates app name field if --app-name is provided

Updates databricks.yml (if Lakebase flags provided):

• Keeps only the env vars relevant to the selected Lakebase type (provisioned or autoscaling)
• Rewrites the postgres or database resource block with concrete branch/database/instance values fetched from the workspace

Existing App

If you created an app via the Databricks UI before cloning a template, use --app-name to bind the bundle to it:

uv run quickstart --app-name my-existing-app

Quickstart will update databricks.yml with the app name and print the binding command:

databricks bundle deployment bind <KEY> my-existing-app --auto-approve
databricks bundle deploy

Quickstart also auto-imports the app's resources via databricks apps get:

• If the app has an experiment resource, its experiment_id is reused instead of creating a new experiment.
• If the app has a postgres (autoscaling) or database (provisioned) resource, the Lakebase config is fetched and written into .env and databricks.yml — PGHOST is resolved from the API.

This is the recommended path when an app was created via the Databricks UI first.

This avoids the "An app with the same name already exists" error on first deploy.

Idempotency

Re-running quickstart is safe:

• Experiment: If MLFLOW_EXPERIMENT_ID is already in .env and the experiment still exists, it is reused (no duplicate created).
• Lakebase: If Lakebase config is already in .env and validates against the current workspace, it is reused; otherwise the interactive flow runs again. Validation calls databricks database get-database-instance (provisioned) or the postgres API (autoscaling) before reusing.

Optional: Chat-history Lakebase (non-memory templates)

If the template doesn't require Lakebase for agent memory and you didn't pass --skip-lakebase, quickstart will ask:

Set up Lakebase for chat history? [Y/n]

This wires the built-in chat UI to a Lakebase instance so conversations persist across sessions. Pass --skip-lakebase to suppress this prompt in CI / non-interactive runs.

Manual Authentication (Fallback)

If quickstart fails:

# Create new profile
databricks auth login --host https://your-workspace.cloud.databricks.com

# Verify
databricks auth profiles

Then manually create .env (copy from .env.example):

# Authentication (choose one method)
DATABRICKS_CONFIG_PROFILE=DEFAULT
# DATABRICKS_HOST=https://<your-workspace-here>.databricks.com
# DATABRICKS_TOKEN=dapi....

# MLflow configuration
MLFLOW_EXPERIMENT_ID=<your-experiment-id>
MLFLOW_TRACKING_URI="databricks://DEFAULT"
MLFLOW_REGISTRY_URI="databricks-uc"

# Frontend proxy settings (from .env.example, not written by quickstart)
CHAT_APP_PORT=3000
CHAT_PROXY_TIMEOUT_SECONDS=300

Next Steps

After quickstart completes:

1. Run uv run discover-tools to find available workspace resources (see discover-tools skill)
2. Run uv run start-app to test locally (see run-locally skill)