01 Clone And Run

Track A Step 1: Clone and Run

Get a working agent running locally in under 30 minutes using the official Databricks app-templates.

When to Use

• Starting a new GenAI agent project on Databricks from scratch.
• Want a working agent with built-in chat UI before writing any custom logic.
• Following the Author an AI agent and deploy it on Databricks Apps workflow.

Foundation Prerequisite Check

Foundation Steps F1 and F2 must be complete before starting Track A. Run these checks to verify:

# F1 check: MLflow installed and autolog available
python3 -c "import mlflow; assert tuple(int(x) for x in mlflow.__version__.split('.')[:2]) >= (3,10); mlflow.openai.autolog(); print('F1: OK')" 2>/dev/null || echo "F1: FAIL"

# F2 check: MLflow experiment exists
# EXPERIMENT_PATH MUST be set to the user-and-use-case-pinned path
#   /Users/<user_email>/mlflow/<APP_NAME>-agent
# (read it from .vibecoding-state.md → mlflow_experiment_path).
# The check fails fast if EXPERIMENT_PATH is unset rather than papering over
# it with a generic /Shared/my-agent/traces default.
python3 -c "import mlflow, os; p = os.environ.get('EXPERIMENT_PATH'); assert p, 'EXPERIMENT_PATH not set — read mlflow_experiment_path from .vibecoding-state.md'; e = mlflow.get_experiment_by_name(p); assert e, f'No experiment at {p}'; print('F2: OK')" 2>/dev/null || echo "F2: FAIL"

If F1 fails: Stop. Load and execute foundation/01-mlflow-genai-foundation/SKILL.md to set up the MLflow environment, autolog, and environment detection.

If F2 fails: Stop. Load and execute foundation/02-experiment-tracing-and-uc-storage/SKILL.md to create the experiment, configure tracing, and set up UC OTEL storage.

Both must print OK before continuing. The template you clone below already contains MLflow setup code — F1/F2 teach the concepts and provision UC resources, and the AI agent will reconcile rather than duplicate.

Prerequisites

Verify before starting:

databricks auth token > /dev/null 2>&1 && echo "CLI: OK" || echo "CLI: FAIL"
uv --version > /dev/null 2>&1 && echo "uv: OK" || echo "uv: MISSING"
python3 -c "import sys; assert sys.version_info >= (3,11)" && echo "Python: OK"
node --version  # Should be v20+

All four must pass. If databricks auth token fails: IDE/CLI — set up a profile per PRE-REQUISITES §11; Genie Code — not applicable (pre-authenticated; uv/node checks also differ — see skills/genie-code-environment).

Step 1a: Pick a Template and Clone

Databricks publishes multiple agent app templates. Pick the one that matches your agent's required capabilities — they share the same @invoke / @stream authoring pattern (Option B from A2: Agent Framework), so picking a more advanced template is not harder; it just adds plumbing you'd write yourself otherwise.

Default for the canonical SkyLoyalty walkthrough: agent-openai-advanced. It already wires the two-layer memory pattern from A5: Lakebase Memory and saves you the work of porting from agent-openai-agents-sdk.

Clone the template

# Option A: Use databricks CLI (when supported)
databricks apps init --template agent-openai-advanced my-agent-app

# Option B: Direct git clone (works for any template)
git clone --depth 1 https://github.com/databricks/app-templates.git /tmp/app-templates
cp -r /tmp/app-templates/agent-openai-advanced ./my-agent-app
cd my-agent-app
rm -rf .git
git init

Replace agent-openai-advanced with whichever template you picked from the table above, and my-agent-app with your project name. Use lowercase-with-hyphens (this becomes the Databricks App name).

PyPI index URL with fallback

Some workspaces sit behind a corporate proxy or firewall that blocks the public PyPI index, and others have intermittent failures resolving pypi.org. To make the clone-and-run path resilient, configure uv / pip with a public-first, Databricks-mirror-second index list before running quickstart:

# state://Track_A.a1_clone_and_run
pypi_index_url_with_fallback:
  - https://pypi.org/simple
  - https://pypi.databricks.com/simple

Apply the same list to the local environment (the order matters — pypi.org is primary, the Databricks mirror is the fallback):

# uv (preferred): primary + fallback indexes
export UV_INDEX_URL="https://pypi.org/simple"
export UV_EXTRA_INDEX_URL="https://pypi.databricks.com/simple"

# pip equivalent (if a sub-step shells out to pip)
export PIP_INDEX_URL="https://pypi.org/simple"
export PIP_EXTRA_INDEX_URL="https://pypi.databricks.com/simple"

If uv run quickstart fails with an index resolution error (Could not connect to https://pypi.org, 403, ReadTimeout from the index), retry once with the fallback as primary:

export UV_INDEX_URL="https://pypi.databricks.com/simple"
export UV_EXTRA_INDEX_URL="https://pypi.org/simple"
uv run quickstart

Capture the index list under state://Track_A.a1_clone_and_run.pypi_index_url_with_fallback exactly as listed above. The list MUST contain both URLs in this order (public first, mirror second) so re-runs and CI redeploys can recover from a transient public-PyPI outage without code changes.

Step 1b: Run Quickstart

The template includes a quickstart script that installs dependencies and configures the local environment:

cd my-agent-app
uv run quickstart

What quickstart does:

1. Creates a virtual environment with uv
2. Installs all Python dependencies from pyproject.toml
3. Creates a .env file from .env.example (or interactive prompts)
4. Validates Databricks CLI authentication
5. Sets up the MLflow tracking URI

If quickstart prompts for values:

After quickstart, verify the .env file was created:

cat .env
# Should contain DATABRICKS_HOST, WAREHOUSE_ID, LLM_MODEL, etc.

Step 1c: Start the Local Dev Server

Genie Code: there is no local dev server — this local-quickstart path is IDE-only. Instead deploy to the Databricks Apps runtime and test the deployed app (the platform builds it server-side). See skills/genie-code-environment §4.

uv run start-app

This starts the MLflow AgentServer locally, typically on http://localhost:8000.

Expected output:

INFO:     Started server process
INFO:     Uvicorn running on http://0.0.0.0:8000

Verify the built-in chat UI

Open http://localhost:8000 in your browser. You should see:

• A chat interface with a text input field
• Streaming responses with markdown rendering
• The ability to send messages and receive agent responses

Verify programmatic access

In a separate terminal:

curl -X POST http://localhost:8000/invocations \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "Hello, what can you do?"}]}'

You should receive a JSON response with the agent's reply.

Step 1d: Tour the Project Structure

Read references/project-structure.md to understand the generated project layout.

DO / DON'T

DO — Run quickstart before start-app

uv run quickstart    # Sets up .env, installs deps
uv run start-app     # Starts the server

DON'T — Skip quickstart

uv run start-app     # Error: missing .env, no dependencies

DO — Check that the agent responds before moving on

curl -s http://localhost:8000/invocations \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "Hello"}]}' | python -m json.tool

DON'T — Move to Foundation Step 1 without verifying the chat UI works

If the agent does not respond, use the checks in references/troubleshooting.md.

Troubleshooting

Read references/troubleshooting.md if uv quickstart or uv start-app fails.

Validation Gate

All must pass before proceeding:

• Project cloned and uv run quickstart completed without errors
• uv run start-app starts the server on http://localhost:8000
• Chat UI at http://localhost:8000 renders and accepts input
• Agent responds to "Hello" with a coherent reply
• You can identify agent.py, start_server.py, app.yaml, and databricks.yml
• pypi_index_url_with_fallback captured in state with https://pypi.org/simple first and https://pypi.databricks.com/simple second

Next Step

After passing this gate, load and execute tracks/A-custom-agent-apps/02-agent-framework/SKILL.md (A2: Agent Framework).

Notes to Carry Forward

Record these values for subsequent steps:

Related Skills

References

• Author an AI agent and deploy it on Databricks Apps
• Databricks app-templates
• MLflow AgentServer
• uv package manager

Version History