Skip to content

0-Vibe: Build your first application

This guide walks you through building an App Framework recipe from scratch, from an empty repository to a deployed DataRobot application. It covers three progressively richer examples: a simple FastAPI application, an LLM with a notebook, and a full agentic workflow.

Prerequisites

This guide assumes that you are creating a new recipe repository and applying App Framework components into it. A recipe is the repository that contains your application code, infrastructure, and component answers files.

Creating a new application from scratch

Every App Framework recipe starts the same way, regardless of what you're building.

Step 1: Create your repository

  1. Head to the datarobot-oss GitHub org and create a new repository.
  2. Name it with a recipe- prefix if you're making a reusable template.
  3. Start from oss-template-repo. It includes the standard project scaffolding.

Important Git settings:

  • Allow merge commits.
  • Set merge commits as the default merge strategy.

Step 2: Clone and prepare

git clone git@github.com:datarobot-oss/recipe-your-app-name.git
cd recipe-your-app-name

Step 3: Install uv

curl -LsSf https://astral.sh/uv/install.sh | sh

Step 4: Bootstrap your recipe

Run the base component copier and answer the interactive questions:

uvx copier copy https://github.com/datarobot-community/af-component-base .

Copier prompts for your recipe name, the components you want, and other configuration. Answer thoughtfully because those choices shape the structure of your recipe. You now have the foundation in place, and everything from here is customization.

Tip

Gave a wrong answer? You can change your answers after the fact without starting over. See Working with Copier for how to re-run prompts, override a single answer, or recover from a mistake.

After this step, expect your repository to contain the shared project files that other components build on, including .datarobot/, a Taskfile.yaml, and Pulumi configuration.

Example 1: Simple FastAPI application

Overview

A two-component App Framework application that covers a wide range of common functionality: a FastAPI backend served as a DataRobot Custom Application, with a React frontend included.

Key components

Implementation

This example assumes that you already completed Steps 1 through 4 and that the base component is present in your repository.

After bootstrapping with af-component-base, add the FastAPI backend:

uvx copier copy https://github.com/datarobot-community/af-component-fastapi-backend .

Accept the defaults. Then compose the task definitions for the generated project:

dr task compose .

Configure your DataRobot credentials:

dr auth set-url
dr dotenv setup

Develop locally

Start the development server:

dr run dev

Visit http://localhost:8080. Changes are auto-discovered and the server restarts on every file change.

The FastAPI autodocs are available at /docs and /redoc:

FastAPI Swagger and ReDoc

Deploy to DataRobot

When you're happy locally, deploy:

dr run deploy

If your generated project exposes deployment commands through dr task, use the project-provided deployment command shown by dr task --list. In this guide, the examples use the command produced by the generated Taskfile for each scenario.

The CLI previews the resources to be created and asks for confirmation:

Pulumi deploy preview

After deployment, your application is live at the URL shown in the output. Open it from your terminal with cmd-click or ctrl-shift-click:

Deployed app on DataRobot

With these steps, you get unit tests, linters, deployments, fast local iteration, and everything needed for team-driven development through GitHub Actions.

Lifecycle commands:

dr task infra:down   # Tear down to save costs.
dr run deploy        # Bring it back.
dr auth set-url      # Switch to a different DR environment.

Example 2: LLM with notebook

Notebook example overview

For fast iteration on LLM use cases that don't yet need a full agentic workflow, use the LLM component and iterate in a Python notebook — locally or in DataRobot.

Notebook example components

Notebook example implementation

Apply the LLM component to your recipe:

uvx copier copy git@github.com:datarobot-community/af-component-llm .

Accept the defaults, then set up your environment:

dr dotenv setup

The setup wizard asks for your LLM integration type. For most use cases, select LLM Gateway with External Model — this is the most production-ready option:

LLM configuration interactive setup

Deploy the LLM

dr task deploy

Deploy preview showing LLM resources

Confirm and deploy. When complete, note the deployment ID because you need it for notebook development. You can retrieve it again at any time with:

dr task infra:info

You can test your LLM in the DataRobot workbench under your use case → LLM Playground.

Notebook development

Create a notebooks directory and start Jupyter:

mkdir notebooks
cd notebooks
uv init .
uv add jupyter dotenv litellm
uv run jupyter notebook

Codespaces

If running in GitHub Codespaces, skip the uv commands. Run pip install dotenv litellm and use the "Create Notebook" button instead.

Create a new notebook named llm_example. Add these cells:

from dotenv import load_dotenv
from litellm import completion
from os import getenv
from urllib.parse import urljoin

load_dotenv()

DATAROBOT_API_TOKEN = getenv('DATAROBOT_API_TOKEN')
DATAROBOT_ENDPOINT = getenv('DATAROBOT_ENDPOINT')

# Adjust these based on `dr task infra:info`.
LLM_DEPLOYMENT_ID = 'DEPLOYMENT_ID'
LLM_DEFAULT_MODEL = 'datarobot/azure/gpt-5-mini-2025-08-07'
DEPLOYMENT_BASE = urljoin(DATAROBOT_ENDPOINT, f'v2/deployments/{LLM_DEPLOYMENT_ID}/chat/completions')

def ask_llm(*messages):
    return completion(
        base_url=DEPLOYMENT_BASE,
        api_key=DATAROBOT_API_TOKEN,
        model=LLM_DEFAULT_MODEL,
        messages=messages
    )

response = ask_llm({"content": "Hi", "role": "user"})
response.choices[0].message.content

Notebook running locally

Deploy the notebook to DataRobot

Create infra/infra/notebook.py:

from pathlib import Path
from pulumi_datarobot.notebook import Notebook

from . import use_case

PROJECT_ROOT = Path(__file__).resolve().parents[2].absolute()
NOTEBOOK_PATH = PROJECT_ROOT / "notebooks" / "llm_example.ipynb"

notebook = Notebook("llm_example_notebook",
    file_path=str(NOTEBOOK_PATH), use_case_id=use_case.id)

Then run dr task deploy again. The notebook appears in your use case in the DataRobot workbench.

Tip

In the DataRobot-hosted notebook, install dependencies via the terminal: pip install litellm dotenv

Example 3: LLM with agent

Agent example overview

For use cases that need agents, tools, and multi-step workflows, the agent component provides a full agentic framework built on top of the LLM component from Example 2.

Agent example components

Agent example implementation

Step 1 — Apply the base component

uvx copier copy https://github.com/datarobot-community/af-component-base .

Accept the defaults.

Step 2 — Apply the LLM component

uvx copier copy https://github.com/datarobot-community/af-component-llm .

Accept the defaults.

Step 3 — Apply the agent component

Use the CLI:

dr component add agent

When prompted, choose your agent framework: CrewAI, LangGraph, LlamaIndex, or the YAML-based NeMo Agent Toolkit.

The component creates the following structure:

agent/
├── agent/myagent.py   # Agent workflow
├── cli.py             # Command-line interface
├── dev.py             # Local development server
├── tests/             # Test suite
└── public/            # UI assets
infra/infra/agent.py   # Pulumi configuration

Step 4 — Configure your environment

dr dotenv setup

The wizard asks for:

  1. Agent port (default: 8842).
  2. DataRobot execution environment.
  3. Execution environment version ID.
  4. Pulumi passphrase.
  5. DataRobot default use case (can leave blank).
  6. LLM Gateway configuration.

Agent environment setup

Develop and test locally

Start the server:

# Terminal 1
dr run dev

In another terminal, run the agent:

# Terminal 2
uv run python cli.py execute --user_prompt "Write a blog post about AI in healthcare" --show_output

Deploy the agent

dr task deploy

Agent deploy preview

The CLI asks you to name your stack, previews the resources (LLM deployment and agent deployment), and deploys both to DataRobot.

After deployment, find your agent in the DataRobot workbench under your use case. Interact with it through the agent playground or via API calls to the deployment endpoint:

Agent interaction in DataRobot

Customize the agent

Edit agent/agent/myagent.py to:

  • Change agent roles and goals.
  • Modify task descriptions.
  • Add more agents to the crew.
  • Integrate additional MCP tools.

Discovering available components

CI/CD setup

Once your recipe is running, set up CI/CD for automated testing and deployment: