datarobot-genai 0.19.2

Generic helpers for GenAI

DataRobot GenAI Library

Homepage · PyPI · Support

A toolkit for building agents on DataRobot.

• Unified LLM layer (DataRobot-compatible)—you use one get_llm() entry point per integration (LangGraph, LlamaIndex, CrewAI, NAT), all backed by the same LiteLLM-based routing to the DataRobot LLM Gateway, LLM deployments, NIM, or external providers.
• Library of agentic tools and DataRobot-compatible MCP server—use drtools and drmcp to give your agent first-class capabilities to interact with the world.
• AG-UI integration—your agents expose a standard AG-UI event stream (RunAgentInput in, lifecycle + text + tool-call events out), so your UIs and the DataRobot platform render runs consistently without bespoke adapters per framework.
• Multi-agent systems out of the box—you get first-class patterns for planner/writer crews, LangGraph multi-node graphs, and LlamaIndex AgentWorkflow handoffs; wrap them with one helper and keep the same streaming contract.
• Orchestration—you build agents from universal pieces in the low-code workflow.yaml interface. Combine and reuse LLMs, tools, agents, and evaluators. The design stays compatible with and draws inspiration from NeMo Agentic Toolkit.
• Serving and evaluating with DRAgent—you run a front-end server to plug your agent into a real-world application. DRAgent supports distributed tracing, generation and evaluation endpoints, async generations, and two-way communication over WebSockets.

Use

Installation

• You need Python 3.11–3.13.
• Install the extra that matches the framework you use:

pip install "datarobot-genai[crewai]"
pip install "datarobot-genai[langgraph]"
pip install "datarobot-genai[llamaindex]"
pip install "datarobot-genai[nat]"

You can also install:

• datarobot-genai[dragent]—serve and orchestrate your agent with DRAgent.
• datarobot-genai[drtools]—use the standard library of agentic tools DataRobot provides.
• datarobot-genai[drmcpbase]—Base class to derive FastMCP servers.
• datarobot-genai[drmcputils]—shared MCP utilities consumed by drmcpbase and drtools.
• datarobot-genai[drmcp]—host a custom MCP server in DataRobot (includes drmcpbase, drtools, and template-server dependencies).
• datarobot-genai[memory]—use the Mem0-backed memory client and NAT memory provider.
• datarobot-genai[eval]—agent evaluation utilities built on the NeMo Evaluator launcher.

Credentials

You need a DataRobot account to use DataRobot-backed features. Export these environment variables:

# Set your DataRobot API token (replace the placeholder).
export DATAROBOT_API_TOKEN=YOUR_DATAROBOT_API_TOKEN
export DATAROBOT_ENDPOINT=https://app.datarobot.com/api/v2

Standalone end-to-end examples

Follow quickstart.ipynb to walk through an experience of setting a LangGraph agent with DataRobot:

• LLM Gateway
• drtools
• Prompt Management
• Conversion to DataRobot agent format
• Running the agent with an AG-UI interface.

In-depth documentation

See docs/README.md for guides on every framework and feature in datarobot-genai.

Develop

You need Python 3.11–3.13, uv, Task CLI, and pre-commit.

uv sync --all-extras --dev
pre-commit install
task test

Semantic versioning

When you change the library, bump the patch version and add an entry to CHANGELOG.md. When you introduce a backward-incompatible change, bump the minor version.

TestPyPI

Comment /build on your PR to build and publish a dev version of the package to TestPyPI.

Excluded Upstream Dependencies

Several transitive dependencies pulled in by upstream packages are not used by this library at runtime. These are explicitly excluded via [tool.uv] exclude-dependencies in pyproject.toml to reduce install size and CVE surface area.

Package	Pulled in by	Reason excluded
build	crewai	Python build system; runtime unnecessary
diskcache	ragas	Optional disk caching backend; not used
flask	nvidia-nat 1.6.0	Web framework; not used
kubernetes	crewai-tools	K8s client; not used
lancedb	crewai	Optional vector DB backend; not used
langchain-milvus	nvidia-nat-langchain	Milvus vector DB adapter; not used
leptonai	nemo-evaluator-launcher	Lepton AI cloud backend; not used, pins httpx==0.27.2 which conflicts with the auth extra
llama-index-cli	llama-index	CLI tool; not needed at runtime
openpyxl	crewai-tools	Excel parser; not used
pymilvus	langchain-milvus	Milvus client; not used
python-docx	crewai-tools	Word doc parser; not used
pytube	crewai-tools	YouTube downloader; not used
scikit-network	ragas	Graph analysis library; not used
stagehand	crewai-tools	Playwright web automation; not used
uv	crewai	Package manager bundled as a runtime dep; not needed
youtube-transcript-api	crewai-tools	YouTube transcripts; not used

Publishing

• Same-repo PRs—comment /build on your PR to publish dev builds to TestPyPI (.devN).
• Merge to main—the release flow creates tag v{version} and publishes to PyPI automatically.
• Version tags—when you push a v* tag, PyPI publish runs as well.
• Local release—optionally run task release:tag-and-push to create and push v{version} from your machine.

Links

• Repository—source and issues.
• PyPI—released packages.
• TestPyPI—dev builds.

License

Apache-2.0—see LICENSE.