One platform.
Every tool you already use.

Managed data plane in your AWS account, or self-managed on EKS.

Managed data plane on Azure, or self-managed on AKS.

GPU-dense deployments on CoreWeave's Kubernetes cloud.

Run on Crusoe's energy-efficient GPU cloud.

Managed data plane on GCP, or self-managed on GKE.

Deploy on Nebius AI cloud infrastructure.

Self-managed deployment on any conformant Kubernetes cluster.

Deploy on Oracle Cloud Infrastructure.

Native S3 support for workflow inputs, outputs, and artifacts.

Native Azure Blob support for workflow data.

Native GCS support for workflow data.

Give Claude-powered agents durable execution, retries, and replay across restarts.

Run CrewAI crews with durable tool calls and automatic recovery.

LangChain Deep Agents with checkpointed planning and sub-agent steps.

Run Google Agent Development Kit agents on Union.ai's durable runtime.

Union.ai's native agent framework, durable by design.

Run LangChain chains and agents with durable, resumable execution.

Persist LangGraph state graphs across failures without changing your graph code.

Serve Model Context Protocol servers as managed, multi-tenant Union.ai apps.

Durable agents on the Mistral Conversations API.

Run OpenAI Agents SDK agents as durable workflows with checkpointed tool calls.

Type-safe Pydantic AI agents running on a durable backend.

Run Ollama model servers as Union.ai apps.

High-throughput LLM serving with SGLang, deployed as a managed app.

Serve open-weight LLMs with vLLM as an autoscaling Union.ai app.

Deploy long-running connectors to external systems as managed apps.

Serve REST APIs as autoscaling Union.ai apps with native FastAPI support.

Deploy Streamlit dashboards alongside the pipelines that feed them.

Trigger workflows from external events with managed webhook endpoints.

Run BigQuery SQL as workflow tasks with typed inputs and outputs.

On-demand Dask clusters for distributed Python, managed per workflow.

Submit Spark jobs to Databricks from your workflows.

Move HF Datasets between tasks and reference Hub datasets directly.

Fast JSONL file and directory types with optional compression and Arrow.

Lance columnar format for DataFrames, streamed lazily between tasks.

Schema-validated DataFrames (pandas, Polars, or PySpark) as task types.

Pass Polars DataFrames and LazyFrames between tasks, backed by Parquet.

Parameterized Snowflake queries, key-pair auth, and DataFrame results.

Ephemeral Spark clusters on Kubernetes, sized per task and torn down after.

Distributed and elastic PyTorch training with automatic fault recovery.

Ephemeral Ray clusters on Kubernetes for distributed training and tuning.

Generations, tool calls, tokens, and cost for every agent run, in Grafana.

Log MLflow experiments and metrics from inside your tasks.

Profile any GPU task under Nsight Systems with a single decorator.

Distributed traces that stay intact across crashes and durable resumes.

Experiment tracking for training runs, including distributed PyTorch.

Compose Hydra configs and launch sweeps as workflow runs.

Use DictConfig and ListConfig as typed task inputs and outputs.

Generate Python from natural language and validate it with sandboxed tests.

Execute untrusted or generated code in isolated sandboxes.

Run parameterized Jupyter notebooks as workflow tasks.