What it is

dstack is an open-source Python project in the AI and machine learning infrastructure ecosystem. It is a unified control plane for GPU provisioning and orchestration that works with GPU clouds, Kubernetes, and on-prem clusters. The project supports NVIDIA, AMD, Google TPU, and Tenstorrent accelerators, and it is compatible with open-source tools and frameworks.

The concrete problem it solves is the operational split between development, training, fine-tuning, and inference workloads on heterogeneous hardware. Instead of managing each cloud, Kubernetes cluster, or on-prem server separately, teams can define fleets, dev environments, tasks, and services in one system. This lets the same workflow target cloud backends, Kubernetes clusters, Slurm, and SSH-based on-prem machines.

Key capabilities

• Provision and manage fleets across GPU clouds, Kubernetes, and on-prem servers, including SSH fleets for self-managed clusters.
• Run development environments, training tasks, batch jobs, and cluster workloads through fleet, task, and service configurations.
• Deploy model inference as secure and scalable endpoints, with replica groups, gateway replicas, and PD disaggregation support.
• Support NVIDIA, AMD, Google TPU, and Tenstorrent accelerators, plus Kubernetes volumes, Slurm backend, and NVIDIA Dynamo integration.
• Provide a server, CLI, and API workflow that requires Git and OpenSSH on Linux, macOS, or Windows through WSL 2.
• Install agent skills so AI agents such as Claude, Codex, and Cursor can create and manage fleets and submit workloads.
• Offer experimental presets for agent-driven inference optimization.

Who uses it and how

• ML platform teams use dstack to provision GPU compute across clouds, Kubernetes, and on-prem clusters without binding workloads to one accelerator vendor.
• Researchers and engineers use dev environments, tasks, and services to develop, train, fine-tune, and serve models, including LLM workloads.
• On-prem operators launch the server and create SSH fleets, avoiding backend configuration for self-managed GPU servers.
• Cloud and Kubernetes operators configure backends, manage multiple Kubernetes clusters, and attach Kubernetes volumes to workloads.
• Agent-driven teams use dstack skills so coding agents can edit configuration and submit fleets, tasks, and services.

Getting started

Install the server with uv tool install "dstack[all]" -U, run dstack server, then install the CLI with uv tool install dstack -U and point it to the server address, user token, and project name.

When to use it — and when not to

Use dstack when vendor-agnostic GPU orchestration across clouds, Kubernetes, and on-prem is the priority, because no paid replacement is named in the facts. Self-hosting requires operating the server, configuring cloud or Kubernetes backends, maintaining SSH fleets, and providing Git and OpenSSH. Do not use it when a mature maintainer record or stable inference optimization is required, because the facts show 0 contributors, 68 open issues, a repo age of 0 years, and experimental presets.

$ uv tool install "dstack[all]" -U
$ dstack server

Applying ~/.dstack/server/config.yml...

The admin token is "bbae0f28-d3dd-4820-bf61-8f4bb40815da"
The server is running at http://127.0.0.1:3000/

$ uv tool install dstack -U

$ dstack project add \
    --name main \
    --url http://127.0.0.1:3000 \
    --token bbae0f28-d3dd-4820-bf61-8f4bb40815da

Configuration is updated at ~/.dstack/config.yml

$ npx skills add dstackai/dstack