NVIDIA AI Workbench
AI Workbench is NVIDIA's free tool for running containerized, Git-managed AI projects on a laptop, a GPU workstation, a remote server or a cloud instance with the same interface. It handles containers, GPU drivers and remote connections; the latest release, 2026.06.5, shipped in July 2026.12
Also known as AI Workbench, nvwb
At a glance
- What is it?
- AI Workbench is a development environment manager. You install a desktop app or a command-line tool (nvwb) on your own computer and add locations, which are machines with Workbench installed, such as your laptop, a GPU workstation, a cloud instance from Brev or a Jetson device. Each project is a Git repository with a .project/spec.yaml file that describes its container, packages, GPU needs and web applications such as JupyterLab. Workbench builds the container, attaches GPUs and opens the apps, so the same project runs the same way on every location.4
- What does it do?
- Workbench removes much of the setup work around AI development. It installs or detects the container runtime and NVIDIA driver on a location, builds the project container, mounts code and data, and starts tools such as JupyterLab, RStudio, VS Code or Cursor. It tracks changes to code and environment in Git, connects to GitHub, GitLab and self-hosted GitLab, pulls private containers and models from NGC, and launches cloud GPU instances through Brev. A Build Assistant explains container build failures, and the CLI returns JSON so scripts and AI agents can drive it. Coding agents such as Claude Code can run inside the project container with sandboxing rules.45
- Who needs it?
- Individual developers and small teams who move AI work between a laptop and remote GPU machines and are tired of rebuilding environments by hand. It also suits teams that want reproducible example projects, such as RAG or fine-tuning, and developers who want agents to work inside an isolated container rather than on their host.
- What does it need?6
- Windows 11 (build 22631 or later), Windows 10 22H2 (build 19045.4052 or later), macOS 14 or later, or Ubuntu 22.04, 24.04 or 26.04
- Desktop app only: about 500 MB disk and at least 8 GB of memory
- Full local install: Docker Desktop or Podman (Docker or Podman on Ubuntu), 16 GB RAM minimum and 32 GB recommended
- An NVIDIA GPU and driver on any location where you want GPU acceleration (no GPU needed for the client)
- SSH access to remote Ubuntu hosts for remote locations
- Optional: GitHub or GitLab account, NGC API key, Brev account
- What it is not
- AI Workbench is not a cluster scheduler or an MLOps platform: it does not share a GPU pool across many users or manage production model serving. It is not a cloud service either; it runs on machines you provide or rent. It does not require a GPU on the machine where you run the desktop app, and it does not check that a container's CUDA version matches the host driver. Free use does not include support, which needs an AI Enterprise license.16
Availability and licensing. NVIDIA states AI Workbench is free to use; support requires an NVIDIA AI Enterprise license, and use is governed by the NVIDIA AI Enterprise End User License Agreement. Releases continue: the current release notes list 2026.06.5 (July 2026), and the version history lists 2026.03.3 (April 2026), 2025.10.14 (17 November 2025), 2025.09.5 (30 September 2025) and 2025.06.4 (13 June 2025). The hybrid RAG example repository's last commit is from August 2025.157
The problem it solves
AI projects often work on one machine and break on the next because of different drivers, CUDA versions, Python packages or container settings. Moving from a laptop to a GPU server or a cloud instance usually means SSH tunnels, manual container commands and rebuilt environments.
Sharing a project with a colleague repeats the same problem.
AI Workbench keeps the environment definition in the project's Git repository and automates container builds, GPU attachment and remote connections, so a project behaves the same on each machine.
How it works
- Install a client. The desktop app runs on Windows, macOS or Ubuntu; the nvwb CLI offers the same functions.
- Add locations. Install Workbench on the local machine (with Docker or Podman) or on remote Ubuntu hosts over SSH; NVIDIA Sync and Brev can set up remote locations for you.
- Create or clone a project. A project is a Git repository with .project/spec.yaml describing the base container, packages, scripts and applications.
- Build and run. Workbench builds the container, runs pre-build, post-build and on-start scripts, attaches the requested GPUs and starts apps such as JupyterLab.
- Work and version. Code and environment changes are shown as Git changes to review, commit or roll back, and can be pushed to GitHub or GitLab.
- Move. Open the same project on another location; the container is rebuilt from the same definition.28
Diagram as a list
Applications & solutions
- Desktop app and nvwb CLIWhere the user or an agent manages projectsConnects to Workbench service on each location
- Git project with .project/spec.yamlDefines environment, scripts and appsConnects to Workbench service on each location
Operations & orchestration
- Workbench service on each locationBuilds containers, attaches GPUs, runs appsConnects to Project container or Compose stack (Docker or Podman)
- Project container or Compose stack (Docker or Podman)Runs code, JupyterLab, IDEs and sandboxed agentsConnects to Local PC, RTX workstation, remote server, cloud or Jetson GPU
Accelerated computing
- Local PC, RTX workstation, remote server, cloud or Jetson GPUHardware that runs the project
Networking, power & facilities
- GitHub/GitLab, NGC, Brev, NVIDIA EndpointsCode hosting, container registry, cloud GPUs, Build AssistantConnects to Workbench service on each location, Git project with .project/spec.yaml
Capabilities
Same projects on local and remote locations46
Locations are hosts with Workbench installed; the interface is the same for a laptop, a cloud instance or a data center server, with SSH tunneling and remote container commands handled for you.
Why it matters: Removes manual setup when moving work to a bigger GPU machine.
Limits: Remote installs are tested on Ubuntu only; Jetson on JetPack 6.1 is expected to work with some adjustments but is not covered by QA.
Containerized, Git-managed environments46
Each project is a Git repository with .project/spec.yaml describing its container, packages, scripts, GPU settings and applications, plus optional multi-container Compose stacks.
Why it matters: Makes environments reproducible and shareable through normal Git hosting.
Limits: Workbench does not validate that the container's CUDA version works with the host driver.
GPU and driver management6
Workbench installs the NVIDIA driver on Ubuntu if none is present, choosing open or proprietary kernel modules for the GPU, and passes GPUs to containers.
Why it matters: Saves driver and container toolkit setup on new machines.
Limits: It does not replace an existing driver, can fail with mixed GPU generations, and does not manage vGPU setup.
Integrations with Git hosts, NGC and Brev29
Connects to GitHub, GitLab and self-hosted GitLab, pulls from private NGC registries with a scoped key, and launches cloud GPU instances through Brev.
Why it matters: Links the tool to where code, containers and cloud GPUs already live.
Limits: Each integration needs its own account and credentials.
Automation and agent support1011
The nvwb CLI matches the desktop app and returns JSON for scripts and agents; guides show how to run Claude Code inside a project container with sandboxing, permissions and hooks.
Why it matters: Lets coding agents work on a project without touching the host system.
Limits: Sandbox settings must be configured per project, and agent changes still need review.
Build Assistant59
Analyzes failed project container builds and suggests fixes, using NVIDIA inference endpoints.
Why it matters: Shortens debugging of broken environments.
Limits: Needs an NVIDIA Endpoints API key; suggestions should be checked before applying.
Example projects and blueprints712
Example projects include hybrid RAG, a downloadable NIM deployment, and AI Blueprint projects such as PDF to Podcast, RAG and AI-Q research.
Why it matters: Gives ready-to-clone starting points.
Limits: The hybrid RAG example repository has had no commits since August 2025; check each example repository's activity before relying on it.
Practical use cases
A data scientist prototypes on a laptop but needs a GPU server for training.
- Approach
- Create the project in Workbench on the laptop, add the GPU server as a remote location, and open the same project there.
- Role of NVIDIA AI Workbench
- Rebuilds the identical container on the server and handles the remote connection.
- Data, infrastructure and skills
- SSH access to an Ubuntu GPU server and a Git host for the project.
- Type of benefit
- Less environment setup time
- Caveats
- Large datasets still need to be moved or mounted separately.
- First step
- Install Workbench on the server through the remote install guide and add it as a location.
A team wants to try a RAG application with NVIDIA models before committing to a stack.
- Approach
- Clone the RAG Blueprint or hybrid RAG example project and run it on a local GPU or a Brev instance.
- Role of NVIDIA AI Workbench
- Provides a ready project with container, apps and settings.
- Data, infrastructure and skills
- A GPU location, API keys for any hosted models used, and test documents.
- Type of benefit
- Faster evaluation
- Caveats
- Example projects are starting points and may lag behind current model versions.
- First step
- Clone the hybrid RAG example project from GitHub in Workbench.
A developer wants a coding agent to work on a project without risking the host system.
- Approach
- Install Claude Code inside the project container and apply a sandbox configuration with permission rules and hooks.
- Role of NVIDIA AI Workbench
- Supplies the isolated container and the documented sandbox setup.
- Data, infrastructure and skills
- A Workbench project, a public repository for the sandbox configuration, and an agent account.
- Type of benefit
- Reduced risk from agent actions
- Caveats
- Sandbox rules need maintenance; agent output still needs review.
- First step
- Follow the Install Claude Code in a Project Container quickstart.
Works with
Optional integration
- NVIDIA AI EnterpriseSupport for AI Workbench requires an NVIDIA AI Enterprise license.
- NVIDIA NIMExample projects include a downloadable NIM deployment, and the RAG Blueprint project uses NIM microservices.
- NVIDIA JetsonJetson devices can be added as remote locations through a CLI install.
Complementary tools
- NVIDIA BlueprintsAI Blueprint example projects (PDF to Podcast, RAG, AI-Q) run in Workbench.
Alternative approaches
- NVIDIA Run:aiFor sharing GPU pools across many users and jobs, a scheduler such as Run:ai fits; Workbench manages individual project environments.
Relationship labels follow NVIDIA's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Getting started
Check the support matrix
Confirm your OS version and choose desktop-only, full local (with Docker or Podman) or remote install.
Check: Your OS and install type appear in the support matrix.
Install AI Workbench
Download and run the desktop app installer for your OS; for a full local install, let it set up Docker or Podman.
Check: The desktop app opens and shows the local location.
Add a GPU location
Add a remote Ubuntu GPU host over SSH, or use NVIDIA Sync or Brev to create one.
Check: The remote location appears and reports its GPUs.
Clone an example project
Clone an example such as hybrid RAG or a blueprint project, then build its container.
Check: The build finishes and JupyterLab or the project app opens.
Connect Git and version your work
Link GitHub or GitLab, commit code and environment changes, and push.
Check: A colleague can clone the project and build the same environment.
Decide on support
Use community forums, or obtain support through an NVIDIA AI Enterprise license.
Check: Support route is agreed before relying on Workbench for team work.
Official resources
Could this technology help you?
Describe your project to the Solution Architect. It starts with NVIDIA AI Workbench as context but recommends independently, including when you do not need it.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
- NVIDIA AI Workbench (product page) (opens in a new tab)
- AI Workbench release notes: current release (opens in a new tab)
- AI Workbench: install on NVIDIA Jetson (opens in a new tab)
- AI Workbench User Guide: introduction (opens in a new tab)
- AI Workbench release notes: version history (opens in a new tab)
- AI Workbench support matrix (opens in a new tab)
- NVIDIA/workbench-example-hybrid-rag repository (opens in a new tab)
- AI Workbench: use NVIDIA Sync (opens in a new tab)
- AI Workbench: NVIDIA integrations (opens in a new tab)
- AI Workbench: CLI for agents (opens in a new tab)
- AI Workbench quickstart: Claude Code sandbox configuration (opens in a new tab)
- AI Workbench example projects: AI Blueprints (opens in a new tab)
- AI Workbench example projects: NIM deployments (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.