What does this skill do, and when should you use it?
This skill guides an agent through installing Holoscan SDK v4.3+ on Linux x86_64 systems with an NVIDIA GPU and CUDA 13 driver. It covers the Python runtime, C++ development headers, and related RAPIDS dependencies, while requiring a check of NVIDIA’s current official installation instructions first. It also provides Python example validation, headless video-example execution, and troubleshooting for Conda, dependencies, channel solving, and startup segmentation faults. The Conda workflow does not support CUDA 12, aarch64, or iGPU hosts.
Checks Conda and the NVIDIA driver; installs Miniforge non-interactively when needed; creates and activates a Python environment; installs Holoscan, the C++ development package, rmm, ucxx, and cuda-version=13 from rapidsai and conda-forge; installs the C++ toolchain; verifies installation through a Python import or header-directory check; downloads and patches Holoscan hello_world and video_replayer examples; sets the recommended stack limit; runs validation; and diagnoses missing runtime dependencies, channel-order issues, CMake discovery failures, and new-shell Conda initialization problems.
- A developer needs the Holoscan Python runtime on a Linux x86_64 host with a CUDA 13 driver.
- A C++ developer needs Holoscan headers, CMake configuration, and runtime libraries.
- A team wants one environment for combined Python and C++ Holoscan development.
- An engineer needs to validate Holoscan examples over SSH or without a display.
- A user is troubleshooting `librmm.so`, `find_package(holoscan)`, solver selection, or startup segmentation faults.
How do you install this skill?
- Before execution, manually review the remote Miniforge installer, Conda channels, and package versions; consider pinning versions or hashes.
- Installation writes to ~/miniforge3 and provides no automated rollback; obtain user consent and prepare a cleanup plan.
- The video example requires external racerx data, and YAML patching relies on exact text matches, so verification can be misleading if the patch silently misses.
- Conda, GitHub raw content, and docs.nvidia.com are core network dependencies; mainland-China users may need reachability, mirror, or proxy planning.
- Shell / CLI
- Network access
- Local filesystem
Linux x86_64NVIDIA GPUCUDA 13 driverConda or MiniforgePython 3.10-3.13conda-forgerapidsai
Install this skill into an Agent Skills-compatible client with npx skills add nvidia/skills --skill holoscan-install-conda --yes. On the target Linux host, check conda --version and nvidia-smi. If Conda is unavailable, run the supplied Miniforge download and silent-install commands, then execute source ~/miniforge3/etc/profile.d/conda.sh in the current shell. Create the environment with conda create -n holoscan python=3.13 -y; after activation, choose holoscan or libholoscan-dev for the intended use and install rmm ucxx cuda-version=13 -c rapidsai -c conda-forge -y. C++ development also requires cxx-compiler cmake ninja. Before installation, fetch NVIDIA’s current official Conda instructions and follow them if package names, channels, or versions differ.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/holoscan-install-conda" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- import holoscan; print(holoscan.__version__)
Example trigger: Install Holoscan with Conda on this Linux x86_64 NVIDIA host with a CUDA 13 driver, then validate both Python and C++ support. Activate the environment and run ulimit -s 32768. Validate Python with python3 -c "import holoscan; print(holoscan.__version__)"; validate C++ support with ls "$CONDA_PREFIX/include/holoscan". Before running video_replayer, the skill downloads the matching examples and changes the configuration to 10 frames, no repeat, no realtime, and headless mode. HOLOSCAN_INPUT_PATH must point to a directory containing a racerx/ subdirectory. The SKILL.md does not define one universal automated method for obtaining that data.
What are this skill's strengths and limitations?
- Covers Python runtime and C++ development dependencies, including the distinct role of `libholoscan-dev`.
- Requires checking the official documentation, helping account for changing package names, channels, and versions.
- Provides copyable installation, validation, example-running, and troubleshooting commands.
- Uses bounded, headless testing suitable for SSH and display-less environments.
- The workflow is limited to Linux x86_64 and CUDA 13 drivers; CUDA 12, aarch64, and iGPU hosts are excluded.
- It depends on network access, Conda channels, NVIDIA documentation, and example data.
- The video example requires input data containing `racerx/`; the source gives guidance but not one complete automated data-install flow.
- No independent test suite or broader platform-coverage evidence is provided.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Holoscan Conda Installer this page ✓ NVIDIA · Official | 52 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Holoscan Python Wheel Installer ✓ NVIDIA · Official | 50 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Holoscan Native Ubuntu Installer ✓ NVIDIA · Official | 52 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Holoscan SDK Source Builder ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Holoscan SDK Setup Guide ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
The skill explicitly redirects CUDA 12 hosts to /holoscan-install-container or /holoscan-install-wheel. The source material does not provide enough information for further product or installation-method comparisons.
How did FollowSkills review this skill?
The prerequisites, network endpoints, and platform scope are disclosed, and the commands are mostly limited to Conda setup and temporary files; no credential theft or covert exfiltration is evident. However, the skill silently downloads and executes a remote Miniforge installer, modifies the user directory, lacks per-source confirmation and rollback/uninstall guidance, so points are deducted.
The workflow is broadly self-consistent and includes prerequisite checks, verification commands, and troubleshooting. Reliability is reduced because it depends on live documentation, remote downloads, and unpinned versions; the YAML text replacement may fail, and this review has no reproducible skill-specific execution evidence, so the static cap applies.
The audience, Linux x86_64/CUDA 13/Conda scenario, and exclusions for CUDA 12, aarch64, and iGPU are fairly clear, with Python and C++ branches. Trigger boundaries, driver compatibility details, and non-GPU handling remain limited, and core network dependencies may be difficult to reach from mainland China, so points are deducted.
Frontmatter, version, Apache-2.0 license, author, purpose, limitations, procedure, verification, and troubleshooting are present and readable. It lacks explicit Instructions/Examples sections, a changelog, clear maintenance ownership/update path, and complete installer/dependency security notes, so points are deducted.
For a typical CUDA 13 Linux Conda installation, it covers environment creation, Python/C++ packages, and basic checks and is operationally useful. However, commands contain placeholder paths, test data must be obtained separately, package names and versions may change with the live documentation, and static evidence cannot confirm successful results; the score stays below the static ceiling.
It cites NVIDIA installation documentation, the product repository, and committed evaluation materials, making key guidance auditable. There is no skill-specific CI or committed execution test, and the benchmark is primarily a report rather than independently reproducible evidence, so points are deducted.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →