NV Tech Assistant is a NemoClaw community example for grounded NVIDIA technical research. It searches authorized NVIDIA sources, GitHub, and arXiv, then answers with citations to evidence retrieved during the current task instead of relying on model memory.
This is an NVIDIA-authored recipe in the public community catalog. Its catalog placement supports discovery; it does not create a separate support or maturity guarantee.
Use it for questions such as:
- How do I use an NVIDIA SDK or find an official sample?
- Which NVIDIA model, NIM, or library fits my use case?
- How do I troubleshoot an NVIDIA SDK error?
- Where is a documented NVIDIA customer success story?
git clone https://github.com/NVIDIA/nemoclaw-community.git
cd nemoclaw-community/examples/recipes/nvidia/nv-tech-assistant
cp .env.example .env # add NVIDIA_INFERENCE_API_KEY; Brave is optional
bash scripts/onboard.sh # create/configure the sandbox
bash scripts/install.sh # apply policies and install the skill
bash scripts/start.sh # ensure the sandbox is runningAsk a question non-interactively:
nemoclaw nv-tech-assistant agent --agent main -m \
"/nv_tech_assistant recommend an NVIDIA model for speech recognition"Or connect to the sandbox and launch the OpenClaw TUI:
nemoclaw nv-tech-assistant connect
openclaw tui- Docker and NemoClaw installed.
- An NVIDIA Endpoints API key from https://build.nvidia.com.
- Optional: a Brave Search API key from https://brave.com/search/api/.
NVIDIA Endpoints is the default inference provider. When BRAVE_API_KEY is
set, onboarding enables Brave Search. When it is blank, web search is disabled
and the skill uses direct retrieval and structured search endpoints from its
allowlisted sources. See .env.example for the exact
variables.
The host-side scripts onboard one NemoClaw sandbox, apply the example's read-only network-policy presets, and install the bundled OpenClaw skill. The skill plans source searches and cited answers inside the sandbox. NemoClaw provides inference through the configured provider; optional Brave Search and the allowlisted public sites remain external services.
Create the sandbox using .env:
bash scripts/onboard.shApply the network policies and install or refresh the skill:
bash scripts/install.shRecover the sandbox gateway and host forwards:
bash scripts/start.shStop the gateway and host auxiliary services while preserving the sandbox, workspace, and credentials:
bash scripts/stop.shRestart it later with scripts/start.sh. To permanently delete the sandbox
instead, run nemoclaw nv-tech-assistant destroy and confirm the destructive
operation.
Override the default sandbox name in .env:
NEMOCLAW_SANDBOX_NAME=my-nv-tech-assistantUse that name in direct nemoclaw commands as well.
scripts/install.sh performs two host-side operations:
- Applies every policy in
policies/withpolicy-add --from-dir. - Deploys
skills/nv-tech-assistant/withskill install.
Both operations target the sandbox named by NEMOCLAW_SANDBOX_NAME. Re-run
the script after editing a policy, the skill, or one of its references.
The assistant receives read-only REST access to the sources it researches:
- NVIDIA documentation, developer resources, blogs, forums, NGC, and build.nvidia.com
- NVIDIA repositories and files on GitHub
- NVIDIA model and dataset pages on Hugging Face
- arXiv pages and its metadata export endpoint
Every endpoint uses access: read-only, protocol: rest, and
enforcement: enforce; ordinary HTTPS endpoints do not skip TLS inspection.
Subdomains are not implied, so each required hostname is listed explicitly.
To inspect the active presets:
nemoclaw nv-tech-assistant policy-listThe skill can also draft a least-privilege policy for a newly requested source. It cannot apply that policy from inside the sandbox; review and apply the YAML from the host.
/nv_tech_assistant recommend an NVIDIA model for speech recognition/nv_tech_assistant show an official TensorRT-LLM sample/nv_tech_assistant I'm getting CUDA error: out of memory with Triton/nv_tech_assistant show a documented customer success story for NVIDIA Riva/nv_tech_assistant draft a read-only policy for medium.com
Answers include links to the exact pages retrieved. Code is included only when needed or requested and is quoted from its identified source.
nv-tech-assistant/
├── .env.example # NVIDIA inference and optional Brave settings
├── policies/
│ ├── arxiv.yaml # arxiv.org and export.arxiv.org
│ ├── github_ext.yaml # GitHub pages, API, and raw files
│ ├── huggingface_readonly.yaml # NVIDIA pages on huggingface.co
│ └── nvidia_ext.yaml # authorized NVIDIA sources
├── scripts/
│ ├── _lib.sh # shared environment and sandbox helpers
│ ├── onboard.sh # create/configure the sandbox
│ ├── install.sh # apply policies and deploy the skill
│ ├── start.sh # start or recover the sandbox
│ ├── stop.sh # stop while preserving workspace state
│ └── tests/
│ └── test_lifecycle_commands.sh
└── skills/
└── nv-tech-assistant/
├── SKILL.md
└── references/
├── nvidia-landscape.md
├── sources-and-search.md
├── nemoclaw-network-policy.md
└── nemoclaw-policy-template.yaml
From this example directory, validate the host-side lifecycle command contracts and shell syntax without creating a sandbox or contacting external services:
bash scripts/tests/test_lifecycle_commands.sh
bash -n scripts/*.sh scripts/tests/*.sh- Retrieval is limited to the explicitly allowlisted hosts. A source that redirects to a new hostname or CDN needs a separately reviewed policy update.
- Results depend on the availability and current contents of external public sources; the recipe does not provide an offline corpus.
- The included checks validate local command and configuration contracts. Live source retrieval still requires the documented credentials and network access.
This example does not vendor or add a package dependency. It runs with the installed NemoClaw and OpenClaw components and contacts the public services listed in the network policies. NVIDIA Endpoints is required for the default inference configuration; Brave Search is optional. Each external service is subject to its own availability and terms.
- The populated
.envis ignored by Git; never commit API keys. - NemoClaw stores inference and Brave credentials through OpenShell provider plumbing rather than writing raw keys into the sandbox configuration.
- Brave is optional and is the only web-search provider configured by these scripts.
- The source policies are read-only and enforce inspected REST methods.
scripts/stop.shpreserves data. Permanent deletion requires the explicitnemoclaw <name> destroycommand.