> This page is for version Dev. > For other versions, use one of these documentation indexes: > - 3.3.0 (default): https://docs.nvidia.com/vss/3.3.0/llms.txt > - Dev: https://docs.nvidia.com/vss/dev/llms.txt > For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/vss/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/vss/_mcp/server. # Agent Skills > **Warning** > > Agent Skills are released as **Early Access** and are not intended for use in production environments. > Agent Skills for VSS were validated using Claude Opus 5. Agent Skills are reusable, self-contained capabilities that coding agents (such as Claude Code and Codex) can load to deploy and operate the VSS Blueprint. Each Skill follows the [agentskills.io](https://agentskills.io) specification and packages prompts, reference data, and helper scripts so the coding agent can deploy a VSS profile, manage cameras and streams, query incidents, run searches, summarize videos, and generate reports. All from natural-language requests. > **Note** > > Agent Skills are a developer-side tool for working with the VSS Blueprint. They are > distinct from the in-product [VSS Agent](/vision-agent/vss-agents), which runs inside a > deployed VSS workflow and orchestrates VSS microservices. ## Skill Anatomy Each VSS Skill is a folder under the `skills/` directory of the VSS Blueprint repository, inside the category folder that describes when you reach for it (`deployment/`, `operations/`, `tools/`, `benchmarking/`) or, for the `vss-build-vision-ai` entry point, at the top level. A VSS Skill folder contains the following: * A `SKILL.md` file with YAML frontmatter declaring the Skill's `name`, `description`, `version`, and `license`. The `description` tells the coding agent when to invoke that VSS Skill. For example, when the user asks to deploy a VSS profile, add a camera, query alerts, or summarize a video. * The body of `SKILL.md` documents the VSS commands, API calls, and patterns the Skill uses. Typically `curl` calls against the VSS REST API, the VA-MCP server, or VIOS. * Optional reference files such as VSS deployment notes, alert-type prompt templates, or example payloads for the workflow the Skill targets. * Optional evaluation specs under `eval/` that are exercised by the VSS Skills Eval CI workflow on every pull request that touches the `skills/` directory of the VSS Blueprint repository. When a coding agent loads its skills directory, it reads each `SKILL.md` frontmatter and adds the VSS Skill to its tool list. The agent then matches user requests against the Skills' descriptions and invokes the appropriate VSS Skill for each task. After installing a new Skill, start a new agent session before invoking it. Symlinked Skills do not need to be reinstalled after a repository update; start a new session to load the updated Skill. ## Usage Patterns VSS Agent Skills are used in two complementary ways: * **Deployment.** A coding agent such as Claude Code or Codex loads the Skills, takes a natural-language request like *"Deploy VSS for video search"*, and invokes `vss-build-vision-ai` to select the profile or capability set, run pre-flight checks, generate the Compose plan, and bring up the stack. The developer does not need to remember exact flags, profile names, or compose commands. * **Runtime.** Once VSS is running, an always-on coding agent such as NemoClaw loads the same Skills and acts as the operational interface to the deployment. The user can manage cameras and streams, submit and verify alerts, run searches, summarize videos, and generate reports through natural-language chat. ## Available Skills The official VSS Agent Skills are published in the `skills/` folder of the [VSS Blueprint repository](https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization/tree/develop/skills) on GitHub, and mirrored to the public NVIDIA Skills catalog at [github.com/nvidia/skills](https://github.com/nvidia/skills). Skills are grouped in the repository by when you reach for them: `deployment/` stands a stack up, `operations/` drives one that is already running, `tools/` and `benchmarking/` are standalone utilities, and `vss-build-vision-ai` sits at the top level as the entry point that composes the rest. **A Skill's invocable name is its own directory name — the category is repository organisation and never appears in the installed path or in the slash command.** ### Entry point | Skill | Description | | ------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | [`vss-build-vision-ai`](/vision-agent/build-vision-ai) | Primary build and deployment Skill for VSS. Deploys stock developer profiles, warehouse industry profiles, and custom capability combinations from natural-language requests. Hands off to the operations Skills once a stack is running, and to a single deployment Skill when only one microservice is wanted. | ### Deployment | Skill | Description | | -------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | [`vss-deploy-warehouse-helm`](/vision-agent/agent-workflows/industry-specific-examples/warehouse-operations-blueprint/helm-deployment) | Deploy or upgrade the warehouse blueprint (2D / 3D / MV3DT) on Kubernetes via Helm, with GPU-aware `NUM_STREAMS` capping so the request never exceeds what the perception pipeline can sustain. | | `vss-deploy-dense-captioning` | Deploy and call the RT-VLM dense captioning microservice on stored video files and live RTSP streams. Captions, alerts, stream management, or OpenAI-compatible completions. | | `vss-deploy-detection-tracking-2d` | Deploy, operate, debug, or tear down the RTVI-CV (Real Time Video Intelligence CV) perception microservice locally, and call its REST API to manage streams, health probes, and metrics. Supports the `warehouse-2d`, `warehouse-3d`, `smartcity-rtdetr`, and `smartcity-gdino` use cases. | | `vss-deploy-detection-tracking-3d` | Deploy and operate the RTVI-CV-3D stack (MV3DT / Multi-View 3D Tracking) — per-camera DeepStream perception plus BEV Fusion over multiple calibrated cameras — on sample, custom-video, or RTSP inputs. Auto-chains to `vss-generate-video-calibration` when calibration data is missing. For sample prompts and expected agent steps, see the [MV3DT Agent Skills walkthrough](/vision-agent/agent-workflows/industry-specific-examples/warehouse-operations-blueprint/development-workflow/deployment-customization/mv3dt-vision-ai-profile#agent-skills-walkthrough). | | `vss-deploy-video-embedding` | Deploy and operate the RT-Embed video-embedding microservice — `/v1` REST API for file, text, and video embeddings and live RTSP, plus Redis, Kafka, and OpenTelemetry integration. | | `vss-setup-behavior-analytics` | Deploy the `vss-behavior-analytics` service standalone, with custom configs or calibration, without the full warehouse stack. | | `vss-setup-video-analytics-api` | Deploy the `vss-video-analytics-api` REST service standalone against custom Elasticsearch and Kafka infrastructure. | ### Operations | Skill | Description | | ------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `vss-search-archive` | Search video archives using natural language with fusion search over Cosmos Embed1 embeddings and computer vision attribute matching. | | `vss-summarize-video` | Summarize a video by routing short clips to the VLM NIM and longer videos to the Video Summarization microservice. | | `vss-ask-video` | Answer a question about previously analyzed or freshly scoped video, routing through hot conversation context, agent Markdown memory, structured VSS memory, bounded memory introspection, or an exact-window `vss vlm run`. Not for retrieval or metadata-answerable questions. | | `vss-generate-video-report` | Produce a video analysis report through one of two documented backends — Mode A per-clip VLM (delegating to `vss-summarize-video` when the clip is 120 seconds or longer) or Mode B incident-range via `vss-query-analytics` — then render markdown output. Never routes through the VSS agent's `/generate` endpoint. | | `vss-generate-video-report-rag` | Generate video summary reports using the `video_search_frag` extension with Video Summarization, Enterprise RAG context, and HITL parameter collection. | | `vss-query-analytics` | Query video analytics data and metrics from Elasticsearch via the VA-MCP server. | | `vss-manage-alerts` | Add, manage, and monitor alerts on streamed video in either CV verification mode or VLM real-time mode. | | `vss-manage-video-io-storage` | Manage video and stream operations, recording timelines, clip extraction, and snapshots using the Video IO and Storage (VIOS) microservices. | ### Tools | Skill | Description | | -------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | `vss-generate-video-calibration` | Calibrate or tune a multi-camera dataset with AutoMagicCalib (AMC) from local MP4s, live RTSP streams, or the bundled sample dataset. See the [Agent Skills walkthrough](/system-components/calibration/auto-calibration/getting-started#agent-skills-walkthrough) for calibration and bounded tuning prompts, or [Auto Calibration Workflow Steps](/system-components/calibration/auto-calibration/workflow-steps) for the complete seven-step UI workflow. | ### Benchmarking | Skill | Description | | ----------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | `vss-benchmark-vlm-qa` | End-to-end video Q\&A accuracy and latency on `vss-devx-base` through `vss vlm run` (CR3 RT-VLM). | | `vss-benchmark-video-summarization` | Long Video Summarization latency and burst throughput on a deployed summarization instance. | | `vss-evaluate-caption-accuracy` | Check whether an RT-VLM configuration change moved caption quality: capture paired baseline and candidate captions, score both against a ground truth with an LLM judge, and emit an accuracy and processing-time table. | > `vss-deploy-profile` was removed in VSS 3.3.0. Use > [`vss-build-vision-ai`](/vision-agent/build-vision-ai) instead — the VSS > 3.3.0 Release Notes map its `-p` / `-m` flags to the equivalent workflows. ## Report Skill Error Handling The `vss-generate-video-report` skill applies the following fail-fast handling: * **Profile probe failure**: stop and ask the user to deploy the required profile via `vss-build-vision-ai` (`base` for Mode A, `alerts` for Mode B). * **Clip URL rewrite env missing**: if any of `VSS_PUBLIC_HTTP_PROTOCOL`, `VSS_PUBLIC_HOST`, or `VSS_PUBLIC_PORT` is unset, fail with the missing variable instead of emitting malformed browser URLs. * **VLM endpoint/model mismatch**: if `${VLM_ENDPOINT}/models` fails or does not list `${VLM_MODEL}`, try the alternate backend once; otherwise surface the concrete error. * **Remote endpoint cannot fetch internal clip URL**: do not issue a doomed VLM request; return a reachability requirement for `VIDEO_URL` instead. * **No incidents in Mode B**: return an explicit zero-results report for the requested range/scope; do not invent incidents and do not auto-switch to Mode A. ## Supported Agent Harnesses VSS Agent Skills can be loaded by any agent harness that follows the [agentskills.io](https://agentskills.io) specification. The following agent harnesses are validated: * [Claude Code](https://docs.anthropic.com/claude/docs/claude-code) * [OpenAI Codex](https://platform.openai.com/docs/codex) * [OpenClaw](https://github.com/openclaw/openclaw) * [NemoClaw](https://github.com/NVIDIA/NemoClaw) For more information on agent harnesses with VSS, see [Agent Harnesses](/vision-agent/agent-harnesses). Each agent harness expects Skills under a standard directory: | Host | Skills directory | | --------------------------------------- | -------------------------- | | Claude Code | `~/.claude/skills//` | | Codex | `~/.codex/skills//` | | Other `agentskills.io`-compatible hosts | `~/.agents/skills//` | ## Discovering and invoking a Skill From a new session opened at the checkout root, ask the agent to read `skills/README.md` and list the installed VSS Skills. Installation takes a Skill directory containing `SKILL.md`; the category name `operations` is not part of the invocable name. If discovery misses a Skill, give its path or use an explicit invocation. | Host | Explicit deployment prompt | | ------------------------ | --------------------------------------------------------------------------------------------------- | | Codex | `$vss-build-vision-ai Deploy the VSS base profile on this machine with NemoClaw.` | | Claude Code | `/vss-build-vision-ai Deploy the VSS base profile on this machine with NemoClaw.` | | Other compatible harness | `Read skills/vss-build-vision-ai/SKILL.md and use it to deploy the VSS base profile with NemoClaw.` | Use a model with tool use and enough context for the deployment plan. Choose high reasoning effort if your host exposes that control. The validation statement above records the tested model; it does not establish that every model/effort combination has been tested. If an intake question tool is unavailable, ask the agent to present the question in ordinary conversation. ## Installing Skills Skills are installed by asking the coding agent itself to read the Skills catalog and register each Skill into the host's skills directory. Symlinks are used so that a `git pull` of the Blueprint repository keeps every Skill up to date automatically. ### Step 1: Clone the VSS Blueprint repository ```bash git clone https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization.git cd video-search-and-summarization git checkout "develop" ``` ### Step 2: Ask the coding agent to install the Skills `vss-build-vision-ai` is the entry point — it composes and deploys a stack from an intent. It deliberately stops at the deploy boundary, routing search, summarization, VIOS, alerts, reports, and video Q\&A to the matching operations Skill. Installing it on its own gives an agent that can build a stack but not drive it, so install the entry point together with the operations Skills: ```text Install these VSS skills for this host: skills/vss-build-vision-ai/ and every skill under skills/operations/. Use the host's standard skills directory, naming each install after the skill's own directory name and not its category. Symlink rather than copying so a git pull here keeps them current, and skip any already pointing at this checkout. List what you registered and where. ``` Add `skills/deployment/` when you need a single microservice on its own or a Helm deployment, `skills/tools/` for standalone camera calibration, and `skills/benchmarking/` for performance measurement. To install the whole catalog: ```text Read skills/README.md and every SKILL.md file under skills/. For each skill in the catalog, install it for this host so I can invoke it from a shell or chat session. Use the host's standard skills directory, naming each install after the skill's own directory name rather than its category. Symlink each skill folder rather than copying it so a git pull here keeps every install up to date. Skip skills that are already installed and pointing at this checkout. When you're done, list the skills you registered and which directory you used. ``` The agent reads each `SKILL.md` frontmatter, creates the symlinks, and confirms what is installed. The Skills become invokable in the next agent session. To install only a single Skill, give the agent the path as it appears in the catalog above — most Skills sit inside a category, while `vss-build-vision-ai` is at the top of `skills/`: ```text Install only skills// for this host the same way — for example skills/operations/vss-search-archive/ or skills/vss-build-vision-ai/. Register it under the skill's own directory name, without any category prefix. ``` ## Using Skills With VSS Agent Skills installed, prompt the coding agent in natural language to deploy a profile or operate the deployment. The agent invokes the matching Skill from the catalog. After installing Skills, see the following guides to get started: * [Quickstart: Agent Skills Walkthrough](/getting-started/quickstart#agent-skills-walkthrough) — deploy and operate the base profile using Skills. * [Build Vision AI Skill](/vision-agent/build-vision-ai) — primary deployment entry point for stock profiles, warehouse profiles, and custom vision stacks. * [Video Summarization](/vision-agent/agent-workflows/video-summarization) — skills-based workflow for video summarization. * [Video Search](/vision-agent/agent-workflows/search) — skills-based workflow for natural language video search. * [Real-Time Alerts](/vision-agent/agent-workflows/real-time-alerts) — skills-based workflow for real-time alert generation. Once the deployment is up and running, Skills can be used to perform workflow tasks or interact with individual microservices. Here are some prompt examples which would call the `vss-manage-video-io-storage` Skill for Video IO and Storage (VIOS): ```text Add an RTSP camera at rtsp://10.0.0.5:554/stream1. ``` ```text Extract a video clip from 30 to 40 seconds (mp4 container) from the uploaded video and return a shareable URL. ``` ## Updating Skills After a `git pull` of the Blueprint repository, the symlinks already resolve to the updated content, so a content change needs no action. Two cases do, because an install is a symlink to a repository path rather than to a name: a **new** Skill has no symlink yet, and a Skill that **moved between category folders** (or was renamed) leaves its old symlink dangling — the name still appears installed while its target no longer exists. To handle both, ask the agent: ```text Re-check this host's skills directory against skills/README.md. Add a symlink for any catalog skill that is missing. Then check every VSS skill symlink already there: if its target no longer exists, or points somewhere other than that skill's current path in this checkout, repoint it. Remove symlinks for skills that are no longer in the catalog. Report what you added, repointed, and removed. ``` To uninstall every Skill, ask the agent: ```text Remove every VSS skill symlink you previously created under this host's skills directory. ``` ## Related Documentation * [VSS Agents](/vision-agent/vss-agents) - The in-product agent that runs inside a deployed VSS workflow. * [Agent Workflows](/vision-agent/agent-workflows) - End-to-end workflows demonstrated through developer profiles.