> This page is for version 3.3.0 (default). > 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 Harnesses An *agent harness* is the coding agent or agent sandbox that is capable of loading the VSS [Agent Skills](/vision-agent/agent-skills) and to easily deploy and operate the vision-based microservices and workflows. Because every VSS Skill follows the [agentskills.io](https://agentskills.io) specification, any compatible harness can load the same Skills folder and drive VSS without bespoke integration code. The following harnesses are validated against the VSS Agent Skills: * **Claude Code** - Anthropic's CLI coding agent. * **OpenAI Codex** - OpenAI's CLI coding agent. * **OpenClaw** - the open agent sandbox that exposes an `agentskills.io`-compatible host with a chat UI. * **NemoClaw** - NVIDIA's open-source reference stack for running always-on AI agents (such as Hermes and OpenClaw) more safely inside NVIDIA OpenShell sandboxes. > **Note** > > Agent harnesses 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 is built on the > NeMo Agent Toolkit (NAT) and runs inside a deployed VSS workflow. For a broader > comparison of ways to drive VSS, including direct MCP and REST integrations, see the > harness comparison in the [Quickstart](/getting-started/quickstart). ## How Harnesses Use VSS Skills Regardless of which harness you choose, the model is the same: 1. **Install the Skills.** Each Skill folder from the `skills/` directory of the [VSS Blueprint repository](https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization/tree/v3.3.0/skills) is registered into the harness's standard skills directory (typically by symlink, so a `git pull` keeps every Skill current). See [Agent Skills](/vision-agent/agent-skills) for the full install flow. 2. **Load the Skills.** On startup the harness reads each `SKILL.md` frontmatter and adds the Skill to its tool list. Start a new session before invoking a newly installed Skill. 3. **Match and invoke.** The harness matches a natural-language request against the Skills' descriptions and invokes the matching Skill - chaining several Skills when a task needs multiple steps or a missing prerequisite. As described in [Usage Patterns](/vision-agent/agent-skills#usage-patterns), harnesses are used in two complementary ways: * **Deployment.** A coding agent such as Claude Code or Codex loads the Skills and stands up a VSS profile from a request like *"Deploy VSS for video search"*. * **Runtime.** An always-on sandbox such as NemoClaw loads the same Skills and acts as the operational chat interface to a running deployment. ## Claude Code [Claude Code](https://docs.anthropic.com/claude/docs/claude-code) loads VSS Skills from `~/.claude/skills//`. After cloning the VSS Blueprint repository, open it in Claude Code and ask the agent to read `skills/README.md` and register each Skill into that directory (see [Installing Skills](/vision-agent/agent-skills)). Once the Skills are installed, prompt Claude Code in natural language - for example, *"Deploy VSS for video summarization"* or *"Add an RTSP camera at rtsp\://10.0.0.5:554/stream1"* - and it invokes the matching Skill such as `vss-build-vision-ai` or `vss-manage-video-io-storage`. ## OpenAI Codex [OpenAI Codex](https://platform.openai.com/docs/codex) loads VSS Skills from `~/.codex/skills//`. The install and invocation flow is identical to Claude Code - point Codex at the cloned repository, have it register the Skills into the Codex skills directory, and then drive VSS through natural-language prompts. The same Skills catalog is used, so the available capabilities and example prompts are unchanged. ## OpenClaw OpenClaw is an open agent sandbox that exposes an `agentskills.io`-compatible host with a chat UI. As a generic compatible host, it loads VSS Skills from `~/.agents/skills//`. OpenClaw is the foundation that [NemoClaw](/vision-agent/agent-harnesses#nemoclaw) builds on: NemoClaw packages OpenClaw together with the VSS plugin, VSS policy preset, and the VSS Orchestrator MCP server so that Skills, safety policy, and host orchestration are preconfigured rather than installed by hand. ## NemoClaw [NVIDIA NemoClaw](https://github.com/NVIDIA/NemoClaw) is an open-source reference stack for running always-on AI agents more safely inside NVIDIA OpenShell sandboxes. If Build Vision AI provisioned NemoClaw, follow the Quickstart's [handoff](/getting-started/quickstart#continue-in-nemoclaw) to connect to its chat and copy videos into the sandbox. The notebooks below are the alternative standalone setup route. Instead of installing Skills by hand, the bundled `deploy_nemoclaw.ipynb` and `deploy_vss_orchestrator.ipynb` notebooks apply the VSS policy, install the repository's VSS Skills and workspace docs into the sandbox, register the VSS Orchestrator MCP, and start the [VSS Orchestrator MCP server](/vision-agent/agent-harnesses/nemoclaw/vss-orchestrator-mcp) on the host. From the chat UI you can deploy profiles, manage cameras and streams, query incidents, run searches, summarize videos, and pull logs - with confirmation prompts before destructive actions. For setup, configuration, and troubleshooting, see [NemoClaw](/vision-agent/agent-harnesses/nemoclaw). > **Note** > > The NemoClaw notebooks are released as an Early Access feature. ## Related Documentation * [Agent Skills](/vision-agent/agent-skills) - The reusable Skills that every harness loads. * [VSS Agents](/vision-agent/vss-agents) - The in-product agent that runs inside a deployed VSS workflow. * [Quickstart](/getting-started/quickstart) - Harness comparison, including direct MCP and REST options.