Skills on NemoClaw — Install, Remove & Network Access
A NemoClaw sandbox runs a normal OpenClaw (or Hermes) agent, so skills are the same SKILL.md folders you'd use outside it. There are two differences. You install local skills from the host with a NemoClaw command. And a skill that calls an outside service only works once the sandbox network policy allows that service.
The earlier version used claw connect, openclaw skill install/verify, a fixed list of "53 official skills" and hand-written ~/.openShell/policies/includes/*.yaml files with openShell policy reload. Current NemoClaw installs local skills with nemoclaw <sandbox> skill install, and opens network access with policy presets.
Install a skill you have locally
Run this on the host. The folder must contain a SKILL.md with a name in its front matter:
nemoclaw my-assistant skill install ./my-skill/
nemoclaw my-assistant skill list
nemoclaw my-assistant skill remove my-skill # use the name from SKILL.md, not the folder name
- NemoClaw checks the folder before uploading: at most 1,024 files and 64 MiB. Dotfiles are skipped, and symlinks or unsafe paths are rejected.
- OpenClaw sandboxes: it hands the folder to OpenClaw's own
openclaw skills install. Removal deletes only/sandbox/.openclaw/workspace/skills/<name>. - Hermes sandboxes: it places the skill at
/sandbox/.hermes/skills/<name>. Start a new Hermes session to pick it up. - NemoClaw keeps no record of what you installed.
skill listasks the agent itself.
Writing the skill in the first place is the same as outside the sandbox. See OpenClaw skills guide or Hermes skills guide.
Skills from ClawHub
The OpenClaw baseline policy already allows clawhub.ai, so ClawHub installs work from inside the sandbox without extra policy:
nemoclaw my-assistant exec -- openclaw skills install <skill-slug>
Read a third-party skill before installing it. The sandbox limits what a bad skill can reach, but it can still misuse anything you've allowed and anything in the workspace.
Skills vs plugins
A skill is instructions in a SKILL.md folder. A plugin is a code package OpenClaw loads, and it's managed by OpenClaw's own CLI inside the sandbox. npm-backed plugins need npm registry access first:
nemoclaw my-assistant policy add npm_registry --yes
nemoclaw my-assistant exec -- openclaw plugins install <package-or-path>
nemoclaw my-assistant exec -- openclaw plugins list
Giving a skill network access
Skills that work only on files in /sandbox, or that just run shell commands, need nothing extra. A skill that calls an outside service will fail at the network call until the policy allows it. Add the matching maintained preset:
| Skill needs | Command |
|---|---|
| GitHub | nemoclaw my-assistant policy add github --yes |
| Gmail over IMAP/SMTP (App Password) | nemoclaw my-assistant policy add gmail --yes |
| Microsoft 365 / Outlook | nemoclaw my-assistant policy add outlook --yes |
| Jira | nemoclaw my-assistant policy add jira --yes |
| Weather lookups | nemoclaw my-assistant policy add weather --yes |
| Python packages | nemoclaw my-assistant policy add pypi --yes |
| Web search | brave (OpenClaw) or tavily |
Always preview with --dry-run first. For a service no preset covers, watch the blocked request in openshell term to see the exact host and program, then write a custom preset limited to that program and those paths.
Gmail note: the gmail preset lets only /usr/bin/python3 open IMAP and SMTP connections, and the App Password lives inside the sandbox where the agent can read it. Use a dedicated App Password and revoke it afterwards. There's no preset for the Gmail REST API or Google OAuth.
Tip: give your coding agent NemoClaw's docs
NVIDIA runs a read-only docs MCP server, which is handy when an agent helps you write skills or policies for NemoClaw:
claude mcp add --transport http fern-docs https://docs.nvidia.com/nemoclaw/_mcp/server
Official references: NemoClaw docs command reference (skill commands) and Install OpenClaw Plugins.
← Back to NemoClaw hub · See also: Skills Guide: Write Your Own · 53 Official Skills Database · OpenShell Policy Configuration