Hermes Agent skills at a glance
Hermes Agent keeps its skills in ~/.hermes/skills/, which holds bundled skills, skills installed from the hub and skills the agent wrote itself. A project can have its own in .hermes/skills/. From a terminal, hermes skills install <id> installs from the hub, and npx skills add owner/repo -a hermes-agent copies a skill from GitHub. Of the 195 repositories we have reviewed, 20 say they support Hermes Agent.
This page follows the Hermes skills documentation, which puts a scanner and trust levels in front of hub installs.
Where Hermes loads skills from
The primary location is ~/.hermes/skills/. Extra folders can be listed under skills.external_dirs in ~/.hermes/config.yaml. Project-local skills in <project-root>/.hermes/skills/ or <project-root>/.agents/skills/ load only after you run hermes skills trust. A cloned repo can’t add instructions to your agent until you trust it.
SKILL.md fields and loading
Each skill has YAML frontmatter with name, description and version. A platforms field limits it to macOS, Linux or Windows. A metadata.hermes block holds tags, a category, conditional-activation rules (for example requires_toolsets) and config settings.
Loading is in three levels to save tokens: skills_list() returns names, descriptions and categories, skill_view(name) loads the whole skill, and skill_view(name, path) loads a single reference file. Every installed skill also runs as a slash command, such as /excalidraw. /reload-skills refreshes the commands.
Installing from the hub
hermes skills browse
hermes skills search <query>
hermes skills inspect <id>
hermes skills install <id>
hermes skills update
hermes skills auditRun inspect before install. hermes skills tap add <owner/repo> adds a custom GitHub source, and hermes skills list --source hub shows what came from the hub.
Trust levels and the built-in scanner
Every hub install passes a scanner that checks for data exfiltration, prompt injection and destructive commands. A dangerous verdict always blocks the install. The --force flag overrides only non-dangerous policy blocks. Sources carry one of four trust levels: builtin, official, trusted and community. An optional NVIDIA SkillEvaluator scan is advisory and never blocks. hermes skills audit re-scans what you already have.
A scanner catches patterns, not intent, so treat a clean result as one input. Our own review of any skill starts from the same kind of static scan and then reads what it finds. See how we review.
Skills the agent writes
Hermes can create skills itself through a skill_manage tool, with actions such as create, patch and delete, and a /learn command builds a skill from a path, URL or description. New skills land in ~/.hermes/skills/. If you want a human check, set skills.write_approval: true: every write is staged, and you approve or reject it with /skills pending, /skills approve <id> and /skills reject <id>. External and project directories are not write-protected, so use filesystem permissions if they must be read-only.
Chaining, bundles and housekeeping
Up to five leading skill commands can be chained in one message. Skill bundles, defined as YAML in ~/.hermes/skill-bundles/, group several skills under a single command. hermes skills check looks for updates, hermes skills uninstall <name> removes a skill, and hermes skills reset <name> clears the “user-modified” flag, with --restore bringing back the bundled copy.
Packs that list Hermes Agent
Among the 20 repositories that list Hermes Agent, 13 are marked “review before installing” and 7 “broad access” in our catalog. obra/superpowers is the most starred of them. Check the security page before installing from any hub, and browse the skills directory for other reviews. The pack list below is filtered to repositories that mention Hermes Agent.