The Skills Hub: Browsing, Installing, and Managing 672 Skills

672 Skills, Four Registries

The Skills Hub is the community layer of the Hermes skills ecosystem. Rather than building every skill from scratch, you can discover and install skills that others have published for common tools and workflows. As of the current release, the Hub contains 672 skills across four registry sources.

GitHub: The GitHub adapter fetches skills from any GitHub repository using the Contents API. Major AI platforms — OpenAI, Anthropic, Hugging Face, and NVIDIA — maintain trusted skill repositories that are automatically recognized as trusted sources. The GitHub adapter is the most general-purpose source, supporting any public repository that follows the SKILL.md format.

skills.sh: A dedicated community skill marketplace with a searchable index. Skills on skills.sh have install-count rankings, versioning, and their own security scanning pipeline. The Hermes skills.sh adapter proxies search and delegates file fetching to the GitHub adapter once the underlying repository is identified.

ClawHub: A dedicated API with search endpoints and exact slug lookups. ClawHub is optimized for precise skill discovery — you can retrieve a specific skill by its exact identifier without a broad search.

LobeHub: Fetches a remote JSON index of skills and filters based on search queries. LobeHub is a community skills directory with a strong focus on AI assistant skills and integrations.

Searching the Hub

Hub operations — searching, browsing, installing, and managing taps — are CLI-only. They are not available from within a chat session via slash commands.

hermes skills search docker

This searches across all configured registries and returns a table of results with the skill name, registry source, description, and trust level. The search is case-insensitive and matches against skill names, descriptions, and tags.

To list all installed skills and their origin:

hermes skills list

This shows installed skills categorized by how they were installed: hub-installed, builtin (bundled with Hermes), or locally-created. The distinction matters because each category has different update and reset behaviors.

The Install Workflow: Fetch, Quarantine, Scan, Install

When you run hermes skills install [source]/[skill-name], the installation follows a four-stage workflow designed to prevent malicious skills from reaching your agent.

Fetch: The skill files are downloaded from the registry source to a temporary location. No execution occurs during this stage.

Quarantine: The downloaded files are placed in a quarantine directory (~/.hermes/hub/quarantine/) while scanning runs. Skills in quarantine are not accessible to the agent.

Scan: The Skills Guard system performs static analysis on every file in the quarantined skill. The scan is the security gate between internet-sourced content and your agent's skill library.

Install: If the scan passes, the skill is moved from quarantine to ~/.hermes/skills/ and becomes immediately available. If the scan detects issues above a threshold determined by the skill's trust level, installation is blocked. The --force flag overrides non-dangerous findings for trusted sources.

Understanding the Security Scan

The Skills Guard system is designed to catch three categories of attack that malicious SKILL.md files might attempt.

Exfiltration attempts: Patterns that would send environment variables, credentials, or session data to external servers. The canonical example is a procedure step instructing the agent to curl a remote URL with environment variables in the request body. Skills Guard recognizes this pattern and flags it as dangerous.

Prompt injection attempts: Text designed to hijack the agent's behavior by overriding its instructions. Phrases like "ignore previous instructions" or "you are now" within skill content are flagged as suspicious. Prompt injection in skills is particularly dangerous because skill content loads with elevated trust relative to user messages.

Obfuscation attacks: Invisible Unicode characters and homoglyph substitutions are tools attackers use to disguise malicious content from human reviewers. Skills Guard scans for these at the byte level and flags any detected obfuscation.

Scan verdicts are categorized as safe, caution, or dangerous. The installation policy varies by trust level: builtin and trusted-source skills with caution verdicts can still install; community skills with caution verdicts are blocked by default and require --force; dangerous verdicts block all sources regardless of trust level.

Trust Levels and What They Mean

Every skill in the Hub is assigned one of four trust levels that determine how the installation policy handles scan verdicts.

Builtin: Skills bundled with the Hermes installation. These bypass standard security scanning because they are reviewed by the Hermes team before distribution.

Trusted: Skills from automatically recognized trusted organizations — currently OpenAI, Anthropic, Hugging Face, and NVIDIA on GitHub, plus any repositories you manually add via tap with the --trust flag. Trusted skills with non-dangerous scan findings can be installed.

Community: Skills from public registries (skills.sh, ClawHub, LobeHub) or public GitHub repositories not in the trusted list. Community skills face stricter installation policy: caution verdicts block installation unless --force is passed.

Agent-created: Skills created autonomously by Hermes using skill_manage. These have the most restricted update path and are the focus of the curator and write approval gate systems.

Tapping Custom Repositories

You can add any GitHub repository as a skills tap, making its skills discoverable via search and installable:

hermes skills tap https://github.com/your-org/team-skills

Taps are stored in ~/.hermes/hub/taps.json. After adding a tap, running hermes skills search includes the tapped repository in results. You can mark a tapped repository as trusted:

hermes skills tap https://github.com/your-org/team-skills --trust

Trusted taps allow skill installation with the same policy as first-party trusted sources. Use the --trust flag only for repositories you control or have audited.

Keeping Hub Skills Current

Hub-installed skills can be updated when the source repository publishes changes:

hermes skills update

This compares content hashes of all hub-installed skills against their remote source and updates those that have changed. Skills you have modified locally are detected by comparing the local hash against the stored origin hash — if the content differs, Hermes treats the skill as user-modified and skips the update to avoid overwriting your changes.

To reset a bundled skill to its upstream default after local modification:

hermes skills reset skill-name

The --restore flag deletes the local copy and re-copies the bundled version, removing all local modifications. Without --restore, the command clears the modification tracking record, allowing future automatic updates without restoring the bundled content.