The Hermes Skills Hub: Publishing and Installing Community Skills

Writing a skill is easy. Distributing it — so other people's agents can install yours, verify it, version it, and trust it — is the problem most runtimes quietly ignored for the first year of this era. The Hermes Skills Hub is Nous Research's answer. Combined with the 100+ skills that ship in the Hermes bundle, it gives the runtime one of the largest ready-to-use procedural libraries in the ecosystem.

This piece walks through where skills live in Hermes, how installation works, and how the Hub fits into the broader agentskills.io world that also includes aiskill.market.

Key Takeaways

• Hermes ships with 100+ bundled skills across 24+ top-level categories including autonomous-ai-agents, software-development, devops, mlops, red-teaming, writing-communication, data-analytics, productivity, automation, information-processing, and research.
• Installed skills land in ~/.hermes/skills/<skill-name>/SKILL.md — a predictable, inspectable directory.
• The Hermes Skills Hub is the community publishing surface for skills that follow the agentskills.io standard.
• Skills are format-compatible with Claude Code, Cursor, and Codex — minor tag tweaks may be required between runtimes.
• aiskill.market's /hermes channel is intended to index compatible skills over time, giving authors a second discovery surface beyond the Hub.
• Versioning uses semver in frontmatter; updates are opt-in, not automatic.

Where Skills Live

Hermes organizes everything skill-related under a single home directory root. After install, the tree looks roughly like this:

~/.hermes/
├── skills/
│   ├── test-driven-development/
│   │   └── SKILL.md
│   ├── systematic-debugging/
│   │   └── SKILL.md
│   └── ...
├── memory/
│   └── *.md  (FTS5-indexed markdown notes)
└── config.yaml

Every skill is its own directory. The directory name is the skill's name field from frontmatter. The SKILL.md is the only required file. A skill may also ship supporting files (templates, example inputs, reference docs) in the same directory, and those are available to the agent when the skill activates.

This layout is deliberate: a skill is a unit you can mv somewhere else, diff in git, share as a tarball, or paste into a PR review. There is no database to migrate, no binary to rebuild.

The Bundled Library

When you install Hermes via the official script, the bundled skill library comes along:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

The bundle covers 24+ top-level categories. A non-exhaustive tour of what is in there:

• autonomous-ai-agents — patterns for multi-step planning, subagent delegation, self-critique.
• software-development — test-driven-development, writing-plans, systematic-debugging, subagent-driven-development.
• devops — deployment rollback patterns, infra-as-code linting flows, incident comms.
• mlops — dataset versioning, experiment tracking, model-card authoring.
• red-teaming — adversarial prompt construction, jailbreak triage, disclosure workflows.
• writing-communication — stakeholder updates, technical spec authoring, commit-message discipline.
• data-analytics — reproducible notebook authoring, SQL review patterns.
• productivity / automation / information-processing / research — inbox triage, meeting-note structure, citation hygiene, literature synthesis.

The exact total count changes per release. The defining property is not the size; it is that the library is curated and version-controlled alongside the runtime. When you upgrade Hermes, you get curated improvements to the bundled skills — similar in spirit to how Linux distributions ship a curated package set on top of upstream sources.

Publishing to the Hermes Skills Hub

The Hub is a community publishing surface. Authors write a SKILL.md following the agentskills.io standard, push it to a public repository, and register it with the Hub so that other users can discover and install it. The exact URL for the Hub changes as the project evolves — reference it by name and rely on the docs link at the end of this piece.

A minimal publishable skill looks like this:

---
name: conventional-commits
description: Use when creating a git commit to enforce Conventional Commits 1.0.0 style.
version: 1.0.0
author: your-handle
license: MIT
metadata:
  hermes:
    tags: [git, commits, conventions]
    related_skills: [writing-plans]
---

Body of the skill — instructions, examples, counter-examples.

The version field matters. The Hub treats skills as semver-versioned packages. Updating a skill means publishing a new version; users pin to a version they trust and opt in to updates explicitly.

For the underlying format details, see The agentskills.io Standard: Why Hermes, Claude Code, and Cursor All Speak It.

Installing a Community Skill

Installation is a file copy with metadata. Conceptually:

# From a local path
cp -r ./my-skill ~/.hermes/skills/my-skill

# From the Hub (conceptual — exact command per docs)
hermes skill install conventional-commits

After install, Hermes re-indexes the skills directory. On the next hermes chat session, the skill's description field is visible to the model's skill selector, and the skill can activate when the user's prompt matches.

Because the file is plain markdown in your home directory, you can also edit a skill after install — to add a project-specific example, for instance — and the change takes effect immediately. This is very different from a traditional plugin model.

Permissions and Trust

Skills are prompts, not executables. A SKILL.md cannot itself run a shell command. But a skill body can instruct the agent to use a tool, and if the agent has permission to run that tool, it will.

Two practical guardrails follow:

1. Read a skill before you install it. It is a markdown file. cat it. Skim the body. Look for anything that tells the agent to run rm -rf, exfiltrate data, or disable safety checks.
2. Keep tool permissions explicit. Hermes, like Claude Code, has a permissions model for tool execution. A malicious skill can ask the agent to do harmful things; the permission layer is what stops the agent from actually doing them.

For the self-modification dimension — skills that the agent itself writes — see Self-Improving Agents: How Hermes Writes Its Own Skills from Experience.

Relationship to aiskill.market

aiskill.market is an independent marketplace for Claude-compatible skills. It already indexes skills across several sources and runtimes. The direction of travel — publicly signaled — is for the /hermes channel on aiskill.market to index skills that follow the agentskills.io standard and work in Hermes.

For a skill author, that means two surfaces going forward: the Hermes Skills Hub (runtime-native, curated, community-driven) and aiskill.market (cross-runtime marketplace with discovery, ratings, and cross-links to MCP servers and other assets). The same SKILL.md can be published to both.

A Practical Workflow

Here is how a solo author typically works end-to-end:

1. Write the skill locally at ~/.hermes/skills/my-skill/SKILL.md.
2. Test in a real session. Iterate on the description — this is the activation hint — until it fires reliably.
3. Move the skill into a public git repo. Tag a release matching the version field.
4. Register with the Hub. Add a README pointing at the raw SKILL.md for reviewers.
5. Cross-publish to aiskill.market's /hermes channel when available.
6. When you ship a new version, bump semver, publish, and let users opt in.

That is the whole loop. No build step, no submission review queue for minor updates, no container to maintain.

Sources

• Hermes Agent repository — https://github.com/NousResearch/hermes-agent
• Hermes documentation — https://hermes-agent.nousresearch.com/docs/
• aiskill.market — https://aiskill.market
• Series: The agentskills.io Standard: Why Hermes, Claude Code, and Cursor All Speak It
• Series: Self-Improving Agents: How Hermes Writes Its Own Skills from Experience
• Series: Writing Your First Hermes SKILL.md