Hermes Skills Hub: 2026 Guide to 100,000+ Agent Skills

Do not index

The Hermes Skills Hub is the built-in system for finding and installing skills in Hermes Agent, the open-source personal agent from Nous Research. Part marketplace, part package manager: as of September 29, 2026 the hub indexes 101,214 skills across 12 registries, and you can browse, search, and install any of them from the hermes skills CLI or straight from chat with /skills. We host Hermes agents for a living at Agent37, so we work inside this ecosystem every day. This guide covers what the hub actually contains right now, how installs and security scanning really work, and the parts every other write-up glosses over.

What the Skills Hub actually is

A Hermes skill is not a plugin and not a binary. The official docs define skills as "on-demand knowledge documents the agent can load when needed." Each one is a folder holding a SKILL.md file, plus optional scripts, references, and assets. The format follows the agentskills.io open standard, originally created by Anthropic and now shared by Claude Code, OpenClaw, Cursor, and dozens of other agent products.

The Skills Hub is the layer that ties registries of those folders together. The docs describe it as one place to "browse, search, install, and manage skills from online registries, skills.sh, direct well-known skill endpoints, and official optional skills."

One naming trap to clear up early: the Hermes README has a "Skills Hub" community link that points to agentskills.io, the standard's own site. The hub you actually install from lives in the product itself: the catalog page in the official docs, the hermes skills CLI, and the /skills command in chat. This guide is about that second, real one.

Two things happen the moment a skill lands in ~/.hermes/skills/, the directory that is the single source of truth. It joins the agent's skill index, a compact list that costs about 3,000 tokens at session start, with full content loaded only on demand. And it becomes a slash command, so /gif-search funny cats works in the CLI and in any connected chat platform. You can even stack up to five: /github-pr-workflow /test-driven-development fix issue #123 and open a PR loads both skills in order.

What's inside the Hermes Skills Hub in 2026

Most articles still quote skill counts from another era. The old "652 skills across 4 registries" figure is still floating around, and one current top-ranking guide runs a 672 count frozen in May 2026. The live catalog is two orders of magnitude past both. Here is what the index showed on September 29, 2026:

Slice	Count
Built-in skills	58
Official optional skills	152
Community skills	101,004
Registries	12
Categories	219
Total skills	101,214

The twelve registries, largest first: ClawHub (78,993), skills.sh (19,993), GitHub (515), LobeHub (505), browse.sh (474), NVIDIA (382), official optional (152), built-in (58), gstack (54), OpenAI (44), HuggingFace (25), and Anthropic (19). These numbers are finally moving again. The catalog rebuilds twice a day (06:00 and 18:00 UTC), but from July 20 to September 12 every scheduled rebuild silently died: a 15-minute CI timeout kept cancelling the job once the ClawHub crawl outgrew it, and the public counts sat frozen for nearly two months while Nous's own freshness watchdog had an open issue about it. Since the fix, the catalog has been catching up fast: ClawHub jumped by almost 10,000 skills, and a new GitHub registry (515 skills) joined the list while a one-skill stub source dropped out. The shape holds, though: two giant community sources (ClawHub is OpenClaw's marketplace, skills.sh is Vercel's directory), a tail of curated vendor registries, and a small core that Nous Research ships itself. Treat any exact count, including ours, as a dated snapshot.

That core deserves attention. The 58 built-in skills arrive with every Hermes install and cover daily-driver territory: productivity, creative work, GitHub, research, email. (The split moved in September: built-ins went from 82 to 58 while the optional catalog grew from 117 to 152.) The 152 optional skills ship in the repo but stay dormant until you run hermes skills install official/<category>/<skill>. The heavier tooling lives there (MLOps alone has 36 entries), including openclaw-migration, which imports an existing OpenClaw setup into Hermes, memories and skills included.

How installing actually works

From the terminal:

hermes skills search kubernetes
hermes skills inspect openai/skills/k8s
hermes skills install openai/skills/k8s
hermes skills install official/security/1password
hermes skills install https://example.com/SKILL.md --name my-skill
hermes skills update
hermes skills uninstall k8s

Every command has a chat equivalent (/skills search, /skills install, and so on), which matters more than it sounds. You can install a skill from your phone, mid-conversation in Telegram or WhatsApp, and the agent picks it up in its next session. Use /reset when you want it active immediately.

Under the hood, an install is not a blind download. The bundle is quarantined, hashed, and run through a security scanner that checks for data exfiltration, prompt injection, destructive commands, and supply-chain signals. The verdict, source URL, and content hash land in a lock file, with an append-only audit log beside it, and hermes skills audit re-scans everything you have installed whenever you ask.

Every source also carries a trust level: builtin, official, trusted (the OpenAI, Anthropic, HuggingFace, and NVIDIA skills repos), or community, which covers everything else, including all of ClawHub. A --force flag can override caution-level findings. It cannot override a dangerous verdict. That hard stop exists for good reason.

Skills that talk to outside services declare their keys in frontmatter: required_environment_variables reads secrets from ~/.hermes/.env, and required_credential_files handles OAuth and service-account files. Keys stay on your machine, never inside the skill folder, and hermes skills inspect shows what a skill will ask for before you commit.

The security paragraph you should actually read

Koi Security's February 2026 audit examined all 2,857 skills then on ClawHub and found 341 outright malicious; a mid-February update raised that count to 824 as the registry passed 10,700 skills. (Koi has since been absorbed into Palo Alto Networks and its original write-up is offline; the linked coverage survives.) ClawHub responded by adding VirusTotal and ClawScan screening on submissions, and it is still not airtight: Palo Alto's Unit 42 found five malicious skills that slipped past the new blocking between February and May 2026, all since taken down. Those figures describe OpenClaw's marketplace, not the Hermes catalog, but ClawHub is the Skills Hub's single largest source, so the practical rules are the same. Prefer built-in, official, and trusted sources for anything that touches credentials, email, or money. Run hermes skills inspect before you install; a skill is a markdown file, and you can just read it. Read any bundled scripts folder with the same eyes: helper scripts are code the agent may run, not documentation. And leave --force alone unless you have read the finding it is overriding.

Hermes writes its own skills too

The hub is only half the story. Hermes's signature feature is a learning loop: it creates skills from its own experience through the skill_manage tool, which the docs call the agent's procedural memory, and it improves those skills during use. Since v0.18.0 there is also /learn: point it at a URL, a directory, or the workflow you just walked it through, and it distills a reusable skill on the spot.

Two controls keep this sane. A write-approval gate (skills.write_approval) stages every agent-authored change under ~/.hermes/pending/skills/ for human review with /skills approve; it ships disabled, so turn it on if a self-editing agent makes you nervous. And the bundled-skill sync never stomps your work: edit any shipped skill and updates skip it from then on, with hermes skills reset <name> --restore as the undo. A built-in Curator watches usage on top of that, surfacing unmanaged skills and flagging ones you never invoke.

Writing your own Hermes agent skills

You do not have to wait for the agent to learn. A skill is a folder with a SKILL.md file: YAML frontmatter over a markdown body of instructions. Drop it in ~/.hermes/skills/ and it is live at the next session. The rule that trips everyone: the agent decides whether to load a skill by matching your request against the frontmatter description field, not the body. Write the description as a trigger condition ("Use when the user asks about invoices or billing exports"), or your perfectly good skill never fires.

Test it without ceremony: hermes chat --toolsets skills -q "your trigger phrase" shows whether the skill loads. When it holds up, publish it with hermes skills publish (the FAQ below has the exact command) or share it with your team as a tap.

The directories worth bookmarking

Three community projects earn a spot next to the official hub:

awesome-hermes-agent (by 0xNyk): the most-starred independent directory, about 250 entries with production/beta/experimental maturity tags, 5,700+ stars, updated a few times a month. Plainly states it is not an official Nous Research project.

awesome-hermes-skills (by ZeroPointRepo): the longer list at 350+ entries, but about 200 of those just catalog the built-in and optional skills that already ship with Hermes. The roughly 170 community projects are the part tagged by maturity, each linking out to per-project install instructions. Updated a couple of times a month, with automated link checks in between. Run by the TranscriptAPI team, with the affiliation disclosed.

Hermes Atlas (hermesatlas.com): 248 ecosystem repos ranked by live GitHub star counts. The number one skill pack, obra/superpowers, sits at 292,000+ stars.

One to skip: the GitHub repo describing itself as "HermesHub, the Skills Hub for Hermes Agent by Nous Research" is neither by Nous Research nor, anymore, a skills hub. It pivoted to an agent-registry experiment and has not shipped a commit since June 2026. It still shows up high in searches for the hub, which is exactly why this disambiguation earns its place.

Teams get something better than a directory: taps. Point hermes skills tap add <owner/repo> at any GitHub repo full of SKILL.md folders and it becomes a private registry. No server, no sign-up, no release pipeline. Set a GITHUB_TOKEN in your environment or the hub runs into GitHub's 60-requests-per-hour anonymous ceiling; with a token it is 5,000.

Do Claude Code and OpenClaw skills work in Hermes?

Mostly, yes. The SKILL.md core (name, description, markdown body, bundled files) is portable by design across everything that speaks the agentskills.io standard, which is why ClawHub and skills.sh can sit inside the Hermes hub at all. What does not carry over is each host's private extension namespace: Claude Code's execution-control frontmatter, OpenClaw's metadata.openclaw dependency gating, Hermes's own metadata.hermes.* toolset wiring. Agents ignore foreign metadata instead of erroring on it, so a Claude Code skill generally just works, minus the Claude-specific behavior. For a full switch, the optional openclaw-migration skill and hermes claw migrate bring an OpenClaw setup across wholesale, and imported user skills land in ~/.hermes/skills/openclaw-imports/. Already keep a folder of Claude Code skills? Point skills.external_dirs at it (say, ~/.claude/skills) and Hermes indexes those skills in place, no reinstall needed. Newer builds go further: hermes import-agent walks an existing Claude Code or Codex CLI setup and brings it across in one command.

Where skills sit in the broader agent stack, next to gateways, memory layers, and orchestrators, is its own topic; our Paperclip vs Hermes vs OpenClaw breakdown maps it, and our Hermes vs OpenClaw comparison walks through the OpenClaw-to-Hermes switch step by step.

Skills vs MCP servers

The other recurring confusion. An MCP server is a live connection that gives the agent new tools: a database it can query, a browser it can drive. A skill is knowledge about how and when to use the tools it already has. They compose rather than compete: wire up an MCP server in config.yaml, let Hermes work out a reliable call sequence, then run /learn and it distills that sequence into a skill it can repeat. Our MCP vs Claude Skills guide maps the distinction across the wider agent ecosystem.

Skills only pay off on an always-on agent

Here is the operational truth the listicles skip: a skill library is only as useful as the agent attached to it. Chat-side installs, scheduled skills, a Curator that learns from usage, an agent that improves its own skills over weeks: all of it assumes a Hermes gateway that is running when the message arrives. On a laptop, that loop dies at lid close.

There is also a first-party option: Hermes Cloud, Nous's own hosting, out of preview since mid-September, from $0.56 per day for the Medium size, about $17 a month if the agent stays up, before model costs, with a $2 credit minimum to deploy. Hosted agents keep the full /skills chat workflow, so you can search and install hub skills from Telegram or Discord just fine. What you do not get is a terminal or file browser, so the shell-side work in this guide (editing skill folders, wiring taps and .env secrets, testing with hermes chat) still needs a machine you control.

Hermes itself is free, MIT-licensed software, and every skill in the hub is a free markdown file. What you pay for is model API usage and somewhere always-on to run the agent. A $5 VPS works if you want to own Docker, updates, and 2 a.m. restarts yourself; our cheap Hermes hosting guide prices out every route honestly. If you would rather skip the maintenance, Agent37 runs your Hermes agent always-on from $3.99/mo, deploys in one click, and gives you a full web terminal for exactly the hermes skills commands in this guide; there is now a free tier too, one personal agent that sleeps when idle and wakes when you message it. Building a product on agents instead? The Agent37 Cloud API provisions persistent Hermes instances from $4.76/mo, metered per minute, behind a single API call. One honest caveat either way: at these prices support is community Discord on the cheap tiers, with email and chat support starting on Pro.

Hermes Skills Hub FAQ

Is the Hermes Skills Hub free?

Yes. The hub is part of open-source Hermes, and every skill in it installs free. Your only costs are model tokens and wherever the agent runs.

How many skills are in the Hermes Skills Hub?

101,214 as of September 29, 2026: 58 built-in, 152 official optional, and 101,004 community skills across 12 registries. The catalog rebuilds twice daily and is live again after a July-to-September freeze, so treat any exact figure, including this one, as a snapshot.

Can the agent install hub skills by itself?

No. Hub installs are user-driven through the CLI or /skills commands. The agent writes and edits its own skills through skill_manage, but it does not shop the hub autonomously.

How do I publish my own skill?

Push it to a GitHub repo with hermes skills publish skills/my-skill --to github --repo owner/repo, then anyone can add that repo as a tap. Getting into the trusted tier requires a Hermes core pull request; community trust is the default for everything else.

Why isn't my installed skill firing?

Nine times out of ten the frontmatter description is the problem. The agent matches your request against each skill's description field, not its body, so a vague description never triggers. Rewrite it as a condition ("Use when..."), then restart the session or run /reset so the refreshed index picks it up.

Hermes Skills Hub: The Complete 2026 Guide to 100,000+ Skills