When Hermes Agent opened to the public on March 12, the release shipped with more than seventy bundled skills across fifteen-plus categories. Four weeks later, a community skills marketplace was live at agentskills.io with hundreds more. This post is about how that happened — and why the Hermes skill architecture is worth understanding even if you never write one.
What a skill actually is
In most AI agent frameworks, a "tool" is a Python function that the developer registers with the agent at import time. You decorate a function, write a docstring, list it in a tools array somewhere. When the agent needs it, the framework injects its description into the prompt and parses the model's tool-call output.
A Hermes skill is not that. It is a declarative bundle with a manifest (skill.yaml), a set of executable scripts or Python entrypoints, a description of its prerequisites, and an activation policy. When Hermes starts, it walks the skill directory, reads every manifest, and decides which skills are available for this session based on whether their prerequisites are actually present on the machine — environment variables, binaries on PATH, config file entries, platform capabilities.
A skill that needs ffmpeg will not be advertised to the agent on a machine that does not have ffmpeg installed. A skill that needs a Telegram bot token will only activate in a session connected to Telegram. The agent's prompt only ever carries the skills you could actually use right now.
This sounds minor. It is the reason an agent can ship with seventy skills on day one without melting the prompt.
The Skills Hub and what it does
Bundled with Hermes from v0.2.0 onwards is a thing called the Skills Hub — a local index of every skill available to this installation, with metadata, source attribution, and a conditional-activation layer. You can run hermes skills list to see what is installed, hermes skills enable <name> or hermes skills disable <name> to turn things on and off, and hermes skills info <name> to see the manifest, the source, and the exact prerequisites.
The Hub is also the abstraction that community skills plug into. A skill author publishes a manifest + source package; Hermes users install it with a single command; the Hub handles prerequisite validation and activation rules the same way it does for bundled skills. There is no special path for "official" versus "community" skills — they are all just skills.
Four weeks after launch, that Hub grew into its public face: agentskills.io. It is a web directory of community-contributed skills with search, categories, popularity metrics, and a standardized install command. Think of it as the npm or pip of agent capabilities, except narrower: each skill is supposed to do one thing, and it ships with a manifest that tells Hermes how to use it safely.
Why the ecosystem grew fast
Four weeks is not a long time to go from zero to hundreds of community skills. A few things in the architecture made that possible.
The manifest is the interface. A skill author only has to get the manifest right. Prerequisites, description, input schema, activation conditions — those go in skill.yaml. The implementation can be Python, a shell script, an executable binary, anything the manifest points to. Contributors do not have to learn a Hermes-specific SDK to write a useful skill. They have to describe what their tool does in a YAML file that already existed for bundled skills.
Prerequisites are structural, not documented. If a skill needs ffmpeg, the manifest says so, and Hermes checks it. You do not install a broken skill that silently fails at runtime because of a missing dependency — the Hub will tell you what is missing and refuse to activate. That means skill authors can assume, and skill users can trust.
Activation is conditional. A skill can declare that it only activates on Telegram, only when a specific env var is set, only when a particular file exists, only in certain working directories. The prompt the agent sees is tailored to the moment. You can install a hundred skills and the agent will never see ninety of them in any given session, which means the prompt does not bloat.
Sandbox-friendly by default. Skills that execute code do it inside Hermes' existing sandbox layers — the git worktree isolation and filesystem checkpoint system from the v0.2.0 launch. A community skill that goes rogue cannot actually destroy your code or your files, because the sandbox is baked into the agent, not into the skill. This drastically lowers the trust cost of installing something new.
What a community skill typically looks like
Community skills from the first four weeks cluster in a few useful patterns:
- •Wrappers around existing CLI tools. Someone takes
ffmpeg, orpandoc, orimagemagick, and writes a thin skill that exposes the common operations — trim video, convert document, resize image. These are cheap to write and immediately useful. - •Integrations with personal services. Skills that talk to a user's Notion account, Obsidian vault, Home Assistant instance, RSS reader, Pocket account, or personal finance app. Most of these are less than 200 lines of Python plus a manifest.
- •Domain-specific helpers. Skills tailored to specific workflows — a paper-reading skill for researchers, a git-log-summarizing skill for engineering leads, a menu-generator for home cooks, a dungeon master assistant for tabletop RPG groups.
- •Model-specific shims. Skills that add provider-specific features (prompt caching headers, reasoning budgets, fine-tuning hooks) as formal agent tools instead of requiring custom client code.
None of these individually is a killer feature. In aggregate, they are the reason Hermes stopped feeling like "a thing Nous Research built" and started feeling like a community.
The interesting consequence
The day agentskills.io crossed a hundred community skills, something subtle changed about the project. It was no longer possible to describe what Hermes Agent is by listing its features, because the feature set had become unbounded. What you can describe is the shape of the thing: a singular agent, talking to any chat platform you like, pulling in skills on demand, under a uniform safety model.
That is not a feature list. That is a platform. And it is the reason the skills ecosystem matters more than any individual capability that ships inside it.