Install
npx skills-atlasWho it is for
You want to look across the public Agent Skills first and then decide which ones to install into your own agent; that is the situation Skills-Atlas fits. It collects public skills in SKILL.md format into a collapsible terminal directory tree organized by topic, and it also provides a local web interface as a second way in: you can select several skills and install them in one go, instead of hunting down a separate repository and running a separate install command for each skill.
If your goal is just to install one skill, the install method that ships with that skill's own repository is faster, and you do not need to browse the catalog first. The other way around: if you want to roll out one consistent set of skills across several agents, what you need is a list you can browse offline plus an install command with fixed parameters, and then "browse first, then install in bulk" is the core action of this tool.
The bulk install example in the documentation fixes a single target agent and a fixed scope, so it fits people who already use one specific agent and are willing to work through the flow the example shows: "specify agent + specify scope + install several at once".
First attempt in 3 minutes
The prerequisite is Node ≥ 20: the documentation states only the version requirement and does not say whether an unmet version is refused outright at startup or comes with a fallback warning, so the only thing that can be asserted with certainty is that when the version is not met, running is not guaranteed.
The command calls skills-atlas through npx, with no global install; the documentation does not describe what this command writes to disk on the machine or how it handles caching, so do not treat it as an action that writes absolutely nothing.
The first step is to browse and inspect:
npx skills-atlas
Inside the terminal directory tree, expand the topic down to the specific skill, press p to view that skill's SKILL.md, and read the risk check it offers at the same time: the risk check lists what the skill will execute, what it will read, and what it will send outward.
The success signals are: the terminal directory tree lists the SKILL.md format skills by topic, viewing any entry lets you read that skill's description and risk notices, and when a skill would replace an already installed skill of the same name, that entry is marked before anything runs.
There is one thing the documentation does not account for: whether pressing p fetches content over the network, and whether it writes to disk on the machine, is not stated; the local web interface entry point by itself means a local service is listening on the machine. So treat the viewing step as an informational action, but do not assume it produces no access at all.
The action that clearly changes the environment is the installation; as for whether the npx invocation itself writes to disk, the documentation does not say. The install form in the documentation is:
skills-atlas install anthropics/skills@pdf supabase/agent-skills@supabase -a claude-code -p -y
Two skills are installed in one go, with -a specifying the target agent, -p specifying the project scope, and -y meaning automatic confirmation. Same-name replacements are marked before the install, so it is worth checking that entry before running the install; if you want to confirm one by one, do not use the command with -y directly.
Use cases
Case one, bulk installing into the same agent by topic. The copyable command given in the documentation is the one above: it installs two skills from two different repositories in one go, the target agent is claude-code, the scope is at the project level, and confirmation is automatic. The observable result is that those two skills appear in that agent's skill directory. Evidence boundary: the documentation gives only this set of skill names, this one agent, and these three parameters; when you switch to other repositories or agents, which parameters carry over as they are and which have to change is not stated.
Case two, looking at a skill's description and risk notices before installing it. After locating a skill in the tree and pressing p, you read its SKILL.md, and the risk check tells you what it will execute, what it will read, and what it will send outward; if it would replace an already installed skill of the same name, that entry is marked before anything runs. The observable result is that you have the description and the notices in hand before installing. Evidence boundary: the documentation does not say whether the risk check results come from a snapshot bundled in the package or are fetched live, and it does not describe the network and disk writing behavior when opening an entry with p.
Limits of use
The directory snapshot is bundled with the package, so browsing the directory can be done offline; but whether the SKILL.md body text and the risk check results are equally available locally is not accounted for in the documentation, so treat it as unknown.
Installing after a multi-selection writes into the target agent's skill directory (the documentation example uses the project scope) and may replace a skill of the same name, although same-name replacements are marked before the install.
Broad coverage does not mean that every client behaves the same: the documentation gives the set of agents that can be covered and two interface entry points (terminal tree, local web interface), without describing their behavioral differences one by one.
Node ≥ 20 is required; when the version is not met, running is not guaranteed, and the documentation does not state the exact fallback or refusal-to-start behavior.