How to Install Claude Skills in 2026: Claude.ai, Claude Code, GitHub, and Skill Managers

Installing a Claude Skill is not a one-command operation that works everywhere. The correct method depends on whether you use Claude on the web or desktop, Claude Code, a plugin marketplace, or a third-party cross-agent installer. This guide separates those paths, gives verifiable commands, and shows how to audit a Skill before trusting it.

Last verified: July 16, 2026. Product menus and CLIs change quickly. Commands below were checked against Anthropic, Agent Skills, Vercel Labs, and GitHub documentation linked in the Sources section.

Quick answer: choose the right Claude Skill installation method

Use this table before copying any command.

Your environment	Best starting method	Where the Skill goes
Claude.ai or the Claude Desktop chat app	Upload the complete Skill folder as a ZIP	Your Claude account under Customize > Skills
Claude Code, available in every project	Personal Skill	`~/.claude/skills/<skill-name>/SKILL.md`
Claude Code, shared with one repository	Project Skill	`.claude/skills/<skill-name>/SKILL.md`
Claude Code extension distributed as a package	Plugin marketplace	Claude's plugin cache and configured scope
Claude Code plus Codex, Cursor, or other agents	`npx skills add` or `gh skill install`	Agent-specific project or user directories
Maximum review and version control	Manual GitHub copy or clone	A directory you inspect and place yourself

There is no universal Anthropic command that installs any arbitrary repository slug as a Skill. In particular, claude install installs or migrates Claude Code itself; it is not a generic Skill installer. A page that tells you to run claude install owner-repo, claude skills install <slug>, or claude skill add without linking to a specific tool's current documentation is likely mixing products or repeating an obsolete command.

Before installing: identify which Claude product you use

The word “Claude” now covers several surfaces with different storage and security models.

Claude.ai and the Claude Desktop chat app

Claude's consumer chat experience manages Skills through Customize > Skills. A custom Skill is uploaded as a ZIP archive containing the Skill folder. You enable or disable it in the interface. Free, Pro, and Max users need Code execution and file creation enabled; Team and Enterprise availability can also depend on organization settings.

This path does not mean copying files into ~/.claude/skills. That filesystem path belongs to Claude Code.

Claude Code CLI and Claude Code Desktop

Claude Code discovers local Skills from personal or project directories and can also receive Skills through plugins. A local Skill is a folder whose required entry point is named exactly SKILL.md. Claude loads the name and description for discovery, then reads the full instructions when the Skill is relevant or explicitly invoked.

Claude Code Desktop is the graphical interface for Claude Code projects. It can use the same project Skills and installed plugins as Claude Code. It should not be confused with the general Claude Desktop chat app, where custom Skills are uploaded through the account interface.

Claude API

The Claude API supports prebuilt and custom Skills through the code execution environment, but that is an API upload and container workflow rather than a local installation. This guide focuses on end-user and developer installation. API users should follow Anthropic's current Skills API documentation instead of copying Claude Code paths.

Skills, plugins, MCP servers, and project instructions are different

Extension	Primary job	Can contain executable behavior?
Skill	Reusable instructions, domain knowledge, and workflows loaded on demand	Yes. A Skill may bundle scripts and may direct Claude to use tools
Plugin	Distributable Claude Code package containing Skills, agents, hooks, MCP, or LSP configuration	Yes; treat it as highly trusted software
MCP server	Gives Claude structured access to external tools or data	Yes; it runs as a service or process
CLAUDE.md	Standing project context loaded broadly	It can influence tool use but is not an installable Skill package

A Skill is therefore not necessarily “just a harmless Markdown prompt.” The open Agent Skills specification permits scripts, reference files, and assets alongside SKILL.md.

Method 1: upload a custom Skill to Claude.ai or Claude Desktop

This is the simplest path for people who do not use a terminal.

1. Inspect the Skill before packaging it

Open the downloaded folder. At minimum it should contain a readable SKILL.md with YAML frontmatter and instructions. A complete Skill may also contain scripts/, references/, assets/, templates, or examples. Review those files before upload; do not assume a ZIP is safe because it came from a directory site.

The basic layout is:

my-skill/
├── SKILL.md
├── scripts/        # optional
├── references/     # optional
└── assets/         # optional

The frontmatter should at least identify the Skill and explain when it applies:

---
name: my-skill
description: Performs a specific workflow. Use when the user asks for...
---

2. Create the ZIP correctly

Compress the complete my-skill folder, not a random parent directory containing several unrelated repositories. Keep supporting files next to SKILL.md with their relative paths intact. Avoid adding secrets, environment files, local caches, or private documents to the archive.

3. Upload and enable the Skill

Open Claude and enable Code execution and file creation if your plan requires it.
Go to Customize > Skills.
Select the plus button, then Create skill and Upload a skill.
Choose the ZIP file.
Read any validation error rather than repeatedly repackaging files at random.
After the Skill appears, make sure its toggle is enabled.

Anthropic documents this workflow in Use skills in Claude.

4. Verify, update, or remove the uploaded Skill

Ask Claude to use the Skill by name for a harmless sample task. Then try a natural request that should match its description. Confirm that the output follows the promised workflow instead of merely producing a plausible generic answer.

To update a personal uploaded Skill, inspect and upload the revised package according to the current Claude interface. To stop using it, toggle it off. To remove it, open the Skill details, use the menu, and choose Delete. Team and Enterprise owners can provision organization Skills; shared Skills follow organization policies and may update automatically for recipients.

Method 2: install a Skill manually in Claude Code

Manual installation is transparent and requires no third-party package manager. It is the best method when you want to review every file and control exactly where it lives.

Choose personal or project scope

A personal Skill is available across your projects:

~/.claude/skills/my-skill/SKILL.md

A project Skill is available in one repository and can be committed for teammates:

.claude/skills/my-skill/SKILL.md

Use personal scope for preferences and workflows that belong to you. Use project scope for repository-specific build, testing, review, or release procedures that the team should version together. Do not place organization secrets inside either Skill.

Copy a reviewed Skill from GitHub

First clone or download the source into a temporary location. Inspect it before placing anything in Claude's discovery path:

git clone https://github.com/OWNER/REPOSITORY.git /tmp/review-skill
find /tmp/review-skill -maxdepth 3 -type f

After identifying the exact Skill directory and reviewing its contents, copy it into the desired scope:

mkdir -p ~/.claude/skills/my-skill
cp -R /tmp/review-skill/path/to/my-skill/. ~/.claude/skills/my-skill/

For project scope, replace the destination with .claude/skills/my-skill/. Copy the whole Skill directory, not only SKILL.md, when the instructions reference scripts, templates, or other resources.

Verify manual installation

Check the exact filename and path:

test -f ~/.claude/skills/my-skill/SKILL.md && echo "Skill entry point found"

In Claude Code, invoke the Skill directly if it is user-invocable, for example /my-skill, or ask a request that closely matches its description. Claude Code watches existing Skill directories for text changes, so adding or editing a Skill normally takes effect in the current session. A restart can be necessary when the top-level Skills directory did not exist when the session began. Plugin component changes have separate reload rules.

The authoritative paths, precedence rules, and live detection behavior are documented in Extend Claude with skills.

Method 3: install with npx skills add

Vercel Labs' skills CLI is a third-party, open-source installer for the broader Agent Skills ecosystem. It supports Claude Code and many other agents. It is not an Anthropic command.

Node.js and npm must be available because npx downloads and runs the CLI package. For an unfamiliar repository, list its Skills before installing:

npx skills add vercel-labs/agent-skills --list

Install interactively from a repository:

npx skills add vercel-labs/agent-skills

Install one named Skill specifically for Claude Code:

npx skills add vercel-labs/agent-skills   --skill frontend-design   --agent claude-code

Add --global or -g only when you want user scope across projects. Project scope is safer for team workflows because the installed configuration stays tied to the repository. The CLI can install by copy or symlink; symlinks provide one canonical copy, while copies are easier to inspect and isolate.

Useful lifecycle commands include:

npx skills list
npx skills find typescript
npx skills update
npx skills remove frontend-design

Avoid --yes on the first installation from an unfamiliar source. Skipping every prompt saves seconds but removes a useful chance to notice the selected Skills, agents, and scope. Also remember that running npx executes package code on your machine; review the package identity and repository before use.

Method 4: install with GitHub CLI gh skill install

GitHub CLI now provides a first-party GitHub command for Agent Skills. It is useful when the source is a GitHub repository and you want source tracking or a pinned Git reference. It supports Claude Code but defaults can target another agent, so specify your intent.

Install a particular Skill for Claude Code at user scope:

gh skill install OWNER/REPOSITORY SKILL_NAME   --agent claude-code   --scope user

Omit --scope user for project scope. To install every discovered Skill, use --all only after reviewing the repository. For reproducibility, pin a tag or commit:

gh skill install OWNER/REPOSITORY SKILL_NAME   --agent claude-code   --scope project   --pin COMMIT_OR_TAG

The command injects source tracking metadata so gh skill update can detect upstream changes. That metadata is valuable for auditing, but an update still deserves review: compare the new revision before accepting changed instructions or scripts.

See the current syntax and supported agents in the GitHub CLI gh skill install manual.

Method 5: install Skills through Claude Plugin Marketplaces

Plugins are the native distribution system when a Claude Code extension needs more than a standalone Skill. A plugin may include multiple Skills plus agents, hooks, MCP servers, LSP configuration, or other executable components.

The official Anthropic marketplace is available in current Claude Code versions. Open /plugin, browse the Discover tab, inspect a plugin, and choose its scope. A direct installation looks like:

/plugin install github@claude-plugins-official

For Anthropic's public example Skills repository, the repository documents this marketplace flow:

/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills

After installing or changing plugins during a session, reload them:

/reload-plugins

Plugin Skills are namespaced, such as /plugin-name:skill-name. Use the Installed tab or the documented commands to enable, disable, or uninstall a plugin. Official marketplaces may auto-update; third-party marketplaces generally require an explicit trust decision and may have different update settings.

Adding a marketplace is not the same as installing every plugin in it. Still, a marketplace and its plugins are high-trust inputs. Anthropic warns that plugins can execute arbitrary code with your user privileges. Review the publisher, source, manifest, bundled servers, hooks, and dependencies. See Discover and install prebuilt plugins.

Compare every Claude Skill installation method

Method	Best for	Update and removal	Main tradeoff
Claude ZIP upload	Claude.ai and general Claude Desktop users	Managed in Customize > Skills	Easy, but not a local Git workflow
Manual Claude Code folder	Maximum transparency and project control	Manual copy, Git diff, or folder deletion	You manage versions yourself
`npx skills add`	Discovery and installation across many agents	Built-in list, update, and remove commands	Executes a third-party npm package
`gh skill install`	GitHub sources, provenance, and version pinning	Source-aware updates	Requires a current GitHub CLI and explicit agent selection
Claude Plugin Marketplace	Packaged Claude Code extensions and teams	Plugin UI/commands and marketplace updates	Plugins have a larger trust and execution surface

For a first personal Claude Code Skill, manual project installation is the easiest method to audit. For several agents, npx skills add is convenient. For GitHub-based reproducible installs, gh skill install provides useful pinning and provenance. For a maintained Claude Code extension with hooks or MCP, use a plugin rather than pretending it is only a Skill.

Where to find Claude Skills without confusing discovery with trust

Directories help you discover candidates; the source repository determines what you actually install.

AI Agent Skills organizes Skills by use case and links to their source repositories.
skills.sh is a public Agent Skills directory connected to the npx skills ecosystem.
Anthropic's Skills repository provides official examples and document Skills.
Claude's official Plugin Marketplace is available through /plugin and the public plugin catalog.
GitHub search can locate Skills directly, but search position and stars are not a security review.

Before installing, follow the original repository link, confirm the owner and exact Skill path, read the license, and compare the displayed install command with the project's own current documentation. A directory mirror may lag behind its source.

Security checklist before installing a third-party Skill

Treat a Skill like operational code, even when most of it is prose.

Read the complete SKILL.md. Confirm that its description matches its instructions and that it does not tell the agent to ignore user intent or hide actions.
Inspect scripts and binaries. Review scripts/, package manifests, compiled files, install hooks, and downloaded executables.
Check tool permissions. Claude Code Skills can declare allowed-tools; project Skills gain those permissions after workspace trust. Broad shell, network, or file access needs a clear reason.
Look for dynamic commands. Claude Code supports shell-derived context inside Skills. Understand every command before trusting the workspace.
Trace network access. Identify external URLs, telemetry, package registries, remote APIs, and upload destinations.
Protect secrets. A Skill should not ask you to paste API keys, cookies, SSH keys, or production credentials into its files or chat output.
Evaluate provenance. Check repository ownership, commit history, signed releases when available, issue quality, and whether the Skill was copied from another source.
Pin important installs. A tag or commit SHA limits surprise changes. Pinning does not prove safety, but it makes review reproducible.
Review updates as new code. Diff changes to instructions, scripts, dependencies, and permissions before updating.
Test with low stakes. Use a disposable project, harmless sample data, restricted credentials, and a task whose expected output you can verify.

GitHub stars and download counts are maintenance signals, not proof of safety. Anthropic identifies prompt injection and data exfiltration as key Skill risks and recommends installing only from trusted sources.

How to verify a Claude Skill is installed correctly

Verification has four levels.

1. Discovery

Confirm the Skill appears in Claude's Skills interface, Claude Code's slash-command menu when user-invocable, or your installer's list. Also confirm the expected scope: project, personal, plugin, or organization.

2. Explicit invocation

Call the Skill by name when supported, or say “Use the [name] Skill to…” with a safe sample. This tests loading without relying on automatic matching.

3. Automatic triggering

Start a fresh request that matches the description without naming the Skill. If Claude never chooses it, improve or inspect the description. If it triggers constantly, narrow the use conditions.

4. Behavioral correctness

Check observable requirements: Did it use the expected template? Did it run only documented tools? Did it create files in the promised location? Did it stop for approval before a destructive action? A generic-looking answer is not evidence that the Skill ran correctly.

Troubleshooting Claude Skills

Problem	Likely cause	What to check
Customize > Skills is missing	Code execution or organization Skills are disabled	Account capabilities and admin settings
ZIP upload fails	Wrong archive root, missing `SKILL.md`, invalid name/frontmatter, or size limit	Unzip locally and inspect the exact tree
Claude Code cannot find the Skill	Wrong scope, filename case, or directory nesting	Exact path ending in `<skill-name>/SKILL.md`
Skill exists but never triggers	Description does not clearly say when to use it	Test explicit invocation, then rewrite description
Skill triggers too often	Description is overly broad	Add concrete situations and exclusions
Change is not detected	Top-level directory was created after session start	Restart once; existing directories normally update live
Plugin Skill is missing	Plugin needs reload, has a load error, or uses a namespace	Run `/reload-plugins` and inspect the Errors tab
Installer targeted the wrong agent	Default agent or scope was assumed	Reinstall with explicit `--agent claude-code` and scope
Team gets different behavior	Personal Skill overrides project Skill, or versions differ	Check precedence, source, and pinned revision
Windows path or permission error	Shell, WSL, home directory, or filesystem permissions differ	Confirm the actual home path and write access

Do not “fix” discovery by copying the same Skill into every possible directory. Duplicates create precedence conflicts and make updates harder to audit.

For a manually installed personal Skill, update it by reviewing and replacing its files; uninstall it by deleting its exact folder. For a project Skill, use Git so changes are reviewed and shared like other repository configuration.

For the Vercel Labs CLI, use npx skills update and npx skills remove <name>. For GitHub CLI, use the gh skill update and list commands documented for your installed version. For plugins, use /plugin or the plugin enable, disable, and uninstall commands. For Claude.ai, manage the Skill under Customize > Skills.

Team and Enterprise sharing should use organization provisioning, shared Skills, project Skills, or controlled plugin marketplaces rather than emailing untracked ZIP files. Keep ownership, version, review date, and rollback instructions with every business-critical Skill.

Frequently asked questions

What is the correct command to install a Claude Skill?

There is no single command for every Claude surface. Claude.ai uses ZIP upload. Claude Code accepts files in ~/.claude/skills or .claude/skills, plugins, and compatible third-party installers. Choose the method based on the product and desired scope.

Does claude install install Skills?

No. Anthropic documents claude install as a Claude Code installation or migration command. Do not append a repository slug and assume it will install a Skill.

Where are Claude Code Skills stored?

Personal Skills live at ~/.claude/skills/<skill-name>/SKILL.md; project Skills live at .claude/skills/<skill-name>/SKILL.md. Plugin Skills live inside the plugin package and use a namespace.

Can I install a Claude Skill from GitHub?

Yes. Review and copy the Skill directory manually, use npx skills add, use gh skill install, or install a repository that is packaged as a Claude plugin marketplace. The repository structure determines which methods apply.

Do Claude Skills work in Claude Desktop?

Custom Skills can be uploaded through Claude's Customize > Skills interface in supported Claude plans and settings. Claude Code Desktop is a different developer surface and uses Claude Code project Skills and plugins.

Do I need to restart Claude Code after installing a Skill?

Usually not when adding or editing a Skill inside an existing watched directory. Restart when the top-level Skills directory was created after the session began or when troubleshooting an older version. Plugins can be refreshed with /reload-plugins.

Can one Skill work in Claude Code, Codex, and Cursor?

Basic Agent Skills follow an open format and can often work across compatible agents. Agent-specific fields, tool names, hooks, dynamic commands, and permission models may not be portable. Test each target rather than assuming full compatibility.

Are third-party Claude Skills safe?

Not automatically. A Skill can influence tool use, execute bundled scripts, install dependencies, or access network resources. Review the source, permissions, scripts, dependencies, and update path, then test with low-risk data.

What is the difference between a Skill and a plugin?

A Skill is one on-demand workflow package. A Claude Code plugin is a distribution container that can bundle multiple Skills and more powerful components such as hooks, agents, MCP servers, and LSP configuration.

How do I remove a Claude Skill completely?

Remove it through Customize > Skills, delete the exact local Skill directory, run the installer's remove command, or uninstall the containing plugin. Then verify it no longer appears at any higher-precedence scope.

Sources and next step

Primary references used for this guide:

Once you know your Claude surface and installation method, browse AI Agent Skills, open the original repository for any candidate, complete the security review, and install only the specific Skill you intend to use.