What is the difference between a plugin, a skill, and a connector?

A skill is knowledge: instructions Claude loads for a kind of work. A connector is access: a live MCP link to an external product, with OAuth and permissions. A plugin is packaging: it can carry skills, MCP configs, hooks, and agents together, installed as one unit. They compose; SiteGPT ships all three, and the plugin is how the skill installs in Claude Code.

How do I install a Claude Code plugin?

Two commands: /plugin marketplace add / registers a marketplace (any git repo with a .claude-plugin/marketplace.json), then /plugin install @ installs the plugin from it. For SiteGPT: /plugin marketplace add sitegpt/agent-skills, then /plugin install sitegpt@sitegpt.

Do plugins work in Claude Cowork?

Yes. Plugins install into Claude Code and Cowork, and the Anthropic Plugin Directory serves both. The same SiteGPT plugin that gives Claude Code the sitegpt-cli skill gives a Cowork session the same workflow knowledge.

What does the SiteGPT plugin do?

It ships the sitegpt-cli skill: the full playbook for building and running AI customer support from Claude Code. Ask for a support chatbot for any website URL and Claude Code builds and trains one with no SiteGPT account, hands back a preview-and-claim link, and from then on manages knowledge, conversations, leads, and settings through the CLI.

Do I need the plugin if I already use the SiteGPT connector?

They do different jobs and pair well. The connector gives Claude live MCP access to your account with OAuth and read/write separation. The plugin adds the workflow knowledge, including the no-signup onboarding lane the connector cannot offer, since MCP access starts at sign-in. Install both: the skill knows the plays, the connector provides the hands.

Can I build my own plugin?

Yes, and the format is small: a git repo with your components in default locations (skills/, agents/, hooks/) and an optional .claude-plugin/plugin.json manifest. Add a marketplace.json and the repo doubles as its own marketplace. The official plugin reference at code.claude.com/docs covers every field.

The plugins guide

Claude plugins, explained properly

Plugins are how Claude Code and Cowork gain new capabilities in one install: skills, tools, and configuration, packaged together. Here is how they work, how they differ from skills and connectors, and how to judge one before you install it.

claude code

> /plugin marketplace add sitegpt/agent-skills  ✓ marketplace added: sitegpt (1 plugin)> /plugin install sitegpt@sitegpt  ✓ installed: SiteGPT v0.2.1 (1 skill) > Add a support chatbot to acme.com  loading skill: sitegpt-cli  ✓ chatbot built and trained, claim link ready

Agents read this page too

First principles

What are Claude plugins?

A plugin is a self-contained package that extends Claude Code or Claude Cowork. Under the hood it is just a directory, usually a git repo, whose contents Claude discovers on install: skills in skills/, agents, hooks, MCP server configs, and an optional manifest describing the bundle.

The point of the packaging is the single step. Before plugins, a team assembling a capable Claude Code setup copied skills around, hand-edited MCP configs, and hoped everyone stayed in sync. A plugin moves all of that into one install command, versioned by its repo, updated like software instead of folklore.

Distribution runs through marketplaces: any git repo with a .claude-plugin/marketplace.json can serve plugins, and Anthropic curates a reviewed Plugin Directory inside Claude for discovery. The directory is young, which is exactly why the plugins worth installing today are mostly vendor-built ones for products you already use.

The trio, untangled

Plugins vs skills vs connectors

The three names get used interchangeably and mean different things. The clean mental model: skills are knowledge, connectors are access, plugins are packaging. They compose rather than compete.

	Plugin	Skill	Connector
What it is	Packaging: a bundle of capabilities installed as one unit	Knowledge: instructions Claude loads for a kind of work	Access: a live MCP link to an external product
Where it lives	A git repo, installed via /plugin	A SKILL.md, often inside a plugin	A server the vendor runs; you approve a grant
Auth involved	None to install	None	OAuth approval, revocable
SiteGPT ships it as	sitegpt/agent-skills, carrying the CLI skill	sitegpt-cli, the full support workflow	The directory-listed MCP connector

The practical pairing for a real product: install the vendor’s plugin for the workflow knowledge, connect their MCP server for live account access. The skill knows the plays; the connector provides the hands.

Mechanics

How marketplaces and installs work

Add the marketplace

/plugin marketplace add sitegpt/agent-skills registers a marketplace from any git repo carrying a marketplace.json. Anthropic’s reviewed directory needs no adding; it is built into Claude.

Install the plugin

/plugin install sitegpt@sitegpt pulls the plugin and wires up everything it ships. Skills appear as namespaced shortcuts (the SiteGPT skill answers to /sitegpt:sitegpt-cli) and load automatically when the work calls for them.

Update like software

Plugins version with their repo. Updating pulls the new revision; uninstalling removes everything the plugin brought, cleanly.

The same mechanism serves Claude Cowork, so a plugin your engineering side installs for Claude Code is one install away for the operators working in Cowork.

Choosing well

What makes a plugin worth installing

Best-plugins lists age badly in a directory this young. Criteria hold up better; apply these three to anything you consider, including ours:

Provenance you can read

Prefer plugins from the vendor of the product they drive, with a public repo you can open. A plugin is code and instructions entering your sessions; you should be able to read every line before it does.

A footprint that matches the job

A focused plugin ships the skills and configs its job needs and nothing else. Be wary of bundles carrying hooks and agents you did not ask for; every component is surface area.

Honest authentication

If the plugin reaches a live product, look at how. Browser OAuth, or no credentials at all, beats a pasted API key in config every time. The strongest version: a plugin that does something real before you even have an account.

A worked example

The SiteGPT plugin, against those criteria

Our plugin is the criteria applied to ourselves. Provenance: it is the official SiteGPT package, one public repo (sitegpt/agent-skills) that also serves the Agent Skills standard, so every line is readable before install. Footprint: one skill, sitegpt-cli, and nothing else; no hooks, no agents, no surprise configs.

Authentication is where it gets unusual: the skill starts working before you have an account. Ask Claude Code for a support chatbot for any website URL and it builds and trains one through SiteGPT’s no-signup onboarding, then hands back a preview-and-claim link. Signing in only becomes necessary when you want to keep it. From there the same skill runs the whole operation: knowledge sync, conversations, leads, settings, reports.

Pair it with the SiteGPT connector from the Claude directory and both halves are covered: the plugin teaches the workflow, the connector provides live MCP access with OAuth and read/write separation. The full Claude Code walkthrough lives on the Claude Code page.

Claude plugins, answered

What are Claude plugins?: Plugins are installable packages that extend Claude Code and Claude Cowork with new capabilities. A plugin is a directory, usually a git repo, carrying any mix of skills (workflow knowledge Claude loads when the work calls for it), agents, hooks, and MCP server configs. Install one and Claude gains everything inside it in a single step.
What is the difference between a plugin, a skill, and a connector?: A skill is knowledge: instructions Claude loads for a kind of work. A connector is access: a live MCP link to an external product, with OAuth and permissions. A plugin is packaging: it can carry skills, MCP configs, hooks, and agents together, installed as one unit. They compose; SiteGPT ships all three, and the plugin is how the skill installs in Claude Code.
How do I install a Claude Code plugin?: Two commands: /plugin marketplace add <owner>/<repo> registers a marketplace (any git repo with a .claude-plugin/marketplace.json), then /plugin install <name>@<marketplace> installs the plugin from it. For SiteGPT: /plugin marketplace add sitegpt/agent-skills, then /plugin install sitegpt@sitegpt.
Do plugins work in Claude Cowork?: Yes. Plugins install into Claude Code and Cowork, and the Anthropic Plugin Directory serves both. The same SiteGPT plugin that gives Claude Code the sitegpt-cli skill gives a Cowork session the same workflow knowledge.
What makes a plugin worth installing?: Three checks. Provenance: is it from the vendor of the product it drives, and is the repo public? Footprint: does it carry focused skills and honest tool configs rather than a grab bag of hooks you did not ask for? Authentication: if it reaches a live product, does it use real auth (OAuth, or no credentials at all) instead of asking you to paste an API key into config? A vendor plugin that starts working before you even have an account is the strongest version of that test.
What does the SiteGPT plugin do?: It ships the sitegpt-cli skill: the full playbook for building and running AI customer support from Claude Code. Ask for a support chatbot for any website URL and Claude Code builds and trains one with no SiteGPT account, hands back a preview-and-claim link, and from then on manages knowledge, conversations, leads, and settings through the CLI.
Do I need the plugin if I already use the SiteGPT connector?: They do different jobs and pair well. The connector gives Claude live MCP access to your account with OAuth and read/write separation. The plugin adds the workflow knowledge, including the no-signup onboarding lane the connector cannot offer, since MCP access starts at sign-in. Install both: the skill knows the plays, the connector provides the hands.
Can I build my own plugin?: Yes, and the format is small: a git repo with your components in default locations (skills/, agents/, hooks/) and an optional .claude-plugin/plugin.json manifest. Add a marketplace.json and the repo doubles as its own marketplace. The official plugin reference at code.claude.com/docs covers every field.

Two commands from a support agent

Install the plugin and ask Claude Code for a support chatbot. No account needed to see it work.

1/plugin marketplace add sitegpt/agent-skills2/plugin install sitegpt@sitegpt

SiteGPT for Claude Code →The MCP server →