SSD Nodes Learn Hosting plans →
How to do am Matt ConnorBy Matt Connor

Claude Code plugins: wetin dem be and how much e cost

Learn wetin Claude Code plugin be, where e dey live, how to install am, and why the mechanism cost nothing while tokens pay for extra context.

Claude Code plugin ki tonu

Claude Code plugin na one directory wey get components wey Claude Code dey load and manage as one unit. The main components na skills, agents, hooks, MCP servers, LSP servers, and background monitors. Plugin fit also carry commands, output styles, themes, workflows, and one bin/ directory wey get executables. When you install plugin, e add all im parts at once under one name. When you disable am, e remove dem the same way.

Plugin mostly no give agent any ability wey e no already get. You fit write skills, agents, hooks, and MCP servers by hand inside .claude/ directory. The exceptions na language servers for code intelligence and mods, because Claude Code only dey load dem from plugin. For every other thing, plugin na packaging layer. E give you way to version those parts, share dem with fifteen people, and update dem later without asking everybody to copy files. Na the whole idea be that. Most confusion about plugins dey happen because people expect dem to be new kind capability.

The optional manifest for .claude-plugin/plugin.json dey specify the plugin name, and that name become namespace. You invoke skill for plugin wey dem call commit-commands as /commit-commands:commit. So two plugins fit each release skill wey dem call commit without one hiding the other. Plugin agents get the same scope for the @-mention list, as plugin-name:agent-name.

Plugin, skill, MCP server, or rules file

People dey use these four words as if dem dey compete. Dem no dey compete, and e good make we explain the boundary once.

  • A skill na one instruction unit wey Claude loads when task need am. See wetin Agent Skill really be.
  • An MCP server na separate process wey expose tools to the agent through protocol. E often be network service wey you run yourself.
  • A rules file like CLAUDE.md na project context wey system read when session start, and e apply to everything.
  • A plugin na directory of skills, agents, hooks, MCP servers, or other components wey Claude Code installs and loads as one unit. Its optional manifest names am and fit add version, description, and other metadata. Distribution no be part of the plugin: marketplace na separate catalog wey list plugins and where to fetch each one.

So the question wey plugin answers no be "wetin agent fit do". Na "how I go ship this give my team and update am next month". If you dey choose between the first three, comparison of skills, MCP servers and rules files explain that decision well. If na the MCP part you care about, how to run your own MCP servers for VPS explain the hosting side.

Where plugins dey live and wetin dey inside one

Most marketplace plugins dey copied into local cache for ~/.claude/plugins/cache, with separate directory for each installed version. Relative-path plugins for marketplace wey you add from local path dey load for the same place, while command-source plugins for link mode dey load through links to their source directory. When you update or uninstall copied plugin, the old version directory dey marked as orphaned and deleted about two weeks later, so session wey don already load the old version go continue to work instead of failing for middle of task. Cleanup dey run only when at least one plugin installation still dey recorded; after you uninstall your last plugin, orphaned directories go remain until you install another one.

Because path for copied plugin dey change whenever you update am, plugin must never hardcode its own location. Hooks and MCP configs inside plugin dey use ${CLAUDE_PLUGIN_ROOT}, wey resolve to the current install directory. State wey must survive update dey go inside ${CLAUDE_PLUGIN_DATA}, wey resolve to stable directory under ~/.claude/plugins/data/.

For copied plugin, na only the plugin own directory dey copied into cache, except symlink targets wey dey elsewhere for the same marketplace; dem dey copied into the correct place. Script wey reads path outside plugin root, like ../shared-utils, go work while you dey develop with local path but go break when plugin dey copied into cache, because those files never get copied. Component path declared that way inside plugin.json go fail earlier: Claude Code rejects any component path wey escapes the plugin root, even during development, with path escapes plugin directory error.

The layout dey look like this.

my-plugin/
├── .claude-plugin/
│   └── plugin.json
├── skills/
│   └── code-review/
│       └── SKILL.md
├── agents/
├── hooks/
│   └── hooks.json
├── .mcp.json
└── bin/

Na only plugin.json dey go inside .claude-plugin/. Everything else dey stay for plugin root. Putting skills/ or hooks/ inside .claude-plugin/ na the most common reason plugin installs cleanly and then no do anything at all: Claude Code dey look for those directories for root, finds none, and loads plugin wey get no components.

The manifest itself small.

{
  "name": "my-first-plugin",
  "description": "A greeting plugin to learn the basics",
  "version": "1.0.0"
}

How to install a Claude Code plugin

Installation na two steps, and the first one no install anything. You go add a marketplace, wey be catalog of plugins, then install individual plugins from there. Anthropic official marketplace, claude-plugins-official, dey already registered for you the first time you start Claude Code interactively, unless managed policy or CLAUDE_CODE_DISABLE_OFFICIAL_MARKETPLACE_AUTOINSTALL block am. You need add other ones yourself.

/plugin marketplace add anthropics/claude-plugins-official
/plugin install commit-commands@claude-plugins-official

Marketplace name dey come from catalog file inside the repository, not from repository path. So read the marketplace name from the Marketplaces tab of /plugin before you type install command. The old claude-code-plugins marketplace for anthropics/claude-code don deprecate; use claude-plugins-official for new installations. Existing plugins from that old marketplace dey update from official repository, so you no need reinstall commit-commands just to change marketplace.

After installation, read the summary line. Plugin is now active. mean say the components don load for this session. Run /reload-plugins to apply. mean say dem no activate during installation; the panel go close and Claude Code go run the reload for you. If the reload warn say e go read the conversation again, e go leave the plugin pending, and you go rerun am as /reload-plugins --force. Then confirm say the plugin dey really there: /plugin show am under the Installed tab, typing / list its skills as /plugin-name:skill-name, and anything wey fail to load go show under the Errors tab with the reason.

Installation go ask for scope, and scope decide who get the plugin. User scope na you, for every project. Project scope write the plugin into the repository's .claude/settings.json under enabledPlugins, wey turn am on for everybody wey clone the repository. E no download am for dem: every collaborator still need install am once for their own machine, for example with claude plugin install name@marketplace --scope project, unless the marketplace entry na relative path or seed directory already hold am. Local scope na you, for this repository only.

For script, Dockerfile, or any session wey no get interactive panel, use shell form instead. E install to user scope unless you pass --scope. For machine wey nobody don open interactive session before, official marketplace no dey registered, so add am first.

claude plugin marketplace add anthropics/claude-plugins-official
claude plugin install commit-commands@claude-plugins-official --scope project
claude plugin list

claude plugin install dey run outside session, so session wey already open no go see the new plugin until you run /reload-plugins or start new session.

Managing wetin you get na the same pattern for both places. /plugin list print wetin dey installed, and e accept --enabled or --disabled. /plugin disable name@marketplace turn plugin off without removing am, /plugin enable turn am back on, and /plugin uninstall remove am. The slash-command forms open plugin panel to apply the change, na why you suppose use the claude plugin ... shell equivalents for scripts.

To give whole team one marketplace, put am inside the project's .claude/settings.json, or run claude plugin marketplace add your-org/claude-plugins --scope project once and commit the file wey e write. Claude Code go register the marketplace for each member once dem trust the repository folder.

{
  "extraKnownMarketplaces": {
    "my-team-tools": {
      "source": {
        "source": "github",
        "repo": "your-org/claude-plugins"
      }
    }
  }
}

While you dey build your own plugin, skip the marketplace completely. claude --plugin-dir ./my-plugin load directory for that session, /reload-plugins pick up your edits without restart, and claude plugin validate ./my-plugin check the manifest, skill and agent frontmatter, and hooks/hooks.json before anybody else see am. Marketplace no required to share one either: you fit send person the folder or a .zip of am to load with --plugin-dir, but dem no go get updates until you send new copy.

Plugin for Claude Code dey cost how much?

The mechanism free. As of November 2026, dem no dey charge you to add marketplace, install plugin, or leave one enabled. The official and community marketplaces na public git repositories, and plugin na directory wey contain text files.

Wetin plugin cost na tokens, and na tokens your subscription usage or API bill dey actually measure. Which one plugin go use depend on how you dey pay for the tool from the beginning, and wetin Claude Code cost for each plan explain the subscription tiers alongside per-token API pricing. This cost dey come through three different ways, and each one dey behave differently.

Standing context cost. Every enabled plugin dey put the names and descriptions of its skills, agents, and commands wey Claude fit invoke by itself inside your context. The system dey read that part again on every turn of the session, whether anything run or not. No be every component dey cost tokens: harness dey run hook, so e no add anything to model context. Before you install, the /plugin detail view dey show one Will install section wey list the commands, skills, agents, hooks, and MCP and LSP servers wey you wan add. For plugin wey dey Anthropic's official marketplace, e also dey show one Context cost section with Every turn and When invoked estimates, as long as you open the plugin through its marketplace name or from the Marketplaces tab, instead of from the Discover list. Read both. Plugin from your own marketplace or third-party marketplace no get Context cost section, but you no need estimate am by hand: claude plugin details name for your shell go print one Always-on token figure and per-component breakdown for any loaded plugin, including one wey you pass with --plugin-dir. This mean you fit clone repository and run claude --plugin-dir ./that-plugin plugin details that-plugin before you install anything. That figure no include MCP servers. Plugin wey bundle MCP server usually na the heaviest, because tool definitions dey large. But for models wey support MCP tool search, dem definitions dey wait until tool become necessary. Run /context for session wey plugin dey enabled, then read the MCP tools line to see wetin dem add.

Invocation cost. When you run plugin skill, e append the skill instructions to the conversation, so na only when you use the skill you go pay for its body. But the body na the cheaper part. Wetin the skill tell agent to do fit cost more: the unlazy skill's Depth Tree method dey spend almost all its tokens on the extra passes wey e force before agent fit mark task as done, not on the file wey you install. Agent different. Subagent dey run its own conversation with its own system prompt and its own cache. E dey start with no cache hits, so plugin wey workflow dey use to spawn agents fit cost much more than its context estimate suggest.

Cache cost. If you enable or disable plugin in the middle of session, e fit force the next request to process the whole conversation again. Skills, commands, agents, hooks, monitors, and themes no dey do this: anything wey dem add dey append after the existing history, so the next request pay for the new content and still read everything before am from cache. The exceptions na plugin wey provide MCP server and code intelligence plugin, wey language server dey give Claude the LSP tool. For MCP, e depend on tool search: if tools dey deferred, cache dey survive. If dem load into prompt prefix, the next request go read the entire conversation again as uncached input. Na exactly why /reload-plugins dey warn and refuse when change to MCP tools or the LSP tool go break the cache, until you pass --force.

You fit monitor this instead of guessing. Every API response dey report cache_read_input_tokens and cache_creation_input_tokens, and one custom statusline wey dey show live token usage go put both for your front. Healthy session dey read much more than e create. If creation remain high turn after turn, something for your prefix dey change every turn. For the bigger picture of wetin dey fill the window, see how to manage the Claude Code context window and wetin those token counts actually mean.

One housekeeping step dey pay for itself. The Installed tab dey group marketplace-installed plugins wey you never use for at least 14 days and 10 sessions under one Not used recently header, with one Last used line for the detail view. The header no include managed and seed-directory plugins, plus plugins wey get theme, output style, monitor, or workflow. LSP server count as used when e deliver diagnostics or answer code navigation request. When strictKnownMarketplaces dey configured, neither the header nor the Last used: line go appear. Those plugins still dey cost startup time and context for every session. Disable or uninstall dem. The cheaper habit na to choose carefully before you install: which of Matt Pocock's skills dey worth installing dey review one 27-skill collection and separate the few wey solo developer need from the ones wey only pay off for team.

Plugin dey run with your permissions

Anthropic documentation talk am plainly: plugin wey you install fit execute arbitrary code for your machine with your user privileges. This no be hypothetical matter. A plugin's hooks run shell commands when session events happen, including before and after tool calls. The plugin's bin/ directory dey added to Bash tool's PATH while the plugin dey enabled. Its stdio MCP servers na processes wey e start, and a mod dey run JavaScript inside Claude Code itself. Hooks, monitors, server processes, and mod processes dey run outside Claude Code's sandbox; but its permission rules still apply to tool calls wey Claude make.

For laptop, this risk dey limited by wetin your desktop user fit reach. For server, e usually no be so. The account wey dey run the agent often get SSH keys, deploy tokens, cloud CLI session, and access to Docker socket. So "arbitrary code as your user" mean say the whole machine fit dey exposed. If Claude Code dey run for VPS, read how to run Claude Code safely on VPS before you install anything. Also read how to keep credentials away from agent before you install plugin wey dey talk to external service. Other harnesses get the same problem for the same rented box. Na why DeepSeek Harness plugins wey worth installing mostly focus on spend caps, tool permission rules, and injection scanning instead of new capabilities.

Some guardrails dey, and e good make you know which ones. Project-scope plugin dey come from repository instead of from you. So e loads only after you trust the workspace. Its MCP servers still need approval for each server, and its background monitors no load at all. Plugin-shipped agents no get permission to declare hooks, MCP servers, or permission mode. When marketplace plugins dey copied into cache, symlinks wey point outside marketplace dey skipped. This stop copied plugin from pulling arbitrary host files through those links.

None of this one replace reading wetin you install. Check the Will install list. Prefer plugins wey source code you fit open and read. Keep your team's plugins for marketplace repository wey you control. Run claude plugin validate on anything wey you write yourself. The thing wey you review fit still change later. Auto-update dey on by default for Anthropic's official marketplaces, except knowledge-work-plugins and first-party-plugins. E dey off for third-party marketplaces wey you add by source. Where e dey on, Claude Code dey update the plugin files for background.

FAQ

Plugin install korile extra taka lage?

No. Plugin system, marketplace add, or plugin install no get any fee. Na token usage dey cost, and dem go bill am against your plan or your API spend like any other context. Plugin dey add standing context for every turn, add more when you invoke one of its skills or agents, and fit force one expensive uncached turn if e provide MCP server wey tools dey load into prompt prefix. For plugin wey dey Anthropic official marketplace, the /plugin detail view go show Context cost estimate before you install am; for any other plugin, claude plugin details name go print its always-on token cost once you install am or load am with --plugin-dir.

Wetin be the difference between plugin and skill?

Skill na one instruction unit. Plugin na package wey fit contain skills, agents, hooks, MCP servers, LSP servers, and monitors, with name, usually version, and usually marketplace wey you fit install am from, though you fit also share am as folder or .zip wey people load with --plugin-dir. Write standalone skill for .claude/ when na for you and this project. Single-purpose skill like Ponytail, wey dey push the agent toward the smallest change wey go work na the clearest example: one file with one rule inside, until the day your team want am too. Turn am into plugin when other people need am and e need updates over time. Plugin skills get namespace, so you invoke skill inside plugin as /plugin-name:skill-name instead of /skill-name.

My plugin install, but its skills no dey appear. Wetin dey wrong?

Check install summary first. If e say Run /reload-plugins to apply., Claude Code go run that reload when panel close, and if reload warn say e go read conversation again, run am as /reload-plugins --force. If e load but show nothing, open /plugin and read Errors tab. The commonest structural mistake na to put skills/, agents/, or hooks/ inside .claude-plugin/, where Claude Code no dey look for dem. Remember say plugin skills get namespace, so na /plugin-name:skill-name you dey find when you type /. If na another person plugin and e load without error, open its details pane for Installed tab: plugin wey list no skills there no get any to offer. If Errors tab say plugin no dey cached, its install directory no dey, so reinstall am from your shell with claude plugin install name@marketplace and run /reload-plugins.

I fit install plugins without interactive panel?

Yes. Use shell command claude plugin install name@marketplace, wey dey install to user scope unless you pass --scope project or --scope local. E work for scripts, images, and non-interactive environments where /plugin panel no dey available. Because e dey run outside session, session wey already open need /reload-plugins before plugin go take effect. For session wey no get interactive terminal, changes to plugin MCP servers go take effect for next session.

E safe to install plugin from marketplace wey I find for GitHub?

Treat am the same way you go treat running that repository install script as yourself, because na almost the same thing. Plugin fit run shell commands through hooks, add executables to Bash tool PATH, and start MCP servers, all with your user privileges. Anthropic no control or verify third-party plugin contents. Install from sources wey you fit read, review the Will install list before you confirm, and use stricter checks for server than laptop, because account for server usually hold keys and tokens wey people fit steal.

Dis article help you?