Claude Code Plugin Usage: Measure Cost, Find Dead Plugins

David IyaDavid Iya November 18, 2026 9 min read
A grid of brass utility gauges with glass dials showing different readings on a dark slate wall, a clipboard with a blank tally sheet hanging beside them in cool light
Original image, Claude Code Club

Why Claude Code plugin usage is worth measuring

Claude Code plugin usage has two sides, and most teams only look at one. The first is cost: every session where a plugin is enabled includes the names and descriptions of its skills, agents and commands in Claude's context, and those tokens count against the user's usage whether or not the plugin gets used. The second is use: whether anyone still invokes the thing. A plugin that costs tokens on every turn and never runs is pure overhead.

Once a company rolls out a marketplace, plugins pile up. Someone installs five to try them, keeps two, and forgets the rest. Multiply that across a team and you are paying a small tax on every session for tools nobody opens. I call the approach below the CCC Plugin Ledger: price each plugin, check who uses it, then decide what stays. It works for one person and scales to a fleet.

Measure what one plugin costs with claude plugin details

Run claude plugin details with the plugin's name in a terminal, not at the prompt of a running session. The plugin has to be loaded: installed, in a skills directory, or passed with --plugin-dir in the same command, for example claude --plugin-dir ./formatter plugin details formatter.

What each part of claude plugin details tells you

Part of the outputWhat it answers
Component inventoryWhat Claude Code found in the plugin: skills, agents, hooks, MCP servers and LSP servers. Commands are counted with skills
Always-onThe tokens the names and descriptions of the plugin's skills, agents and commands add to every session where it is enabled. This is the number every user carries
Per-componentEach skill, agent or command split into its always-on share and its on-invoke cost, which is the body that loads only when that component runs

Two blind spots. Hooks get no cost estimate, and the output marks them as having no model context cost. MCP servers are not counted either, because their tool schemas are resolved at runtime. To see what a plugin's MCP tools add, run /context in a session with the plugin enabled and read the MCP tools category. If a plugin is mostly an MCP server, its details output will look cheaper than it is.

Plugins in the official marketplace show this cost before install. In /plugin, the details pane for a plugin has a Context cost section with an Every turn line and a When invoked line, and the Every turn line is highlighted when the always-on figure is 2,000 tokens or more. A plugin in your own marketplace has no Context cost section, so for internal plugins, claude plugin details is the only way to see the number before people install it.

Lower the always-on figure if you maintain the plugin

The always-on figure counts each component's name plus its description and when_to_use frontmatter. There are two levers: shorten skill and agent descriptions, or split a large plugin so people install only the components they need.

My take: if you run an internal marketplace, put the claude plugin details always-on number in each plugin's review. It costs nothing to read, it makes a bloated description visible before rollout, and it gives you a fair way to say no to a plugin whose per-session cost is out of proportion to what it does.

Find plugins nobody uses on one machine

Claude Code does not report a plugin's usage back to its author. Usage is recorded on the machine of each person who installed it. If those people are teammates you can ask, each of them has four places to look, all run at the Claude Code prompt in their own session:

Where a user sees their own plugin usage

WhereWhat it shows
/plugin, Installed tabA marketplace plugin moves under a Not used recently header once it has gone unused for at least 14 days and 10 sessions. Its details show a Last used line
/skill-doctorWhat each skill costs and how often it runs, flagging skills that are listed but have never been invoked, including skills from plugins. Needs v2.1.252 or later
/doctorLists each user-installed skill, MCP server and plugin, and recommends disabling the ones that were not used
/usageOn Pro, Max, Ultra, Team and Enterprise plans, attributes recent usage to skills, subagents, plugins and MCP servers as a share of the total

The Not used recently header has gaps worth knowing before you rely on it. It never appears for plugins loaded with --plugin-dir or from a skills directory, for plugins enabled through managed settings or mounted from a seed directory, or for plugins that include a theme, output style, monitor or workflow, because those are in use without a tracked invocation. And if your organization sets strictKnownMarketplaces, neither the header nor the Last used line appears at all.

That last one catches admins out. The tighter your marketplace policy, the less your users see about their own unused plugins, so the cleanup job moves to you. A language server plugin, by contrast, counts as used whenever it delivers diagnostics or answers a code navigation request, so an active one will not be flagged.

Measure plugin usage across an organization

If you administer Claude Code for a team, you have two fleet-wide sources. OpenTelemetry events go to your own backend once you configure an exporter. The Analytics API is served from Anthropic's records with no exporter, on the Enterprise plan only. The OpenTelemetry guide covers the exporter setup.

Which OpenTelemetry signal answers which plugin question

QuestionEvent or attribute
Which plugins get installed, and from whereclaude_code.plugin_installed, one per install, with install.trigger set to cli or ui
Which plugins are active in how many sessionsclaude_code.plugin_loaded, one per enabled plugin at session start
Which skills activate, and which plugin owns themclaude_code.skill_activated, with plugin.name and marketplace.name for plugin skills
What a plugin's hooks reportclaude_code.hook_plugin_metrics, emitted only for hooks in official-marketplace plugins
What a plugin costs in API spendplugin.name and marketplace.name on the cost counter, set when the active skill or subagent belongs to a plugin

plugin_loaded is the most useful of the five for a cleanup. Compare it with skill_activated and you get the plugins that load in many sessions but whose skills almost never fire. Two of its attributes make the picture sharper. plugin.scope says where the plugin came from: official, community, org, user-local or default-bundle. enabled_via says how it got turned on: default-enable, org-policy, admin-install, seed-mount or user-install. Before v2.1.246, plugins set to required or auto-install in your organization settings were reported as user-install or seed-mount, so filter out older versions before you trust that split.

The redaction tradeoff with OTEL_LOG_TOOL_DETAILS

Here is the part that decides whether your dashboard is useful. Plugins from the official marketplace report their names verbatim. Every other plugin's name is redacted or left out by default, including a plugin from your organization's own marketplace.

Plugin names in your backend, by default and with OTEL_LOG_TOOL_DETAILS=1

EventDefaultWith OTEL_LOG_TOOL_DETAILS=1
plugin_loadedplugin.name and marketplace.name are the literal string third-partyReal names
plugin_installed, skill_activatedName fields omitted, and skill.name on skill_activated is custom_skillReal names
Cost counterplugin.name is third-party, marketplace.name absentReal plugin.name, marketplace.name still absent

Without the flag you can still count distinct third-party plugins, because plugin_loaded carries plugin_id_hash, a deterministic hash of the plugin and marketplace name sent only to your exporter. That gives you how many plugins are out there, but not which ones.

My recommendation: start with the hash. Count distinct plugins and sessions per plugin_id_hash for a month. If the numbers show a real sprawl problem, turn on OTEL_LOG_TOOL_DETAILS for a short, announced window, get the names you need, and turn it off. Leaving it on permanently to answer a question you only ask quarterly is a poor trade.

Query the Analytics API on Enterprise

On the Enterprise plan, GET /v1/organizations/analytics/plugins answers which plugins your organization installs and invokes from Anthropic's records, with no exporter. It returns per-plugin, per-day install and invocation counts across Claude Code and Cowork, which you can group by user, RBAC group or product. The API is not available on the Team plan.

  • Authenticate with an API key that has the read:analytics scope. A Primary Owner creates it.
  • Plugins from the official and community tiers appear by name. Every other plugin is rolled into one aggregate third-party row.
  • That means an internal marketplace shows up as part of third-party here too. For per-plugin numbers on your own plugins, OpenTelemetry with names turned on is the route.

For the wider picture of who uses Claude Code and how much, the analytics dashboard guide covers the dashboard side of the same data.

The CCC Plugin Ledger checklist

  1. Price every plugin you allow with claude plugin details and record the always-on figure. Check MCP-heavy plugins in /context, since their tools are not counted.
  2. Ask maintainers of internal plugins to trim descriptions, and to re-run their trigger evals afterward.
  3. Tell users about /skill-doctor and /doctor. If you set strictKnownMarketplaces, know that the Not used recently header disappears for them.
  4. Export plugin_loaded and skill_activated. Count by plugin_id_hash first.
  5. Turn on OTEL_LOG_TOOL_DETAILS only for a defined window, and only if your backend is cleared to hold Bash commands and tool input.
  6. On Enterprise, pull the Analytics API plugin report for official and community plugins, grouped by RBAC group.
  7. Remove or stop requiring plugins that load in many sessions and almost never activate.

Token cost is the reason this matters for a single user too. What actually eats your Claude Code limit shows how small per-turn overheads add up over a long working day.

Free Claude Code drops, straight to your inbox

Short, practical drops on skills, MCP, agents, prompts, and more. No spam, unsubscribe anytime.

Frequently asked questions

Do Claude Code plugins use tokens when I don't use them?

Yes. Every session where a plugin is enabled includes the names and descriptions of its skills, agents and commands in Claude's context, and those tokens count against your usage whether or not the plugin runs.

How do I see what a Claude Code plugin costs?

Run claude plugin details followed by the plugin's name in a terminal. It shows the always-on token cost added to every session and splits each skill, agent or command into its always-on and on-invoke cost. MCP tools are not counted, so check those with /context.

How do I find Claude Code plugins I no longer use?

Open the Installed tab in /plugin. A marketplace plugin moves under Not used recently after at least 14 days and 10 sessions unused. /skill-doctor flags skills that have never been invoked, and /doctor recommends disabling unused plugins.

Can a plugin author see how many people use their plugin?

No. Claude Code does not report a plugin's usage back to its author. An organization's admin can measure it across their own machines with OpenTelemetry events or, on Enterprise, the Analytics API.

Why do my plugin names show as third-party in telemetry?

Only official-marketplace plugins report their names by default. Every other plugin, including your own organization's, is redacted to third-party or omitted unless OTEL_LOG_TOOL_DETAILS is set to 1, which also logs Bash commands and tool input.

Is the plugin Analytics API available on the Team plan?

No. The Analytics API, including the plugins endpoint, is available on the Enterprise plan. Team plan admins can measure plugin usage with OpenTelemetry events instead.

Last reviewed by David Iya on November 18, 2026

David Iya

Written by

David Iya

Forbes 30 Under 30 · Y Combinator

Keep reading

Claude CodeTroubleshooting

Why does Claude Code keep writing buggy code, and how do I fix it?

Claude Code writes buggy code mostly because it can't see what you see and has no way to check its own work, so it stops when the result merely looks done. Here is a six-step debugging routine anyone can follow, a copy-paste prompt that makes Claude find the cause before it touches anything, and how to undo a fix that made things worse.

Duncan Rogoff 9 min
Read article
Claude CodeSetup

Claude Code Mods for Organizations: Control What Runs

Claude Code mods are on by default from v2.1.286, and a mod runs inside Claude Code with the user's own permissions. Here is what your users get when you change nothing, the one managed setting that stops user-installed mods, how to make a mod count as your organization's, and how to review a mod before anyone installs it.

David Iya 10 min
Read article

Ready to build it yourself?

Join Claude Code Club, the #1 community for learning claude code, for $9/month.

← Back to the blog