Claude Code Plugin Marketplace: What 172 Plugins Cost

I run this blog from a Mac mini with three Claude Code plugins installed. Answering a simple question — is it worth adding a fourth? — I found that the plugin marketplace tells you what a plugin does and, for some plugins, what it costs. The two sets are not the same, and the plugins I actually run fall in the gap.

So I read the file Claude Code already keeps on my disk: ~/.claude/plugins/plugin-catalog-cache.json, 238,527 bytes, a cached copy of the official marketplace catalog. It carries a number the listings never show you in aggregate — how many tokens each plugin adds to every single turn.

What is in the catalog, and what bounds it

The cache holds 172 plugins. Each has tokens.always_on (charged every turn, forever), tokens.on_invoke (charged only when something fires), unique_installs, and a component inventory. Three limits before any number:

• It is stamped fetchedAt: 2026-05-15T04:54:50Z — 101 days before I read it. A picture of May, not today.
• Its models array is ['claude-opus-4-7', 'claude-sonnet-4-6']. Neither is the model I run now. I use the Opus column and label it.
• The units are a trap: plugin-level figures are tokens, the per-skill breakdown is chars. Across the 116 plugins with skills the median ratio is 2.68 (range 0.32–3.37). Adding them together produces nonsense, so I use plugin-level tokens only.

The most-installed plugins are the cheapest ones

Median always_on is 239 tokens; the mean is 777, dragged up by a tail ending at 8,752 (twilio-developer-kit, 55 skills). Install all 172 and you carry 133,679 tokens before typing a character — the kind of standing charge that decides how soon you hit the auto-compact threshold. The top ten are 36.1% of it.

Spearman’s rho between cost and installs is −0.311 across 171 plugins. Negative — cheaper plugins get installed more. At the extremes:

Group	Summed always_on	Summed installs
Ten most-installed	1,448	3,415,976
Ten most expensive	48,276	45,996

The ten most expensive ask 33.3 times the context for 1.3% of the adoption. The single most-installed plugin, frontend-design, has 739,578 installs and costs 91 tokens. The most expensive one has 199 installs.

43 of the 172 cost exactly zero, and every one ships no skills, commands, or agents — the MCP, LSP, and output-style plugins: context7 (322,649 installs), github (241,860), playwright (225,178), typescript-lsp (164,992). Those 43 hold 29.1% of all installs in the catalog.

Update, 2026-09-11: the other cost of a plugin is the hook it installs. In a census of 294 Stop hook error issues, the largest named families are hooks that never started: a plugin path that moved, a missing /bin/sh, a node not on PATH. The breakdown and ten probe hooks are in Claude Code Stop hook error: 294 issues, 10 probes, 6 fired.

Zero always_on is not free. An MCP server bills you through its tool schemas instead — a line the catalog does not track. Anthropic’s MCP documentation describes tool search as the way to defer that cost, and in my own sessions the MCP block appears as a separate (deferred) row. The catalog measures one lane, not the whole road.

Skill count is almost the whole story

Pearson’s r between skill count and always_on is 0.917. Across the 116 plugins shipping at least one skill, the median is 136 tokens per skill.

That yields a rule you can apply to plugins the catalog never rated: always_on ≈ 136 × skill count. Count the SKILL.md files and multiply.

My own plugins are not small

Plugin	Marketplace	Skills	On disk	Catalog always_on
oh-my-claudecode 4.13.7	third-party	38	52,832 KB	not listed
toprank 0.19.0	third-party	22	2,220 KB	not listed
discord 0.0.4	official	2	50,312 KB	160

Two of my three have no published cost anywhere. By the rule, oh-my-claudecode lands near 5,176 tokens — 4th most expensive in a 173-entry catalog — and toprank near 2,996, ranking 11th. I carry two top-eleven plugins that no listing rates.

That is a regression estimate, so I checked it against a direct reading. When I measured /context in an earlier session while tracking down what Claude Code’s cost command actually reports, the skills block came to 8,200 tokens. The rule predicts 5,176 + 2,996 + 160 = 8,332, a 1.6% gap. Two unrelated sources landing that close is why I trust the multiplier.

The table also killed an intuition I walked in with. discord takes 50,312 KB of disk and costs 160 tokens; oh-my-claudecode takes 52,832 KB and costs an estimated 5,176. Nearly identical on disk, 32× apart in context. Disk size predicts nothing.

Why my plugins have no number

I started this assuming the marketplace hides the cost. That was wrong, and the correction changes the advice. Anthropic’s install-side documentation says the detail pane shows “a Context cost estimate so you can see how many tokens the plugin will add to your context window every turn.” It is there.

The next sentence is the one that matters: “Not every plugin provides the data behind these fields. For plugins from local or custom marketplaces, you may not see the Context cost and Last updated rows.”

So the number exists exactly for official-catalog plugins and is missing exactly for third-party ones — which, in my install, are the big ones. The separate marketplace-authoring guide never mentions context cost at all; grep it for token and you get git authentication. The shopper sees a price; the publisher is never told to set one.

The second thing nobody told me: my plugins had stopped updating

Marketplace	Type	lastUpdated
claude-plugins-official	official	2026-08-24T05:00:19Z
omc	third-party	2026-05-15T04:54:45Z
nowork-studio	third-party	2026-05-15T06:58:22Z

The official marketplace refreshed this morning; the other two had not moved in 101 days. That is documented, not broken: “claude-plugins-official and most other official Anthropic marketplaces have auto-update enabled by default. Third-party and local development marketplaces have auto-update disabled by default.”

Checked against the GitHub API today: I run v4.13.7, released 2026-05-09. Twenty releases have shipped since. Upstream is v5.0.0, tagged 2026-08-24T04:37:07Z — roughly three hours before I wrote this. My local marketplace.json still reads 4.13.7, so the client does not know a newer version exists.

For scale on what I was missing: the May snapshot held 172 plugins, today’s live marketplace.json holds 286. That is 118 added and 4 removed in 101 days, up 66.3%. One removal was aws-dev-toolkit — previously the 4th most expensive plugin in the catalog.

Two people on Hacker News were both right

The argument has been running in public, and the catalog settles it. In December, Jimmc414 argued skills are cheap by construction: “Claude sees a one line index entry per skill. You can create hundreds. The full instructions only load when invoked… With skills, you only pay for what you use.” This August, km144 said the opposite: “Adding skills or plugins bloats the context window.”

The mechanism is Jimmc414’s. Median on_invoke is 5,716 against median always_on of 239 — lazy loading is real, a 23.9× saving. But “only pay for what you use” fails at the index layer, because the index is not free: 133,679 tokens across the catalog, and in my own measurement of skills against agents, 74 skills never once invoked still cost 6,920 tokens every session. The two were describing different layers.

A third comment names the structural cause. In April, michiosw wrote that “the permission model is ‘all or nothing.’ Once you install a plugin, it gets full context on every session, every prompt,” and argued plugins should declare which project types they apply to. Until that exists, every skill in every installed plugin is billed against every unrelated task — and on a fleet that runs unattended, that lands straight in the per-post cost of running this blog.

What I changed

1. Count skills before installing. Multiply by 136. A 40-skill plugin is a ~5,400-token standing charge, not a free addition.
2. Prefer the zero-cost shapes when they fit. MCP and LSP plugins carry no always-on index. Not a reason to install more of them — their tool schemas bill separately — but between two plugins that solve your problem, component shape decides the standing cost.
3. Check third-party marketplaces by hand. Mine were frozen 101 days with auto-update off by default and nothing surfaced it. /plugin marketplace update <name> is the command, and the Not used recently list in the /plugin Installed tab is where to find plugins still charging for skills you stopped invoking.

I have not updated to 5.0.0 yet. A major version across 20 releases will move the skill count, and therefore the number this whole post is about, so I would rather measure it deliberately than mid-publishing-run. That measurement is the follow-up.

The 172-plugin figures come from one file — plugin-catalog-cache.json, dated 2026-05-15 — read on my own machine on 2026-08-24, using the claude-opus-4-7 column and plugin-level tokens only. One snapshot, one machine, 101 days stale, and today’s live marketplace already lists 286 plugins rather than 172, so treat the distribution as directional and per-plugin values as of May. My own plugin costs are estimates from the 136-tokens-per-skill regression, not direct readings; the one figure I measured directly, 8,200 tokens of skills via /context, agreed within 1.6%. Versions and release dates were checked against the GitHub API on 2026-08-24 and the documentation quotes read the same day. I could not use Reddit — it blocks the user agent I search with — so the community evidence here is Hacker News only.