The token cost of claude-code-setup is two numbers, not one

People hear "plugin" and assume installing one drops 10K tokens on every turn. For claude-code-setup that's wrong by about two orders of magnitude. The plugin is structurally one skill, and Claude Code skills are designed for progressive disclosure: the discovery payload (name plus description) loads every turn so the model knows the skill exists, the body loads only when the skill is invoked.

So "what does it cost" is really four numbers: idle, first invocation, full reference fan-out, and the second-order tax on whatever the recommender talks you into installing. Below is each one, with the actual bytes you can verify from the public repo.

The plugin is one skill (and a PNG)

The repository at anthropics/claude-plugins-official/plugins/claude-code-setup contains exactly six top-level entries: .claude-plugin/plugin.json (273 bytes of manifest), a LICENSE file (11,358 bytes, not shipped to context), a README (1,015 bytes, also not in context), an example PNG that's only there for the marketplace listing, and one skill folder. That's the entire plugin.

Two things matter for cost. The 545KB PNG never enters context (image embeds in marketplace listings don't flow into the model's system prompt). And the 273-byte plugin.json is also not in the system prompt: it's read by the Claude Code client to register the skill, not by Claude itself. The only bytes the model ever sees from this plugin are inside the skill folder.

State 1 · idle (every turn): the frontmatter only

When a skill is installed but not invoked, Claude Code lists it in the available-skills section of the system prompt with its name and description so the model knows when to reach for it. For claude-automation-recommender, that's the YAML block at the top of SKILL.md:

The whole block is ~310 characters once whitespace is normalized, which lands at ~78 tokens. Add the wrapper formatting Claude Code puts around the available-skills list (the heading, the bullet markers, the "invoke with /skill-name" hint) and call it 80–100 tokens per turn for this one skill. That is the entire per-turn cost of having the plugin installed.

State 2 · invoked (SKILL.md body)

When a user types /claude-automation-recommender or asks something like "help me set up Claude Code" and Claude routes to the skill, the full SKILL.md body loads into context. The body is 10,856 bytes on disk, so call it ~2,714 tokens.

The body is structured in two phases. Phase 1 issues a handful of shell commands (the relevant ones are ls -la package.json pyproject.toml Cargo.toml go.mod pom.xml, cat package.json | head -50, and ls -la .claude/ CLAUDE.md). Each of those Bash calls eats a few hundred input tokens of tool I/O depending on your repo size. Phase 2 emits the recommendations: 1–2 picks across five categories.

State 3 · full reference fan-out

SKILL.md uses relative links like [references/mcp-servers.md](references/mcp-servers.md) to pull in detailed pattern catalogs. Whether all five reference files load depends on what Phase 1 found. A polyglot repo with frontend, database, and CI signals will probably trigger four or five of them. A small Python script will trigger one or two. The worst case totals ~31,412 bytes, ~7,900 tokens, on top of the SKILL.md body.

State 4 · the cost the plugin asks you to take on

The above is the cost of running the plugin. The recurring cost is what the plugin convinces you to install. Its output is a list of 1–2 MCP servers, 1–2 skills, 1–2 hooks, 1–2 subagents, and 1–2 slash commands. Some of those (especially MCP servers and skills) have their own discovery payloads that load on every turn after you install them.

A worked example: imagine the recommender suggests Playwright MCP, the frontend-design skill, an auto-format hook, a security-reviewer subagent, and a /test slash command, and you adopt all five. The plugin itself drops back to idle (~78 tokens per turn). The five new entries add their own descriptions to the discovery payload: roughly 60–120 tokens each. Call it 450 tokens per turn permanently, 5.7x the plugin's own cost.

That second-order cost is invisible at install time and almost never measured. It's also where ccmd's analyzer at src/lib/analyzer.ts earns its keep: paste your CLAUDE.md or the skill descriptions you've installed and it scores total tokens, three cost states (cold, cached, cache-busted), and a per-finding savings estimate. The plugin tells you what to install; the analyzer tells you what it costs to keep installed.

The manifest (for completeness)

The 273-byte plugin manifest, in full, looks like this. None of it is in the model's context window. It exists so the Claude Code client can register the skill and surface it in /plugin.

Same answer as the skill's tools field (Read, Glob, Grep, Bash, no Write/Edit): the plugin is structurally a recommender, not a mutator. That's why its per-turn cost can stay this small and why we don't flag installing it as a hot move. We flag the configs it tells you to install.

FAQ

Related on ccmd: CLAUDE.md token cost audit, when installed skills become bloat, what fires on every turn vs only on invoke.