# Claude Code Skills vs Plugins: Token Cost Breakdown

> Skills and plugins cost tokens at different layers — discovery, frontmatter, and invocation. A line-by-line cost breakdown for both, with the cases where the difference actually matters on a solo repo's context budget.

Canonical: https://www.agentickit.co/compare/claude-code-skills-vs-plugins-token-cost
Last updated: 2026-11-18

A skill costs a small, fixed amount of context on every turn (its frontmatter description) plus a larger one-time hit only when it actually fires (its body). A plugin costs the sum of whatever skills, commands, and agents it bundles — it has no separate token cost of its own, but it can quietly multiply your fixed tax by installing more skills than you'd add by hand one at a time. This page is narrower than the full primitives comparison: it's only about where the tokens go, not which primitive to reach for. Last updated 2026-11-17 by Rakshit Yadav.



## Three layers, three different costs
"How much does a skill cost" and "how much does a plugin cost" aren't the same kind of question, because the tokens get spent at three distinct points, not one: **discovery** (what the agent has to read to know a skill or command exists at all), **frontmatter** (the short description that's resident in context so the agent can decide whether to open something), and **invocation** (the full body that loads only once something actually fires). A plugin doesn't add a fourth layer — it just bundles however many skills, commands, and agents it ships, and each of those pays the same three costs independently. There's no plugin-level token tax beyond the sum of its parts.

## Cost at each layer
**Discovery.** Claude Code scans `.claude/skills/*/SKILL.md` and `.claude/commands/*.md` at session start to build its index of what's available. This is cheap per item — a few dozen tokens to register a name and path — but it runs for every skill and command on disk whether or not the session ever uses them, including everything a plugin installed that you don't personally touch this session.
**Frontmatter.** Each skill's YAML `description` (and a command's one-line summary) stays loaded in context for the whole session once discovered, because that's the text the agent matches against the current task to decide whether to open the full file. A terse, well-written description runs 15-40 tokens; a sprawling one describing five unrelated use cases can run past 100. This is the layer that scales with *how many* skills you have installed, regardless of which ones ever actually fire — it's paid on every single turn, not just the turns that use the skill.
**Invocation.** The full SKILL.md body — the actual playbook, often several hundred to a couple thousand tokens — loads only on the turn the agent decides the description matches. This is the layer most people picture when they say "skills are cheap," and it's true in isolation: an unopened skill's body costs nothing beyond its frontmatter line. But it's also the layer that's easy to undercount, because a skill that fires on most tasks (a stack convention, a CLAUDE.md-adjacent rule) is paying its full body cost on most turns too.

## Cost profile comparison table
<table><thead><tr><th></th><th>Skill</th><th>Command</th><th>Plugin</th></tr></thead><tbody><tr><td>**Discovery cost**</td><td>One scan at session start, per skill folder</td><td>One scan at session start, per command file</td><td>No cost of its own — equals the sum of what it installs</td></tr><tr><td>**Per-turn resident cost**</td><td>Frontmatter description, every turn, regardless of use</td><td>Near zero — commands are typed, not matched by description</td><td>Equals the sum of every skill's frontmatter it bundles, every turn</td></tr><tr><td>**Invocation cost**</td><td>Full SKILL.md body, only on the turn it fires</td><td>Full command body, only on the turn it's typed</td><td>Same as above, per-primitive — a plugin install doesn't change this</td></tr><tr><td>**Scales with**</td><td>How many skills are installed × how often each fires</td><td>How often you type it (explicit, so no guessing cost)</td><td>How many skills/commands/agents it happens to bundle</td></tr><tr><td>**Easy to undercount because**</td><td>"It's free until it fires" ignores the frontmatter tax paid every turn it doesn't</td><td>N/A — commands have the most predictable cost of the three</td><td>Installing one plugin can add a dozen skills' frontmatter at once, not one</td></tr></tbody></table>Commands are the cheapest of the three on a per-turn basis because nothing about them sits resident in context waiting to be matched — you type `/audit` or you don't, so there's no standing frontmatter tax the way there is for a skill the agent has to keep deciding whether to open.

## When the cost difference matters for a solo repo
On a repo with five or six skills, the frontmatter layer is noise — a few hundred tokens total, not worth optimizing. The gap starts to matter once a plugin install adds more skills than you'd have written by hand: a marketplace plugin aimed at a broad audience often ships ten or fifteen skills to cover cases most of its installers don't need, and every one of those pays its frontmatter cost on every turn whether or not this particular repo ever triggers it. A solo founder running Claude Code against a single Next.js SaaS codebase rarely needs a plugin built for five different frameworks.
The practical check: after installing a plugin, open `.claude/skills/` and count what actually landed. If half of it is irrelevant to this repo's stack, that's resident frontmatter tax on every turn for the life of the project, not a one-time cost. Deleting the unused skill folders (or the matching entries if the plugin's manifest lets you select a subset at install time) removes that frontmatter tax the same way deleting an unused section from CLAUDE.md would — see [CLAUDE.md vs Skills vs Subagents vs Hooks](https://www.agentickit.co/compare/claude-code-skills-vs-subagents-vs-hooks) for the equivalent logic applied to the always-loaded file instead of the on-demand ones.
MCP servers are a separate and usually bigger line item on this same budget — a server's full tool schema loads every turn the same way skill frontmatter does, but tool schemas tend to run longer than a skill description. If you're auditing total fixed cost rather than just the skills-vs-plugins question, run the [Context Budget Estimator](https://www.agentickit.co/tools/context-budget-estimator) to see CLAUDE.md, MCP servers, and skills added together.



## FAQ

### Does installing a plugin cost more tokens than installing the same skills one at a time?
No — the per-turn cost is identical either way, because it's the sum of each bundled skill's and command's own frontmatter and invocation cost. The practical difference is that a plugin install adds all of its bundled skills in one step, which makes it easier to end up with more installed skills than you intended, each paying its own frontmatter tax.

### If a skill's body never loads, does it cost anything?
Yes — its frontmatter description stays resident in context every turn so the agent can keep deciding whether to open it, even on turns where it never fires. Only the (usually larger) body cost is avoided until invocation.

### Are commands cheaper than skills?
On a per-turn basis, yes — a command has no standing frontmatter tax the way a skill does, because it's invoked explicitly by name rather than matched by a resident description. Its full body only loads on the turn it's typed.

### How do I find out my actual number instead of estimating?
Run the Skill Token-Cost Estimator on an individual SKILL.md or CLAUDE.md file, or the Context Budget Estimator to add CLAUDE.md, installed MCP servers, and skills together into one combined total.

## Related
- [CLAUDE.md vs Skills vs Subagents vs Hooks](https://www.agentickit.co/compare/claude-code-skills-vs-subagents-vs-hooks)
- [Claude Code Plugins vs Skills vs MCP: What to Use When](https://www.agentickit.co/compare/claude-code-plugins-vs-skills-vs-mcp)
- [CLAUDE.md Context Budget Estimator](https://www.agentickit.co/tools/context-budget-estimator)
- [Skill Token-Cost Estimator](https://www.agentickit.co/tools/skill-token-cost)
- [Skills pillar](https://www.agentickit.co/skills)
- [Claude Code Plugins vs Commands](https://www.agentickit.co/compare/claude-code-plugins-vs-commands)
