Skills vs MCP: we measured 6,722 Agent Skills

How each one reaches the model

An MCP server sends the definition of every tool it offers to the model on every request, whether or not the tools get used (clients that load tools on demand pay less). A skill loads in three steps: its name and description at startup, the SKILL.md body when the agent decides the skill applies, and bundled files only when the agent opens them. That is the Agent Skills specification's progressive disclosure, and it is why the two compare so differently.

Findings

• Idle cost: 58 tokens for a skill, 1,362 for an MCP server. The always-on part of a skill (name and description) has a median of 58 tokens and a 90th percentile of 152. The median MCP server we measured adds 1,362 tokens per turn and the 90th percentile 8,335. Popular skills write longer descriptions: the most-installed ones have a median of 90.
• In use, a skill costs about as much as a server. The median SKILL.md body is 1,418 tokens (1,950 for the most-installed skills) and the 90th percentile 4,073. 398 skills (5.9%) exceed the specification's 5,000-token guideline; the largest is 142,335 tokens.
• Installing a whole collection erases the advantage. Every installed skill's description is always loaded. The biggest collections we crawled would add at least 24,406 tokens (wondelai/skills, 150+ skills), at least 20,796 tokens (NVIDIA/skills, 150+ skills), 17,401 tokens (aaron-he-zhu/aaron-marketing-skills, 120 skills) before any of them is used: more than a 90th-percentile MCP server. Install the skills you need, not the pack.
• 4.9% of skills can't be picked by an agent. 329 have no frontmatter or no description, and the description is what an agent reads to decide when a skill applies. Another 419 have a name that breaks the spec or doesn't match its folder, and 52 have a description over the 1,024-character limit.
• 15% bundle executable scripts, and 50 tell the agent to pipe a download into a shell. 10 of those are among the most-installed skills, led by higgsfield-ai/skills and 101-skills/superpowers. A curl … | sh step runs whatever the URL serves at that moment, with your permissions.
• 6 repositories reuse the name “superpowers” and publish the same skill set. None of their skills come from obra/superpowers, the well-known project with that name. Each has at most 9 GitHub stars, yet skills.sh counts 14,441,771 installs across them. Check the owner before you install a skill by its name.
• 13% of the SKILL.md files we found are byte-identical copies of another skill in the set. We count each body once in the figures above.
• We found no hidden instructions in popular skills. Every prompt-injection match we checked by hand was a security guide quoting an attack, a test fixture, or a rule against acting silently.

Method

We took every skill on the skills.sh all-time leaderboard (the 464 distinct ones we could match carry its install counts), the official anthropics/skills and openai/skills repositories, and the most-starred GitHub repositories tagged agent-skills, claude-skills or claude-code-skills (100+ stars, forks excluded), up to 150 skills per repository, so totals for the largest collections are minimums. For each SKILL.md we parsed the frontmatter and counted tokens with the o200k tokenizer, the same one we use for MCP servers: “always on” is name: description, “once used” is the body. Bundled files come from the repository tree. Spec checks follow agentskills.io. Copies are byte-identical bodies; we keep the official or most-installed one. Exact token counts vary by model tokenizer and by how each client formats its skill list; the comparison holds because both sides use the same tokenizer.

Download CSV (7,755 rows, CC BY 4.0)