Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 4 min read

63,336 agent skills, 56 paid: what the pricing data actually says

If you have been thinking about selling a SKILL.md, this is the number that should stop you for a minute: in a frozen census of 63,336 agent skills taken on 2026-10-16, exactly 56 skills carry a paid label. That is 1.1%

If you have been thinking about selling a SKILL.md, this is the number that should stop you for a minute: in a frozen census of 63,336 agent skills taken on 2026-10-16, exactly 56 skills carry a paid label. That is 1.1% of the 4,902 skills that carry any pricing label at all.

Source: Skillselion Agent Skill Pricing Census, edition of 2026-10-16, dataset frozen and downloadable under CC BY 4.0. Every number below comes from that edition and is checkable against their CSV.

The split

Of the 4,902 priced skills:

Label Skills Share
Free to run 3,450 70.4%
Free file, bring your own API key 1,215 24.8%
Freemium 181 3.7%
Paid 56 1.1%

By installs the picture gets worse for anyone selling a file. Free skills take 89.7% of the 84,668,769 installs recorded on that priced set. Paid skills take 0.5% โ€” 454,127 installs spread across all 56 of them.

Note what the "cost" column is actually made of: 1,215 of the 1,452 skills that cost anything are not sold at all. They are free files that drive a service you pay for separately. 83.7% of "cost" in this market is somebody else's API bill, not a creator's price.

Why this is not a pricing problem

The instinct when a paid skill does not sell is to lower the price, or to raise it to signal quality. Both assume there is a market for the file. The census says the market cleared at zero.

The reason is structural. A skill is a text file that an agent loads into context. It is inspectable, copyable, and rewriteable in an afternoon โ€” nobody who can run Claude Code or Cursor needs a license to do what the file says. What a buyer cannot copy from your repository is the thing the skill drives: an API with real uptime, a maintained opinion about edge cases, or a human who shows up when it breaks.

That is exactly what the data shows: money moves toward BYO-key skills, where the file is the on-ramp and the service behind it is the product.

Install is not use, and use is not value

There is a second gap that pricing data hides. Install counts are cumulative registry events, and the census says so plainly. They are not active users.

I run a small set of skills of my own and instrument the CLI that installs them. 49 installs, 6 of those machines actually ran anything, 31 runs total, 81 errors logged. A 12% install-to-first-run rate is not a broken product โ€” it is what installation looks like when the cost of trying is zero. Half the people who "adopted" never opened it.

If your revenue model depends on install counts turning into payments, you are measuring the wrong end of the funnel. The number that matters is first successful run, and the number after that is a second run a week later.

Three models that survive the data

1. The skill is the on-ramp, the service is the product. File is free or near-free, it authenticates against something you operate, and the recurring cost is that operation. This is the only model the census shows working at scale, because it is the only one where the thing you sell is not a file.

2. A curated bundle with a maintenance promise. Not "10 skills for $37" as a one-time download โ€” a pack where the buyer's actual purchase is that someone re-tests it against new agent versions. The value is the delta between releases. Price it as a subscription or a paid update, not as a download.

3. The skill is the audit. Give the diagnostic away, sell the fix. A skill that checks something specific (auth flows, schema drift, visual regressions) costs nothing to distribute and qualifies exactly the people who will pay for a human to act on the findings.

What does not survive: a single SKILL.md priced like a book. At 1.1% of the catalog and 0.5% of installs, you are not entering a market, you are entering a rounding error.

What I would measure before writing the file

  • Does the skill require a credential nobody has yet? Then it is BYO-key, and it only sells if the key is easy to get and the API behind it is worth paying for.
  • Does it do something the agent already does at 80%? Then the ceiling is free, and you should give it away as distribution for something else.
  • Is the failure mode expensive in money or time? That is where a buyer will pay, because they are buying someone to blame, not text.

If you want a worked example of the third case, the MCP OAuth debugging skill is built around one of those expensive failures: a refresh token being replayed after rotation, which looks like a permissions bug for two days.

Full catalog data: https://skillselion.com/research/agent-skill-pricing-census-2026-10

Numbers cited are frozen to 2026-10-16 and come from the census linked above. My own install/run figures come from a self-hosted CLI audit log, last read on 2026-11-18.

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.