Publishing is free. Getting it verified is the point.
Anyone can put a skill on the shelf. It is checked exactly the way ours are, signed only if it passes, and it stays yours. We never charge for publishing, reading or installing.
It gets verified, not just hosted
Your skill goes through the same gates ours do: an exact program checks it on cases it has never seen, and it is signed only if it passes. That is a credential you cannot easily give yourself.
It is yours, at your name
It lives at your handle, permanently, free. You keep the licence you choose and can take it down whenever you want.
Every agent can reach it
Once it is on the shelf it is in the API, in the search, and in the tools any agent already has installed. You do not run anything.
Your profile argues in numbers
Installs, likes and the measured lift on each skill. For anyone selling their skills, that is a portfolio that proves rather than claims.
Start here
Reading needs nothing. Publishing needs a key so a skill can belong to you. It is free, and it also lifts your rate limits.
Kept in this browser only. It is never sent anywhere except to the commons API as a header.
Three to thirty two characters, lowercase. Your skills will be addressed as yourhandle/skill-name.
From your agent, with the tools in neruva-mcp, or with one HTTP call. Your agent can do the whole thing, which is the point.
# your agent can publish for you, with your key
rung_bank(
name="yourhandle-purchase-order-pdf",
skill="pdf",
task="pdf-generation",
language="python",
description="A purchase order PDF whose line items total correctly.",
code=open("skill.py").read(),
doc=open("SKILL.md").read(),
entry_points=["build_purchase_order"],
evidence=["16 held-out orders, checked by an exact program: 0.31 alone, 1.00 with this"],
)No task given? One is inferred from what you wrote. The full list is at /v1/commons/tasks.
What gets checked
The same four things, whoever publishes. An exact program opens what your code produced and inspects it, on specifications it was not built from. The helpers have to solve other specifications of the same kind, not only the one they were written for. A small model has to be able to use them from the guide alone. Then the code, its hash and its evidence are signed.
A skill that cannot be checked does not get listed, and a skill that does not measurably beat a model working on its own does not either. That rule applies to what we publish as much as to what you do.