AgentStatus
We tested every public MCP server we could find. Calling is solved. Delivery is not.
Read the reportMCP standardized how an agent calls a tool. It did not standardize how the agent is supposed to behave once it has one.
For the last year, the way a product reached an agent was a list of tools. A host connects, the server lists what it can do, and the agent calls a function.
Rooms, courts, and catalogs left the website and showed up inside ChatGPT, Claude, Cursor, and Gemini. That was real progress.
It was also an incomplete description of the product.
A tool list is an interface. It is not a way of working. The SaaS era could ignore that, because a person sat in the product and already knew the business.
The agent era cannot. The agent is the user, and it arrives with no muscle memory.
APIs worked because the client was written by someone who had read the contract. An agent does not have that author. It infers.
Inference is how a hotel search becomes a product search, and how a room the tool just named gets described as unavailable.
It is also how a long job gets reported as a failure, because the first call only started it. None of those are transport failures. The server did what it was asked.
The agent did not know the policy. Which call comes first. Which names to repeat. What to say when nothing matched, and what a start actually means.
In the old model, that policy lived in the UI, in training, and in the habits of the people who used the product every day.
Strip the UI away and the policy has to be written down, or every host reconstructs it.
The ecosystem's answer has been the skill. A short markdown file with a name, a description, and the instructions for the job.
That is the right abstraction. A skill is the standard operating procedure, made legible to a model. People write them because a tool list is not enough.
The cost of a wrong inference is silent. The agent does not file a ticket. It gives a worse answer and moves on.
The failure is distribution. Most skills sit beside the agent, in a repo or a folder owned by whoever connected the server.
The company that owns the product does not own the instructions its agents follow. Every new host starts blank. When the tools change, the skill does not.
Two customers end up maintaining two private versions of how the same business works.
That is the fragmentation early APIs had before a shared client contract. Every integrator writes a private adapter, and the adapter rots.
The vendor cannot see it, cannot update it, and cannot tell a new host what the last host learned.
If agents are a real distribution channel, the behavior has to ship with the product.
The person who owns the catalog should own the note.
A host that has never seen the server should get the current instructions on the first connection, the same way it already receives the tools.
Starting today, every MCP we host carries one skill. One file, written from the tools that server already has, and from nothing else.
A hotel skill talks about rooms and starts with the room search. A court skill talks about courts and names a court that server actually stores.
A store skill is not pasted onto a hotel. There is no second catalog, and no skill library sitting beside the product.
When an agent connects, the server says it has a skill. A host that knows how to ask can list it and read the file.
The file says which tool to start with, what to repeat, and what to admit when nothing matched. Some hosts never ask.
Those hosts still get a short version in the instructions we send at connection time. Missing a skills client should not mean a blank page.
The owner can replace the file. Upload a SKILL.md whose name matches that server.
Your wording becomes the note agents read. Remove it, and the generated skill returns.
That is the whole control. The server stays the source of truth for the tools and for the policy.
Tool discovery was the first step. It made a product callable. A server-owned skill is the second.
It makes the product behave the same way on the tenth host as on the first.
Nobody has to re-author the business for each new agent. When you change a tool, the note can change with it. A new agent does not inherit someone else's stale copy.
It is already on the MCPs we host. Same subscription. Nothing to install.
If the note we wrote is close enough, leave it. If you want it tighter, replace the file.
Open yours on the dashboard, or read how the platform hosts the server the skill sits on.
The tools already travel. The instructions should travel with them.
One skill per hosted server, written from the tools that server already has. Read it, or replace the file.