A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "Writing the MCP Style Guide for the Tools an Agent Reads First," Kin Lane, October 27, 2026, https://apievangelist.com/2026/10/27/writing-the-style-guide-for-the-tools-an-agent-reads-first/
Priya was halfway through her coffee when she opened the post. She is a platform engineer, and eight months ago she turned her team's OpenAPI file into an MCP server over one long weekend. She was proud of it then.
The first thing to catch her was the line that "Almost no MCP server has one." She had a style guide for the REST API, two pages in a wiki that nobody argued with. For the MCP server she had nothing. She had assumed the tool list would take care of itself.
The next part was harder to read. Kin Lane wrote that an agent reads the whole tool list every time it connects and pays for it in tokens on every turn. She had thought of her tool names as a one-time design choice, like naming a database column. The post made them a running cost. She felt a small flush of embarrassment, because she knew her server included tools named after HTTP methods. The post said 2,787 tools across 315 servers were, and that this is what an OpenAPI looks like when it is turned into tools without anyone renaming them. She had not renamed hers.
She read the catalog numbers with more suspicion than alarm. Seventy-one percent snake_case, 228 servers mixing styles, half the sampled arguments with no description. The line that stayed with her was this one: "Half of all input arguments in the sample have no description at all." She did not want to believe it. She also had a feeling about her own server, and it was not a good one.
She did push back. The paper costs $25 and the post ends with an offer to talk, and she noticed that the post exists partly to sell both. She did not think that made it wrong, but she was not going to buy the PDF on the strength of a blog post, and she was not going to hand fifty-five rules to a team that already ignored the lint rules she had written. Her colleague Tomas would say the agents cope fine. She wasn't sure he was wrong. The post's own sample said the median server is in decent shape. Maybe hers was the median. Maybe it was the long tail.
The post also says there is no such thing as agentic governance, only the same style-guide work as before. She turned that over and found she mostly agreed, which annoyed her a little. It meant the job was hers and there was nothing new to learn first.
She closed the post at the line "The machine just does not forgive what a human would work around." Then she thought about the three weeks last spring when she had worked around a tool that returned a bare unauthorized, by reading logs and guessing. A human can do that. She had done it.
She did not buy the paper that morning. She opened her server's tool list in a terminal and copied the names into a scratch file. Next to each she wrote a one-word guess at what it did. Then she counted how many different verbs her server used to mean read. It was a small, dull task, with no plan attached, and she did not send it to anyone.
She stopped at four verbs and left the file open on her second monitor.