An OWASP MCP proposal meets a simpler answer: start with a skill
A proposal to feed current security guidance into coding assistants prompted project leaders to favor a small skill and better search before adding server code.
A proposal for an official OWASP Cheat Sheet MCP server has prompted a practical counterproposal from the project’s own leaders: start with a skill and good search. The October 21 exchange is a small but concrete example of maintainers questioning how much new infrastructure coding agents actually need.
Jakub Mackowski, posting as mackowski, proposed a read-only Model Context Protocol server so assistants such as Claude, Copilot and Cursor could retrieve maintained security guidance rather than depend on what a model remembers from training. The suggested first version would list cheat sheets, fetch a sheet or section, and search relevant passages with source links.
The proposal favors original passages over model-written summaries, heading-based chunks and basic full-text or BM25 retrieval before embeddings. Responses would carry a title, canonical URL and the content’s CC BY-SA 4.0 license. Local stdio would come first; a hosted HTTP transport is a possible later step.
The maintainers’ smaller starting point
Gabriel Corona, randomstuff, asked whether a simple SKILL.md plus good website search, and possibly a search API, would work instead. That route could also help visitors who are not using agents. Jim Manico agreed with starting with a skill, writing that he wanted to “minimize code infrastructure here that could get us in trouble.”
Mackowski replied that a good skill could work and was closer to the project’s work and easier to maintain. He also noted that his proposed local MCP server would not itself require hosted infrastructure. The distinction matters: the discussion is about additional code and maintenance as well as operating a service.
The issue remained open when reviewed. These comments establish a preference to explore a smaller approach, not a finalized project decision or an available official skill. There is no released package, installation command, compatibility matrix or new paid service established by this thread.
Read-only still needs a trust boundary
The proposal explicitly recognizes that community-edited Markdown returned as tool output can become an indirect prompt-injection surface. It identifies pull-request review as a trust boundary and says responses should be marked as data. A read-only server would avoid action tools, but would not make all retrieved text safe instructions for an agent to follow.
For developers choosing between a skill and a new MCP service, the exchange supplies a useful design question: can source discovery and attribution solve the immediate problem before another executable integration is added? It does not establish that either approach has already improved coding-assistant security.