Codex can now read skills directly from MCP servers. The tools and the instructions that go with them can now live on the same MCP server. Teams no longer have to install the same skill on every machine and maintain each copy separately. Before, teams had to copy SKILL.md to every machine by hand. As soon as a tool changed, those local copies started falling out of sync. The read and update flow: → skills/list: shows which skills are available on the server → skills/get: returns the selected skill’s metadata, file list, and digest → resources/read: loads only the files it needs → Updates: after a refresh, Codex sees the new digest and fetches the updated files That lets Codex read the name and description first, then load the actual files only when a task needs them, without stuffing every skill into the context window at once.