alt Hacker News> My biggest complaint right now is that OpenAI has still refused to implement the MCP Prompts spec
MCP servers provide three things: tools, resources, and prompts. Of these, tools seem to be the only part implemented consistently across major clients like ChatGPT, Claude.ai, Claude Code, etc.
For prompts and resources, there doesn't seem to be a common understanding of how clients are supposed to consume them.
For example, if an MCP server exposes resources, Claude Code can discover them and consume them when needed without you explicitly asking for a specific resource. Claude.ai behaves differently. It doesn't automatically discover and consume those resources. Instead, it gives you a way to manually add an MCP resource to the prompt.
So while MCP defines tools, resources, and prompts at the protocol level, the actual user experience for resources and prompts varies quite a bit across clients.
Replies
In practice, the problem with resources is that many resource collections are too large to list exhaustivley, and if that's the case you will need to need to implement a proper search tool anyways (as the Completions utility isn't a good fit), at which point there is little use in also implementing all of that as a resource, rather than `list_`,`search_`,`get_` for a resource.
The idea of "model controlled", "application controlled" and "user controlled" for tool, resources and prompts (respectively) was aligned with the chat interface. It breaks for the autonomous agent paradigm where the agency is the user and the lines are blurred. Unfortunately MCP has been mostly relegated to tool calling leaving potentially powerful capabilities on the table due to lac of client support for them.
I disagree on Prompts since virtually all of the mainstream harnesses implement them: Cursor, Claude, OpenCode, Copilot. Prompts are very clearly just a remotely delivered `/` command and it is easy to see why this is really powerful (single entry point, no need to update/sync skills, dynamic sets by audience, dynamic construction of the payload by audience, etc). For all intents and purposes, it should be viewed as an analog to local, text-only skills.
Codex is the only mainstream harness that does not implement this in the client.