[BUG] MCP prompts invisible to agents — prompts/list and prompts/get not surfaced in agentic context

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

MCP defines three core primitives: resources, tools, and prompts. The MCP spec requires
clients to call prompts/list at connection time and make prompts accessible. Claude Code does
this for the human REPL (exposing prompts as /mcp__server__promptname slash commands) but
not for agents. In an agentic session, there is no ListMcpPromptsTool or
ReadMcpPromptTool — prompts are completely invisible.

As a result, agents systematically ignore guidance that server authors have correctly placed in
MCP prompts, exactly where the spec says behavioral guidance belongs.

What Should Happen?

Agents should be able to discover and read MCP prompts, just as they can discover and read
resources via ListMcpResourcesTool / ReadMcpResourceTool. Ideally, registered prompts from
connected MCP servers should be automatically applied as system context at session start — the
intended MCP workflow. At minimum, agents need ListMcpPromptsTool and ReadMcpPromptTool.

Steps to Reproduce

1. Create an MCP server that registers a prompt with behavioral guidance (e.g., via FastMCP

@mcp.prompt("my-guide")):
``python
@mcp.prompt("usage-guide")
def usage_guide() -> str:
return "IMPORTANT: always include partition key h0 in joins..."
``

1. Connect Claude Code to this server (via .mcp.json)
2. Start an agentic session and ask Claude to use the server's tools
3. Observe: Claude never discovers or reads the usage-guide prompt; the guidance is ignored

Claude Code Version

_(fill in: claude --version)_

Platform

Anthropic API

Operating System

Linux (also reproducible on macOS)

Terminal/Shell

Non-interactive/CI environment (agentic)

Additional Information

A compounding issue: server authors work around the missing prompt support by embedding guidance
in tool descriptions. This also fails in Claude Code because tool schemas are deferred —
descriptions are not loaded until ToolSearch is called for that specific tool. An agent
planning work without yet executing a tool never fetches the schema. Both the spec-correct
mechanism (prompts) and the workaround (tool descriptions) are broken for autonomous agents.

Related issues:

• #30738 — MCP prompts not surfaced as slash commands in VS Code (same gap, different surface)
• #31893 — broader MCP spec compliance gaps (prompts/list_changed, etc.)

MCP spec — Prompts: https://modelcontextprotocol.io/docs/concepts/prompts