[BUG] MCP Prompts are not visible to Claude in conversation 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 prompts are successfully discovered and can be invoked as slash commands by users, but they are not visible to Claude in the conversation context. This prevents Claude from proactively suggesting or using MCP prompts, unlike MCP tools which are visible.

• MCP prompts appear as slash commands to users (e.g., /mcp__my-mcp-server__my_prompt)
• Users can successfully invoke these prompts
• The prompt content is correctly injected into the conversation
• However, Claude cannot see these prompts in its available capabilities

What Should Happen?

MCP prompts should be visible to Claude in the same way that:

• MCP tools are visible (they appear as mcp__servername__toolname in Claude's available tools)
• Regular slash commands are visible (they appear in <available_skills>)

This would allow Claude to:

1. Suggest using relevant MCP prompts when appropriate
2. List available MCP prompts when asked
3. Proactively incorporate them into workflows

Error Messages/Logs

Steps to Reproduce

1. Create an MCP server that exposes both tools and prompts:

McpServer.sync(transportProvider)
      .serverInfo("my-mcp-server", "0.0.1")
      .capabilities(McpSchema.ServerCapabilities.builder()
              .tools(true)
              .prompts(true)  // Prompts capability enabled
              .build())
      .tools(syncToolSpecification)
      .prompts(getSyncPromptSpecification())  // Prompt registered
      .build();

1. Configure the server in .claude.json:

{
    "mcpServers": {
      "my-mcp-server": {
        "command": "java",
        "args": ["-jar", "/path/to/server.jar"]
      }
    }
  }

1. Start Claude Code and verify:

• MCP tools appear in Claude's context (✅ Working)
• MCP prompts do NOT appear in Claude's context (❌ Not working)
• User can manually invoke the prompt via slash command (✅ Working)

Verification

• Tested with MCP inspector: prompts are correctly exposed by the server
• MCP tools from the same server are visible to Claude
• The prompt executes successfully when invoked by user

Impact

This limits Claude's ability to leverage MCP prompts effectively. Since Claude cannot see what prompts are available, it cannot:

• Recommend their use in relevant contexts
• Build workflows that incorporate them
• Inform users about available prompt capabilities

Environment

• Claude Code: 2.0.33
• OS: Ubuntu 24.04.3 LTS
• MCP Server Implementation: Java-based using io.modelcontextprotocol SDK

Related Documentation

• https://docs.claude.com/en/docs/claude-code/mcp states: "MCP prompts are dynamically discovered from connected servers"
• However, "discovered" appears to only mean "available to users as slash commands," not "visible to Claude"

Claude Model

None

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.0.33

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Xterm

Additional Information

_No response_