[BUG] MCP Server Tools Not Exposed to AI 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 servers successfully connect and show as "✓ connected" in the /mcp interface, but their tools are never made available to the AI assistant. The AI cannot see or use any tools provided by connected MCP servers.

What Should Happen?

Summary

MCP servers successfully connect and show as "✓ connected" in the /mcp interface, but their tools are never made available to the AI assistant. The AI cannot see or use any tools provided by connected MCP servers.

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce

1. Configure MCP servers in mcp.json:

``json
{
"mcpServers": {
"mie": {
"command": "wsl",
"args": [
"bash",
"-c",
"export MIE_CONFIG_PATH='/mnt/c/Users/NNNNN/OneDrive/VSCode/Clients/project-workspace/.mie/config.yaml' && mie --mcp"
],
"env": {
"MIE_CONFIG_PATH": "/mnt/c/Users/NNNNN/OneDrive/VSCode/Clients/project-workspace/.mie/config.yaml"
}
},
"microsoft-learn": {
"type": "http",
"url": "https://learn.microsoft.com/api/mcp"
}
}
}
``

1. Restart Claude Code / Reload VSCode window
2. Run /mcp command - servers show as "✓ connected"
3. Ask the AI "What MCP servers do you have access to?"

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

v2.1.34

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

MCP Server Tools Not Exposed to AI Conversation Context

Summary

MCP servers successfully connect and show as "✓ connected" in the /mcp interface, but their tools are never made available to the AI assistant. The AI cannot see or use any tools provided by connected MCP servers.

Environment

• Claude Code Version: v2.1.34
• VSCode Version: 1.109.0
• OS: Windows 11
• Platform: VSCode Extension

Steps to Reproduce

1. Configure MCP servers in mcp.json:

``json
{
"mcpServers": {
"mie": {
"command": "wsl",
"args": [
"bash",
"-c",
"export MIE_CONFIG_PATH='/mnt/c/Users/nnnn/OneDrive/VSCode/Clients/project-workspace/.mie/config.yaml' && mie --mcp"
],
"env": {
"MIE_CONFIG_PATH": "/mnt/c/Users/nnnn/OneDrive/VSCode/Clients/project-workspace/.mie/config.yaml"
}
},
"microsoft-learn": {
"type": "http",
"url": "https://learn.microsoft.com/api/mcp"
}
}
}
``

1. Restart Claude Code / Reload VSCode window
2. Run /mcp command - servers show as "✓ connected"
3. Ask the AI "What MCP servers do you have access to?"

Expected Behavior

The AI should have access to:

• Mie MCP Server: 9 tools (semantic search, graph traversal, fact storage, etc.)
• Microsoft Learn MCP Server: 3 tools (doc search, page fetch, code samples)

Actual Behavior

The AI reports it can only see standard Claude Code built-in tools:

• Task, TaskOutput, TaskStop
• Bash, Glob, Grep, Read, Edit, Write, NotebookEdit
• WebFetch, WebSearch
• TodoWrite, AskUserQuestion
• Skill, EnterPlanMode, ExitPlanMode

Zero MCP tools are visible to the AI, despite servers being connected.

Verification Steps Taken

1. Verified Mie MCP Server Works

wsl bash -c 'export MIE_CONFIG_PATH="/mnt/c/Users/nnnn/OneDrive/VSCode/Clients/project-workspace/.mie/config.yaml" && mie --version'
# Output: mie version v0.1.2

wsl bash -c 'export MIE_CONFIG_PATH="/mnt/c/Users/nnnn/OneDrive/VSCode/Clients/project-workspace/.mie/config.yaml" && mie status'
# Successfully shows memory graph status

2. Tested Both Local and Remote MCP Servers

• Local: Mie via WSL (command-based)
• Remote: Microsoft Learn via HTTP endpoint

Neither type exposes tools to the AI.

3. Multiple Restarts

• Reloaded VSCode window multiple times
• Restarted Claude Code extension
• Servers reconnect successfully each time

4. Configuration Iterations

• Tried different args formatting (inline env vars vs explicit export)
• Added env section to configuration
• Used both WSL and HTTP connection types

Impact

This makes MCP servers completely non-functional in Claude Code. While servers can connect at the system level, the AI cannot use any of their capabilities. This defeats the entire purpose of MCP integration.

Screenshots

<img width="719" height="770" alt="Image" src="https://github.com/user-attachments/assets/188d5322-aab1-4b1a-ba43-579d0ee7d41c" />

<img width="1024" height="768" alt="Image" src="https://github.com/user-attachments/assets/7bfafe1a-d66a-43d0-8177-859d11baa228" />

<img width="1024" height="768" alt="Image" src="https://github.com/user-attachments/assets/2d7f1d43-9f79-40e2-98ff-316abadbde08" />

<img width="746" height="487" alt="Image" src="https://github.com/user-attachments/assets/ad129c05-5265-418e-a0f0-d0f1eab54be6" />

Additional Context

• The AI explicitly states in tool descriptions that "If an MCP-provided web fetch tool is available, prefer using that tool" - suggesting MCP tools should appear in its toolkit
• The /mcp interface works correctly and shows server status
• The issue appears to be in the bridge between MCP server connections and the AI's conversation context
• This affects all MCP servers regardless of type (local/remote) or implementation

Possible Root Cause

There appears to be a missing integration layer that should:

1. Query connected MCP servers for their available tools
2. Add those tools to the AI's callable function list
3. Route tool calls from the AI to the appropriate MCP server

The first two steps appear to be missing or broken.