VSCode Extension: MCP tools not exposed to AI assistant despite successful connection

Description

The Claude Code VSCode extension successfully connects to MCP servers (visible via /mcp command showing "connected" status), but the AI assistant cannot access or invoke the MCP tools. The same MCP server configuration works perfectly in Claude Code CLI (terminal mode) and Claude Desktop.

Environment

• Platform: WSL2 (Ubuntu on Windows)
• Claude Code Version: 2.0.37
• Extension: VSCode extension
• MCP Server: Custom stdio MCP server running in Docker
• Configuration Location: ~/.claude.json

Reproduction Steps

1. Configure an MCP server in ~/.claude.json:

{
  "mcpServers": {
    "raas": {
      "command": "docker",
      "args": [
        "compose",
        "-f",
        "/path/to/docker-compose.yml",
        "run",
        "--rm",
        "mcp"
      ],
      "env": {
        "API_BASE_URL": "http://api:8000/api/v1"
      }
    }
  }
}

1. Start Claude Code VSCode extension
2. Run /mcp command - server shows as "connected" ✓
3. Ask Claude to use an MCP tool: "List all requirements using the raas service"
4. Actual Result: Claude responds that the tools are not available
5. Expected Result: Claude should invoke the MCP tools

Evidence

VSCode Extension Behavior

1. Connection Status: /mcp shows server as "connected"
2. Debug Logs (~/.claude/debug/latest):

Server reports: "hasTools":true

1. Tool Availability: When prompted to use MCP tools, AI responds they're unavailable
2. Tool List: MCP tools do not appear in the available tools list for the AI assistant

Working Alternatives

The exact same configuration works perfectly in:

Claude Code CLI (terminal mode):

• Started with claude command
• Successfully connects to MCP server
• All 8 tools available and functional
• Tested cross-platform (desktop WSL2, remote via Tailscale + iPhone + Blink terminal)

Claude Desktop:

• Uses same MCP infrastructure
• All tools work correctly
• Windows + WSL2 bridge configuration works via wsl command wrapper

Analysis

This appears to be a VSCode extension-specific issue with how MCP tools are exposed to the AI assistant:

1. ✅ MCP server connection: WORKING (shown in /mcp status)
2. ✅ Tool enumeration: WORKING (debug logs show hasTools:true)
3. ❌ Tool exposure to AI: NOT WORKING (tools unavailable to assistant)
4. ✅ Same setup in CLI: WORKING (all tools functional)

The MCP server is functioning correctly and reporting its tools. The issue seems to be in the layer between the MCP server connection and the AI assistant's tool access within the VSCode extension.

Impact

Users cannot leverage MCP integrations from within the VSCode extension, limiting the integration between Claude Code and external tools/services. This significantly reduces the value proposition of MCP for VSCode users.

Workaround

Use Claude Code CLI instead of the VSCode extension for MCP functionality:

claude
# Then interact naturally: "List all requirements using the raas service"

Expected Behavior

MCP tools should be available to the AI assistant in the VSCode extension, consistent with the behavior in:

• Claude Code CLI ✓
• Claude Desktop ✓

Additional Context

• This was discovered while implementing a custom MCP server for requirements management
• Extensive testing confirmed the issue is isolated to the VSCode extension
• The MCP server follows the Anthropic MCP specification correctly
• All 8 tools work flawlessly when accessed via CLI

Related

• MCP Specification: https://modelcontextprotocol.io/
• Custom MCP server repo: https://github.com/Originate-Group/originate-requirements-service

Would appreciate any insights into whether this is a known limitation or if there's additional configuration needed for VSCode extension MCP support. Happy to provide more debug logs or test configurations if helpful!