MCP servers configured but tools not exposed to AI assistant in Copilot CLI

[chinlung]

GitHub Copilot MCP Integration Issue Report

Date: 2025-10-03
Version: GitHub Copilot CLI 0.0.333
Platform: macOS (Darwin)

---

🐛 Issue Summary

MCP servers are configured but their tools are not exposed to the AI assistant in GitHub Copilot CLI sessions.

---

📋 Environment Details

MCP Configuration File Location

~/.copilot/mcp-config.json

Configured MCP Servers (3 total)

1. Serena

• Type: local
• Command: uvx
• Args: --from git+https://github.com/oraios/serena serena start-mcp-server
• Tools: * (all tools enabled)
• Status: Can be manually started ✅
• Tools Exposed to AI: ❌ None visible

2. Context7

• Type: local
• Command: npx
• Args: -y @upstash/context7-mcp@latest
• Tools: * (all tools enabled)
• Status: Unknown
• Tools Exposed to AI: ❌ None visible

3. Sequential Thinking

• Type: local
• Command: npx
• Args: -y @modelcontextprotocol/server-sequential-thinking
• Tools: * (all tools enabled)
• Status: Process running ✅
• Tools Exposed to AI: ❌ None visible

---

🔍 What We Observed

Tools Currently Visible to AI Assistant

The AI assistant can see and use:

• ✅ bash - Execute shell commands
• ✅ str_replace_editor - File editing operations
• ✅ github-mcp-server-* - 30+ GitHub-related tools (e.g., get_pull_request, search_code, etc.)

Tools NOT Visible Despite Configuration

Despite proper MCP configuration, the AI cannot see:

• ❌ Any serena-* tools (26 tools available in Serena)
• ❌ Any context7-* tools
• ❌ Any sequential-thinking-* tools

---

🧪 Testing Performed

1. Verified MCP Configuration File

$ cat ~/.copilot/mcp-config.json
# Configuration is valid JSON with correct structure

2. Verified Serena Can Start

$ uvx --from git+https://github.com/oraios/serena serena start-mcp-server
INFO  Starting Serena server (version=0.1.4)
INFO  Available projects: [PROJECT_NAME]
INFO  Loaded tools (36): read_file, create_text_file, find_symbol, ...
INFO  Number of exposed tools: 26
INFO  Serena web dashboard at http://127.0.0.1:24282/dashboard/index.html
# Server starts successfully ✅

3. Verified Serena Tools List

$ uvx --from git+https://github.com/oraios/serena serena tools list
# Returns 26 tools including:
- activate_project
- find_symbol
- find_referencing_symbols
- read_file
- create_text_file
- write_memory
- execute_shell_command
- replace_regex
- get_symbols_overview
- insert_after_symbol
- insert_before_symbol
- list_memories
- onboarding
- think_about_collected_information
- think_about_task_adherence
- think_about_whether_you_are_done
# ... and more

4. Checked Running Processes

$ ps aux | grep -E "(serena|context7|sequential-thinking|mcp)"
# Sequential Thinking is running ✅
# Serena was not running initially (can be started manually)

5. Verified Environment

$ which uvx
# uvx is available in PATH
$ uvx --version
uvx 0.8.13

$ which npx
# npx is available in PATH
$ npx --version
10.9.2

---

🎯 Expected Behavior

When MCP servers are configured in ~/.copilot/mcp-config.json, the AI assistant should:

1. ✅ Automatically connect to configured MCP servers
2. ✅ Expose MCP server tools in the AI's tool list
3. ✅ Allow the AI to invoke these tools when appropriate
4. ✅ Show tool names prefixed with server name (e.g., serena-read_file)

---

❌ Actual Behavior

1. ❌ MCP servers are configured but not connected
2. ❌ MCP tools are not visible in AI's tool list
3. ❌ AI cannot invoke MCP tools even when explicitly requested
4. ❌ No error messages or warnings about MCP connection failures

---

💬 User Experience Impact

What the User Experienced

1. User installed and configured 3 MCP servers
2. User asked AI to "use Serena to analyze the project"
3. AI performed analysis using only built-in tools (bash, str_replace_editor)
4. AI did NOT use Serena despite explicit request
5. User questioned why Serena wasn't used
6. Investigation revealed tools are not exposed to AI

User's Confusion

• ✅ Configuration file exists and is valid
• ✅ MCP servers can be started manually
• ✅ MCP servers report their tools correctly
• ❌ But AI cannot see or use these tools
• ❓ No feedback to user about what went wrong

---

🔧 Reproduction Steps

1. Install MCP servers:

   # Install uvx (for Serena)
   pip install uv
   
   # Install npx (for Context7 and Sequential Thinking)
   # Usually comes with Node.js

2. Configure MCP in ~/.copilot/mcp-config.json:

   {
     "mcpServers": {
       "serena": {
         "tools": ["*"],
         "type": "local",
         "command": "uvx",
         "args": ["--from git+https://github.com/oraios/serena serena start-mcp-server"]
       },
       "context7": {
         "tools": ["*"],
         "type": "local",
         "command": "npx",
         "args": ["-y @upstash/context7-mcp@latest"]
       },
       "sequential-thinking": {
         "type": "local",
         "command": "npx",
         "tools": ["*"],
         "args": ["-y @modelcontextprotocol/server-sequential-thinking"]
       }
     }
   }

3. Start a GitHub Copilot CLI session

4. Ask AI to list available tools or explicitly use MCP tools

5. Observe that MCP tools are not available

---

💡 Suggested Solutions

For GitHub Team

1. Add MCP Connection Diagnostics

   • Show MCP server connection status in Copilot output
   • Provide clear error messages when MCP servers fail to connect
   • Add a command like gh copilot mcp status to check configuration

2. Auto-start MCP Servers

   • Automatically start configured MCP servers when Copilot CLI initializes
   • Show startup progress/failures
   • Provide logs for debugging

3. Better Documentation

   • Document how MCP integration works in Copilot CLI
   • Explain expected tool naming conventions
   • Provide troubleshooting guide
   • Show examples of working configurations

4. Tool Visibility

   • Add command to list all available tools: gh copilot tools list
   • Show which tools come from which MCP server
   • Indicate tool source (built-in vs MCP server)

5. Error Handling

   • Show warnings if MCP servers are configured but not running
   • Provide helpful error messages with resolution steps
   • Log MCP connection attempts and failures

For Users (Workarounds)

1. Manual Server Start

   • Start MCP servers manually before using Copilot
   • Use Serena's web dashboard for direct access: http://127.0.0.1:24282/dashboard/

2. Use Built-in Tools

   • Continue using bash and str_replace_editor for now
   • Wait for MCP integration improvements

3. Check Logs

   • Look for Copilot CLI logs
   • Check MCP server logs (e.g., Serena logs in ~/.serena/logs/)

---

📊 System Information

OS: macOS (Darwin)
GitHub Copilot CLI Version: 0.0.333
MCP Config: ~/.copilot/mcp-config.json

MCP Servers Configured: 3
- Serena (uvx) - v0.1.4
- Context7 (npx)
- Sequential Thinking (npx)

Tools Visible to AI: ~35
- bash, str_replace_editor, write_bash, read_bash, stop_bash
- github-mcp-server-* (30+ tools)

Tools NOT Visible: All MCP server tools
- serena-* (26 tools expected)
- context7-* (unknown count)
- sequential-thinking-* (unknown count)

---

📎 Additional Context

Why This Matters

MCP (Model Context Protocol) is designed to extend AI capabilities with specialized tools. In this case:

• Serena: Provides advanced code navigation, symbol search, project understanding, and intelligent code editing
• Context7: Offers semantic memory and vector search capabilities
• Sequential Thinking: Enables structured reasoning and step-by-step problem solving

Without these tools working, users lose significant value from their MCP investment and cannot leverage advanced AI capabilities.

What We Tried

1. ✅ Verified configuration syntax (valid JSON)
2. ✅ Manually started Serena (works correctly)
3. ✅ Checked Serena tools list (26 tools available)
4. ✅ Verified executables (uvx, npx) are in PATH
5. ✅ Confirmed Sequential Thinking process is running
6. ❌ Could not find way to connect MCP to Copilot session
7. ❌ Could not find diagnostics for MCP connection issues
8. ❌ No error logs or warnings about MCP integration

GitHub MCP Server Works

Notably, the github-mcp-server-* tools ARE visible and working, which suggests:

• The MCP infrastructure exists in Copilot CLI
• GitHub's own MCP server is properly integrated
• Third-party MCP servers may need different integration approach
• Or there's a configuration/discovery issue with third-party servers

---

🙏 Request

Please investigate and fix the MCP tool integration in GitHub Copilot CLI, or provide:

1. ✅ Clear documentation on how MCP integration works
2. ✅ Diagnostic commands to troubleshoot issues
3. ✅ Error messages when MCP servers fail to connect
4. ✅ Expected behavior and known limitations
5. ✅ Examples of working MCP configurations
6. ✅ Troubleshooting guide for common issues

---

📧 How to Submit

Option 1: GitHub Issues

• Repository: https://github.com/github/copilot-issues
• Category: GitHub Copilot CLI
• Label: MCP Integration

Option 2: IDE/Editor Feedback

• Use the feedback mechanism in your IDE/editor
• Mention "MCP tool integration issue"

Option 3: Community Forum

• GitHub Community Discussions
• Tag: copilot-cli, mcp

---

🔗 Related Links

• MCP Specification: https://modelcontextprotocol.io/
• Serena GitHub: https://github.com/oraios/serena
• Context7 MCP: https://www.npmjs.com/package/@upstash/context7-mcp
• Sequential Thinking: https://www.npmjs.com/package/@modelcontextprotocol/server-sequential-thinking

---

Generated: 2025-10-03
Report Version: 1.0 (Anonymized)

[cioannid]

I'm also experiencing the very same behavior with Figma's remote MCP server. The MCP configuration is recognized in Copilot CLI, but it isn't connected to the MCP server at all.

[kiliancm94]

@chinlung There is an issue with the args in your mcp-config.json defintions, you have to split args with commas ",".
For me it works:

There is an issue in your mcp-config.json file:

  {
    "mcpServers": {
      "serena": {
        "tools": ["*"],
        "type": "local",
        "command": "uvx",
        "args": ["--from", "git+https://github.com/oraios/serena", "serena", "start-mcp-server"]  # --> here
      }
    }
  }

[PradeepSanTmeic]

Having very similar issues with mcp server connections (which are currently otherwise working in claude-code2.0, copilot in vs code etc), running windows 11

[svnscha]

I experience the same issue, e.g:

● Logged in as user: svnscha

 ● Configured MCP servers: mcp-windbg

 ● Connected to GitHub MCP Server

 ╭────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
 │ > Use the windbg mcp server to list dumps                                                                          │
 ╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

 ● I don't have access to a WinDbg MCP server tool in my available tools. The tools
   I have access to are primarily for GitHub operations, file editing, and
   PowerShell command execution.

   If you have a WinDbg MCP server set up, you could use PowerShell to interact
   with it directly. Would you like me to help you run a PowerShell command to list
   dumps, or do you need help setting up access to the WinDbg MCP server?

 ● MCP Server Configuration:

   • mcp-windbg (local): Command: python

   Total servers: 1
   Config file: C:\Users\svnscha\.copilot\mcp-config.json

 ● MCP server "mcp-windbg" is already enabled.

[castlebbs]

I have the same issue.

% copilot --allow-all-tools
Welcome to GitHub Copilot CLI
Version 0.0.339 · Commit 66f4dfe

Copilot can write, test and debug code right from your terminal. Describe a task to get started or enter ? for help. Copilot uses AI, check for mistakes.

● Logged in as user: castlebbs

● Configured MCP servers: SecurityArchitect-AI

● Connected to GitHub MCP Server

│ > Use the SecurityArchitect-AI mcp server to list the current threats

● I don't have access to the SecurityArchitect-AI MCP server in my available
tools. The tools I have access to are primarily for GitHub operations, file
management, and bash commands.

To use the SecurityArchitect-AI MCP server, you would need to ensure it's
properly configured and connected. Could you verify that the MCP server is
running and configured in your environment?

[EvanBoyle]

Hey folks, thanks for the several reports and repros here. We are investigating this issue as a part of a push on MCP quality and will report back.

[chinlung]

@kiliancm94 Thank you for your response, but I tested it and found it ineffective for me. 🥹

^^^^^
I will correct my previous response. I just reviewed it again, and it wasn't set up properly before. Your settings can be effective now. Thank you! 👍

[phamminhtuan1999]

This config is working for me:

{
  "mcpServers": {
    "serena": {
      "tools": ["*"],
      "type": "local",
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/oraios/serena",
        "serena",
        "start-mcp-server",
        "--context",
        "ide-assistant"
      ]
    }
  }
}

[chinlung]

This config is working for me:

{
  "mcpServers": {
    "serena": {
      "tools": ["*"],
      "type": "local",
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/oraios/serena",
        "serena",
        "start-mcp-server",
        "--context",
        "ide-assistant"
      ]
    }
  }
}

Thank you for your reply. I also succeeded, and other similar MCPs configured in this way have also worked.

Here are my settings and the response from Copilot.

{
  "mcpServers": {
    "serena": {
      "tools": [
        "*"
      ],
      "type": "local",
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/oraios/serena",
        "serena",
        "start-mcp-server",
        "--context",
        "ide-assistant"
      ]
    },
    "context7": {
      "tools": [
        "*"
      ],
      "type": "local",
      "command": "npx",
      "args": [
        "-y",
        "@upstash/context7-mcp@latest"
      ]
    },
    "sequential-thinking": {
      "type": "local",
      "command": "npx",
      "tools": [
        "*"
      ],
      "args": [
        "-y",
        "@modelcontextprotocol/server-sequential-thinking"
      ]
    }
  }
}

What MCP tools are currently enabled?

● Based on your environment configuration, the currently enabled MCP tools are:

Core Development Tools:

 - bash, write_bash, read_bash, stop_bash - Command execution and interaction

File Operations:

 - str_replace_editor - View, create, and edit files

Code Analysis (Serena):

 - serena-list_dir, serena-find_file, serena-search_for_pattern - File exploration
 - serena-get_symbols_overview, serena-find_symbol, serena-find_referencing_symbols - Code navigation
 - serena-replace_symbol_body, serena-insert_after_symbol, serena-insert_before_symbol - Code editing
 - serena-write_memory, serena-read_memory, serena-list_memories, serena-delete_memory - Project memory
 - serena-activate_project, serena-get_current_config, serena-check_onboarding_performed - Project management
 - serena-think_about_collected_information, serena-think_about_task_adherence, serena-think_about_whether_you_are_done - Reflection tools

GitHub Integration:

 - Repository search, code search, issue/PR management
 - Workflow runs, commits, branches, releases, tags
 - Code/secret scanning alerts, file contents
 - github-mcp-server-web_search - AI-powered web search

Documentation (Context7):

 - context7-resolve-library-id, context7-get-library-docs - Library documentation lookup

Reasoning:

 - sequential-thinking-sequentialthinking - Multi-step problem solving

[asw101]

This looks like a similar issue to #117 which can happen when multiple arguments are provided without commas between them at the time of configuration, so they are not split out to an array in args: [], which you can see in the first example. This can mean the MCP server itself does not run correctly, and therefore will not expose tools.

[summer88123]

{
  "mcpServers": {
    "gitlab": {
      "type": "local",
      "command": "npx",
      "tools": [
        "*"
      ],
      "args": [
        "-y",
        "@zereight/mcp-gitlab"
      ],
      "env": {
        "GITLAB_PERSONAL_ACCESS_TOKEN": "my-access-token",
        "GITLAB_API_URL": "https://my-gitlab-api-url/api/v4",
        "GITLAB_READ_ONLY_MODE": "false",
        "USE_GITLAB_WIKI": "false",
        "USE_MILESTONE": "false",
        "USE_PIPELINE": "false"
      }
    }
  }
}

This is my mpc-config.json. args should be correct, but the agent still cannot load the tool. Can anyone see what’s wrong?