A powerful MCP server that allows Claude Code to consult Gemini for complex coding problems with full code context and conversation persistence.

• Session Management: Maintain conversation context across multiple queries
• File Attachments: Read and include actual code files in conversations
• Hybrid Context: Combine text-based code_context with file attachments
• Follow-up Questions: Ask follow-up questions without resending code context
• Context Caching: Code context and file content are cached per session
• Automatic Processing: Files are processed and formatted automatically
• Multiple Sessions: Run multiple parallel conversations for different problems
• Session Expiry: Automatic cleanup of inactive sessions after 1 hour
• Latest Model: Uses Gemini 2.5 Pro (stable) by default

When Claude Code encounters difficult problems or needs a second opinion, it can:

• Send complete code files by reading them from the local filesystem
• Include text-based code context alongside file attachments
• Have multi-turn conversations about the same problem
• Get different perspectives without repeating context
• Work on multiple problems in parallel sessions
• Process files locally and include content in conversations

1. Clone this repository
2. Create a Python virtual environment:

   python3 -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

4. Copy .env.example to .env and add your Gemini API key:

   cp .env.example .env
   # Edit .env file and set your GEMINI_API_KEY

5. Add to Claude Code:

   claude mcp add gemini-coding -s user -- /path/to/gemini-mcp/start_server.sh

   Replace /path/to/gemini-mcp/ with the actual path to this directory.

Start or continue a conversation with Gemini about complex coding problems.

Parameters:

• session_id (optional): Continue a previous conversation
• problem_description: Description of the problem (required for new sessions)
• code_context: All relevant code (required for new sessions, cached afterward)
• attached_files (optional): Array of file paths to read and include in the conversation
• file_descriptions (optional): Object mapping file paths to descriptions
• specific_question: The question you want answered
• additional_context (optional): Updates or changes since last question
• preferred_approach: Type of help needed (solution/review/debug/optimize/explain/follow-up)

List all active Gemini consultation sessions.

End a specific session to free up memory.

/consult_gemini 
  problem_description: "I need to implement efficient caching for a React application"
  code_context: "[paste entire relevant codebase]"
  specific_question: "What's the best approach for implementing LRU cache with React Query?"
  preferred_approach: "solution"

/consult_gemini 
  problem_description: "I need to optimize this React component for performance"
  attached_files: ["/absolute/path/to/src/components/Dashboard.jsx", "/absolute/path/to/src/hooks/useData.js", "/absolute/path/to/package.json"]
  file_descriptions: {
    "/absolute/path/to/src/components/Dashboard.jsx": "Main dashboard component with performance issues",
    "/absolute/path/to/src/hooks/useData.js": "Custom hook for data fetching", 
    "/absolute/path/to/package.json": "Project dependencies"
  }
  specific_question: "How can I improve the rendering performance of this dashboard?"
  preferred_approach: "optimize"

/consult_gemini 
  problem_description: "Complex authentication flow needs debugging"
  code_context: "// Additional context or pseudocode here"
  attached_files: ["/absolute/path/to/auth/login.js", "/absolute/path/to/middleware/auth.js"]
  specific_question: "Why is the token refresh failing?"
  preferred_approach: "debug"

Response includes a session ID for follow-ups.

/consult_gemini
  session_id: "abc123..."
  specific_question: "I implemented your suggestion but getting stale data issues. How do I handle cache invalidation?"
  additional_context: "Added the LRU cache as suggested, but users see old data after updates"
  preferred_approach: "follow-up"

/list_sessions
# Shows all active sessions with IDs and summaries

/end_session
  session_id: "abc123..."
# Frees up memory for completed conversations

1. Initial Context: Include ALL relevant code via code_context or attached_files
2. File Organization: Use attached_files for multiple files, code_context for snippets
3. File Descriptions: Provide clear descriptions for each attached file
4. Follow-ups: Use the session ID to continue conversations
5. Additional Context: When asking follow-ups, explain what changed
6. Session Management: End sessions when done to free memory and clean up files
7. Multiple Problems: Use different sessions for unrelated problems
8. File Types: Supports JavaScript, Python, TypeScript, JSON, and other text-based files

You can test the server directly before adding it to Claude Code:

# Make sure your .env file has a valid GEMINI_API_KEY
./start_server.sh

The server will start and display:

Gemini Coding Assistant MCP Server v3.0 running (Python)
Features: Session management, file attachments, context persistence, follow-up questions
Ready to help with complex coding problems!

• Maximum combined input: ~50,000 characters per message
• Maximum response: 8,192 tokens (~16,000 characters)
• Session timeout: 1 hour of inactivity
• Rate limiting: 1 second between requests

1. First Message: Creates a new session, caches code context
2. Follow-ups: Reuses cached context, maintains conversation history
3. Session Storage: In-memory storage (use Redis for production)
4. Cleanup: Automatic expiry after 1 hour of inactivity

• Efficiency: Code context sent only once per session
• Context: Gemini remembers previous questions and answers
• Natural Flow: Have real conversations about complex problems
• Cost Savings: Reduced token usage for follow-up questions

• API key is never exposed
• Rate limiting prevents abuse
• Sessions expire automatically
• No persistent storage of code

• v2.1.0: Added file attachment system with automatic cleanup
• v2.0.0: Added session management and follow-up support
• v1.0.0: Initial stateless implementation