Deep Code Reasoning MCP Server
An MCP server that pairs Claude Code with Google's Gemini AI for complementary code analysis. This server enables a multi-model workflow where Claude Code handles tight terminal integration and multi-file refactoring, while Gemini leverages its massive context window (1M tokens) and code execution capabilities for distributed system debugging and long-trace analysis.
Core Value
Both Claude and Gemini can handle deep semantic reasoning and distributed system bugs. This server enables an intelligent routing strategy where:
- Claude Code excels at local-context operations, incremental patches, and CLI-native workflows
- Gemini 2.5 Pro shines with huge-context sweeps, synthetic test execution, and analyzing failures that span logs + traces + code
The "escalation" model treats LLMs like heterogeneous microservices - route to the one that's most capable for each sub-task.
Features
- Gemini 2.5 Pro Preview: Uses Google's latest Gemini 2.5 Pro Preview (05-06) model with 1M token context window
- Conversational Analysis: NEW! AI-to-AI dialogues between Claude and Gemini for iterative problem-solving
- Execution Flow Tracing: Understands data flow and state transformations, not just function calls
- Cross-System Impact Analysis: Models how changes propagate across service boundaries
- Performance Modeling: Identifies N+1 patterns, memory leaks, and algorithmic bottlenecks
- Hypothesis Testing: Tests theories about code behavior with evidence-based validation
- Long Context Support: Leverages Gemini 2.5 Pro Preview's 1M token context for analyzing large codebases
Prerequisites
- Node.js 18 or later
- A Google Cloud account with Gemini API access
- Gemini API key from Google AI Studio
Key Dependencies
- @google/generative-ai: Google's official SDK for Gemini API integration
- @modelcontextprotocol/sdk: MCP protocol implementation for Claude integration
- zod: Runtime type validation for tool parameters
- dotenv: Environment variable management
Installation
Quick Install for Cursor
Note: After installation, you'll need to update the file path to your actual installation directory and set your GEMINI_API_KEY.
Manual Installation
- Clone the repository:
git clone https://github.com/Haasonsaas/deep-code-reasoning-mcp.git
cd deep-code-reasoning-mcp
- Install dependencies:
- Set up your Gemini API key:
cp .env.example .env
# Edit .env and add your GEMINI_API_KEY
- Build the project:
Configuration
Environment Variables
GEMINI_API_KEY (required): Your Google Gemini API key
Claude Desktop Configuration
Add to your Claude Desktop configuration (~/Library/Application Support/Claude/claude_desktop_config.json):
{
"mcpServers": {
"deep-code-reasoning": {
"command": "node",
"args": ["/path/to/deep-code-reasoning-mcp/dist/index.js"],
"env": {
"GEMINI_API_KEY": "your-gemini-api-key"
}
}
}
}
How It Works
- Claude Code performs initial analysis using its strengths in multi-file refactoring and test-driven loops
- When beneficial, Claude escalates to this MCP server - particularly for:
- Analyzing gigantic log/trace dumps that exceed Claude's context
- Running iterative hypothesis testing with code execution
- Correlating failures across many microservices
- Server prepares comprehensive context including code, logs, and traces
- Gemini analyzes with its 1M-token context and visible "thinking" traces
- Results returned to Claude Code for implementation of fixes
Note: The tool parameters use snake_case naming convention and are validated using Zod schemas. The actual implementation provides more detailed type safety than shown in these simplified examples. Full TypeScript type definitions are available in src/models/types.ts.
The server now includes AI-to-AI conversational tools that enable Claude and Gemini to engage in multi-turn dialogues for complex analysis:
start_conversation
Initiates a conversational analysis session between Claude and Gemini.
{
claude_context: {
attempted_approaches: string[]; // What Claude tried
partial_findings: any[]; // What Claude found
stuck_description: string; // Where Claude got stuck
code_scope: {
files: string[]; // Files to analyze
entry_points?: CodeLocation[]; // Starting points
service_names?: string[]; // Services involved
}
};
analysis_type: 'execution_trace' | 'cross_system' | 'performance' | 'hypothesis_test';
initial_question?: string; // Optional opening question
}
continue_conversation
Continues an active conversation with Claude's response or follow-up question.
{
session_id: string; // Active session ID
message: string; // Claude's message to Gemini
include_code_snippets?: boolean; // Enrich with code context
}
finalize_conversation
Completes the conversation and generates structured analysis results.
{
session_id: string; // Active session ID
summary_format: 'detailed' | 'concise' | 'actionable';
}
get_conversation_status
Checks the status and progress of an ongoing conversation.
{
session_id: string; // Session ID to check
}
escalate_analysis
Main tool for handing off complex analysis from Claude Code to Gemini.
{
claude_context: {
attempted_approaches: string[]; // What Claude tried
partial_findings: any[]; // What Claude found
stuck_description: string; // Where Claude got stuck
code_scope: {
files: string[]; // Files to analyze
entry_points?: CodeLocation[]; // Starting points (file, line, function_name)
service_names?: string[]; // Services involved
}
};
analysis_type: 'execution_trace' | 'cross_system' | 'performance' | 'hypothesis_test';
depth_level: 1-5; // Analysis depth
time_budget_seconds?: number; // Time limit (default: 60)
}
trace_execution_path
Deep execution analysis with Gemini's semantic understanding.
{
entry_point: {
file: string;
line: number;
function_name?: string;
};
max_depth?: number; // Default: 10
include_data_flow?: boolean; // Default: true
}
cross_system_impact
Analyze impacts across service boundaries.
{
change_scope: {
files: string[];
service_names?: string[];
};
impact_types?: ('breaking' | 'performance' | 'behavioral')[];
}
Deep performance analysis beyond simple profiling.
{
code_path: {
entry_point: {
file: string;
line: number;
function_name?: string;
};
suspected_issues?: string[];
};
profile_depth?: 1-5; // Default: 3
}
hypothesis_test
Test specific theories about code behavior.
{
hypothesis: string;
code_scope: {
files: string[];
entry_points?: CodeLocation[]; // Optional array of {file, line, function_name?}
};
test_approach: string;
}
Example Use Cases
Conversational Analysis Example
When Claude needs deep iterative analysis with Gemini:
// 1. Start conversation
const session = await start_conversation({
claude_context: {
attempted_approaches: ["Checked for N+1 queries", "Profiled database calls"],
partial_findings: [{ type: "performance", description: "Multiple DB queries in loop" }],
stuck_description: "Can't determine if queries are optimizable",
code_scope: { files: ["src/services/UserService.ts"] }
},
analysis_type: "performance",
initial_question: "Are these queries necessary or can they be batched?"
});
// 2. Continue with follow-ups
const response = await continue_conversation({
session_id: session.sessionId,
message: "The queries fetch user preferences. Could we use a join instead?",
include_code_snippets: true
});
// 3. Finalize when ready
const results = await finalize_conversation({
session_id: session.sessionId,
summary_format: "actionable"
});
Case 1: Distributed Trace Analysis
When a failure signature spans multiple services with GB of logs:
// Claude Code: Identifies the error pattern and suspicious code sections
// Escalate to Gemini when: Need to correlate 1000s of trace spans across 10+ services
// Gemini: Processes the full trace timeline, identifies the exact race window
When performance degrades but the cause isn't obvious:
// Claude Code: Quick profiling, identifies hot paths
// Escalate to Gemini when: Need to analyze weeks of performance metrics + code changes
// Gemini: Correlates deployment timeline with perf metrics, pinpoints the exact commit
Case 3: Hypothesis-Driven Debugging
When you have theories but need extensive testing:
// Claude Code: Forms initial hypotheses based on symptoms
// Escalate to Gemini when: Need to test 20+ scenarios with synthetic data
// Gemini: Uses code execution API to validate each hypothesis systematically
Development
# Run in development mode
npm run dev
# Run tests
npm test
# Lint code
npm run lint
# Type check
npm run typecheck
Architecture
┌─────────────────┐ ┌──────────────────┐ ┌─────────────────┐
│ Claude Code │────▶│ MCP Server │────▶│ Gemini API │
│ (Fast, Local, │ │ (Router & │ │ (1M Context, │
│ CLI-Native) │◀────│ Orchestrator) │◀────│ Code Exec) │
└─────────────────┘ └──────────────────┘ └─────────────────┘
│
▼
┌──────────────────┐
│ Code + Logs + │
│ Traces + Tests │
└──────────────────┘
Security Considerations
- API Key: Store your Gemini API key securely in environment variables
- Code Access: The server reads local files - ensure proper file permissions
- Data Privacy: Code is sent to Google's Gemini API - review their data policies
Troubleshooting
"GEMINI_API_KEY not found"
- Ensure you've set the
GEMINI_API_KEY in your .env file or environment - Check that the
.env file is in the project root
"File not found" errors
- Verify that file paths passed to the tools are absolute paths
- Check file permissions
Gemini API errors
- Verify your API key is valid and has appropriate permissions
- Check API quotas and rate limits
- Ensure your Google Cloud project has the Gemini API enabled
Validation errors
- The server uses Zod for parameter validation
- Ensure all required parameters are provided
- Check that parameter names use snake_case (e.g.,
claude_context, not claudeContext) - Review error messages for specific validation requirements
Best Practices for Multi-Model Debugging
When debugging distributed systems with this MCP server:
- Capture the timeline first - Use OpenTelemetry/Jaeger traces with request IDs
- Start with Claude Code - Let it handle the initial investigation and quick fixes
- Escalate strategically to Gemini when you need:
- Analysis of traces spanning 100s of MB
- Correlation across 10+ services
- Iterative hypothesis testing with code execution
- Combine with traditional tools:
go test -race, ThreadSanitizer for race detection- rr or JFR for deterministic replay
- TLA+ or Alloy for formal verification
Contributing
- Fork the repository
- Create a feature branch
- Make your changes
- Add tests for new functionality
- Submit a pull request
License
This project is licensed under the MIT License - see the LICENSE file for details.
Author
Jonathan Haas - GitHub Profile
Acknowledgments
- Built for integration with Anthropic's Claude Code
- Powered by Google's Gemini AI
- Uses the Model Context Protocol (MCP) for communication
Support
If you encounter any issues or have questions: