Vibe Coder MCP

Vibe Coder MCP Server

Vibe Coder is an MCP (Model Context Protocol) server designed to supercharge your AI assistant (like Cursor, Cline AI, or Claude Desktop) with powerful tools for software development. It helps with research, planning, generating requirements, creating starter projects, and more!

Overview & Features

Vibe Coder MCP integrates with MCP-compatible clients to provide the following capabilities:

• Semantic Request Routing: Intelligently routes requests using embedding-based semantic matching with sequential thinking fallbacks.
• Tool Registry Architecture: Centralized tool management with self-registering tools.
• Direct LLM Calls: Generator tools now use direct LLM calls for improved reliability and structured output control.
• Workflow Execution: Runs predefined sequences of tool calls defined in workflows.json.
• Research & Planning: Performs deep research (research-manager) and generates planning documents like PRDs (generate-prd), user stories (generate-user-stories), task lists (generate-task-list), and development rules (generate-rules).
• Project Scaffolding: Generates full-stack starter kits (generate-fullstack-starter-kit).
• Code Map Generator: Recursively scans a codebase, extracts semantic information, and generates either a token-efficient, context-dense Markdown index with Mermaid diagrams or a structured JSON representation with absolute file paths for imports and enhanced class property information (map-codebase).
• Asynchronous Execution: Many long-running tools (generators, research, workflows) now run asynchronously. They return a Job ID immediately, and the final result is retrieved using the get-job-result tool.
• Session State Management: Maintains basic state across requests within a session (in-memory).
• Standardized Error Handling: Consistent error patterns across all tools.

(See "Detailed Tool Documentation" and "Feature Details" sections below for more)

Setup Guide

Follow these micro-steps to get the Vibe Coder MCP server running and connected to your AI assistant.

Step 1: Prerequisites

1. Check Node.js Version:
   • Open a terminal or command prompt.
   • Run node -v
   • Ensure the output shows v18.0.0 or higher (required).
   • If not installed or outdated: Download from nodejs.org.
2. Check Git Installation:
   • Open a terminal or command prompt.
   • Run git --version
   • If not installed: Download from git-scm.com.
3. Get OpenRouter API Key:
   • Visit openrouter.ai
   • Create an account if you don't have one.
   • Navigate to API Keys section.
   • Create a new API key and copy it.
   • Keep this key handy for Step 4.

Step 2: Get the Code

1. Create a Project Directory (optional):
   • Open a terminal or command prompt.
   • Navigate to where you want to store the project:
     cd ~/Documents # Example: Change to your preferred location
2. Clone the Repository:
   • Run:
     git clone https://github.com/freshtechbro/vibe-coder-mcp.git
     (Or use your fork's URL if applicable)
3. Navigate to Project Directory:
   • Run:
     cd vibe-coder-mcp

Step 3: Run the Setup Script

Choose the appropriate script for your operating system:

For Windows:

1. In your terminal (still in the vibe-coder-mcp directory), run:
   setup.bat
2. Wait for the script to complete (it will install dependencies, build the project, and create necessary directories).
3. If you see any error messages, refer to the Troubleshooting section below.

For macOS or Linux:

1. Make the script executable:
   chmod +x setup.sh
2. Run the script:
   ./setup.sh
3. Wait for the script to complete.
4. If you see any error messages, refer to the Troubleshooting section below.

The script performs these actions:

• Checks Node.js version (v18+)
• Installs all dependencies via npm
• Creates necessary VibeCoderOutput/ subdirectories (as defined in the script).
• Builds the TypeScript project.
• Copies .env.example to .env if .env doesn't already exist. You will need to edit this file.
• Sets executable permissions (on Unix systems).

Step 4: Configure Environment Variables (.env)

The setup script (from Step 3) automatically creates a .env file in the project's root directory by copying the .env.example template, only if .env does not already exist.

1. Locate and Open .env: Find the .env file in the main vibe-coder-mcp directory and open it with a text editor.
2. Add Your OpenRouter API Key (Required):
   • The file contains a template based on .env.example:
     # OpenRouter Configuration ## Specifies your unique API key for accessing OpenRouter services. ## Replace "Your OPENROUTER_API_KEY here" with your actual key obtained from OpenRouter.ai. OPENROUTER_API_KEY="Your OPENROUTER_API_KEY here" ## Defines the base URL for the OpenRouter API endpoints. ## The default value is usually correct and should not need changing unless instructed otherwise. OPENROUTER_BASE_URL=https://openrouter.ai/api/v1 ## Sets the specific Gemini model to be used via OpenRouter for certain AI tasks. ## ':free' indicates potential usage of a free tier model if available and supported by your key. GEMINI_MODEL=google/gemini-2.0-flash-thinking-exp:free
   • Crucially, replace "Your OPENROUTER_API_KEY here" with your actual OpenRouter API key. Remove the quotes if your key doesn't require them.
3. Configure Output Directory (Optional):
   • To change where generated files are saved (default is VibeCoderOutput/ inside the project), add this line to your .env file:
     VIBE_CODER_OUTPUT_DIR=/path/to/your/desired/output/directory
   • Replace the path with your preferred absolute path. Use forward slashes (/) for paths. If this variable is not set, the default directory (VibeCoderOutput/) will be used.
4. Configure Code-Map Generator Directory (Optional):
   • To specify which directory the code-map-generator tool is allowed to scan, add this line to your .env file:
     CODE_MAP_ALLOWED_DIR=/path/to/your/source/code/directory
   • Replace the path with the absolute path to the directory containing the source code you want to analyze. This is a security boundary - the tool will not access files outside this directory.
   • Note that CODE_MAP_ALLOWED_DIR (for reading source code) and VIBE_CODER_OUTPUT_DIR (for writing output files) are separate for security reasons. The code-map-generator tool uses separate validation for read and write operations.
5. Review Other Settings (Optional):
   • You can add other environment variables supported by the server, such as LOG_LEVEL (e.g., LOG_LEVEL=debug) or NODE_ENV (e.g., NODE_ENV=development).
6. Save the .env File.

Step 5: Integrate with Your AI Assistant (MCP Settings)

This crucial step connects Vibe Coder to your AI assistant by adding its configuration to the client's MCP settings file.

5.1: Locate Your Client's MCP Settings File

The location varies depending on your AI assistant:

• Cursor AI / Windsurf / RooCode (VS Code based):
  1. Open the application.
  2. Open the Command Palette (Ctrl+Shift+P or Cmd+Shift+P).
  3. Type and select Preferences: Open User Settings (JSON).
  4. This opens your settings.json file where the mcpServers object should reside.
• Cline AI (VS Code Extension):
  • Windows: %APPDATA%\Cursor\User\globalStorage\saoudrizwan.claude-dev\settings\cline_mcp_settings.json
  • macOS: ~/Library/Application Support/Cursor/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json
  • Linux: ~/.config/Cursor/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json
  • (Note: If using standard VS Code instead of Cursor, replace Cursor with Code in the path)
• Claude Desktop:
  • Windows: %APPDATA%\Claude\claude_desktop_config.json
  • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
  • Linux: ~/.config/Claude/claude_desktop_config.json

5.2: Add the Vibe Coder Configuration

1. Open the settings file identified above in a text editor.
2. Find the "mcpServers": { ... } JSON object. If it doesn't exist, you may need to create it (ensure the overall file remains valid JSON). For example, an empty file might become {"mcpServers": {}}.
3. Add the following configuration block inside the curly braces {} of the mcpServers object. If other servers are already listed, add a comma , after the previous server's closing brace } before pasting this block.
   // This is the unique identifier for this MCP server instance within your client's settings "vibe-coder-mcp": { // Specifies the command used to execute the server. Should be 'node' if Node.js is in your system's PATH "command": "node", // Provides the arguments to the 'command'. The primary argument is the absolute path to the compiled server entry point // !! IMPORTANT: Replace with the actual absolute path on YOUR system. Use forward slashes (/) even on Windows !! "args": ["/Users/username/Documents/Dev Projects/Vibe-Coder-MCP/build/index.js"], // Sets the current working directory for the server process when it runs // !! IMPORTANT: Replace with the actual absolute path on YOUR system. Use forward slashes (/) even on Windows !! "cwd": "/Users/username/Documents/Dev Projects/Vibe-Coder-MCP", // Defines the communication transport protocol between the client and server "transport": "stdio", // Environment variables to be passed specifically to the Vibe Coder server process when it starts // API Keys should be in the .env file, NOT here "env": { // Absolute path to the LLM configuration file used by Vibe Coder // !! IMPORTANT: Replace with the actual absolute path on YOUR system !! "LLM_CONFIG_PATH": "/Users/username/Documents/Dev Projects/Vibe-Coder-MCP/llm_config.json", // Sets the logging level for the server "LOG_LEVEL": "debug", // Specifies the runtime environment "NODE_ENV": "production", // Directory where Vibe Coder tools will save their output files // !! IMPORTANT: Replace with the actual absolute path on YOUR system !! "VIBE_CODER_OUTPUT_DIR": "/Users/username/Documents/Dev Projects/Vibe-Coder-MCP/VibeCoderOutput", // Directory that the code-map-generator tool is allowed to scan // This is a security boundary - the tool will not access files outside this directory "CODE_MAP_ALLOWED_DIR": "/Users/username/Documents/Dev Projects/Vibe-Coder-MCP/src" }, // A boolean flag to enable (false) or disable (true) this server configuration "disabled": false, // A list of tool names that the MCP client is allowed to execute automatically "autoApprove": [ "research", "generate-rules", "generate-user-stories", "generate-task-list", "generate-prd", "generate-fullstack-starter-kit", "refactor-code", "git-summary", "run-workflow", "map-codebase" ] }
4. CRUCIAL: Replace all placeholder paths (like /path/to/your/vibe-coder-mcp/...) with the correct absolute paths on your system where you cloned the repository. Use forward slashes / for paths, even on Windows (e.g., C:/Users/YourName/Projects/vibe-coder-mcp/build/index.js). Incorrect paths are the most common reason the server fails to connect.
5. Save the settings file.
6. Completely close and restart your AI assistant application (Cursor, VS Code, Claude Desktop, etc.) for the changes to take effect.

Step 6: Test Your Configuration

1. Start Your AI Assistant:
   • Completely restart your AI assistant application.
2. Test a Simple Command:
   • Type a test command like: Research modern JavaScript frameworks
3. Check for Proper Response:
   • If working correctly, you should receive a research response.
   • If not, check the Troubleshooting section below.

Project Architecture

The Vibe Coder MCP server follows a modular architecture centered around a tool registry pattern:

Directory Structure

Tool Registry Pattern

The Tool Registry is a central component for managing tool definitions and execution:

Sequential Thinking Process

The Sequential Thinking mechanism provides LLM-based fallback routing:

Session State Management

Workflow Execution Engine

The Workflow system enables multi-step sequences:

Workflow Configuration

Workflows are defined in the workflows.json file located in the root directory of the project. This file contains predefined sequences of tool calls that can be executed with a single command.

File Location and Structure

• The workflows.json file must be placed in the project root directory (same level as package.json)
• The file follows this structure:
  { "workflows": { "workflowName1": { "description": "Description of what this workflow does", "inputSchema": { "param1": "string", "param2": "string" }, "steps": [ { "id": "step1_id", "toolName": "tool-name", "params": { "param1": "{workflow.input.param1}" } }, { "id": "step2_id", "toolName": "another-tool", "params": { "paramA": "{workflow.input.param2}", "paramB": "{steps.step1_id.output.content[0].text}" } } ], "output": { "summary": "Workflow completed message", "details": ["Output line 1", "Output line 2"] } } } }

Parameter Templates

Workflow step parameters support template strings that can reference:

• Workflow inputs: {workflow.input.paramName}
• Previous step outputs: {steps.stepId.output.content[0].text}

Triggering Workflows

Use the run-workflow tool with:

Detailed Tool Documentation

Each tool in the src/tools/ directory includes comprehensive documentation in its own README.md file. These files cover:

• Tool overview and purpose
• Input/output specifications
• Workflow diagrams (Mermaid)
• Usage examples
• System prompts used
• Error handling details

Refer to these individual READMEs for in-depth information:

• src/tools/fullstack-starter-kit-generator/README.md
• src/tools/prd-generator/README.md
• src/tools/research-manager/README.md
• src/tools/rules-generator/README.md
• src/tools/task-list-generator/README.md
• src/tools/user-stories-generator/README.md
• src/tools/workflow-runner/README.md
• src/tools/code-map-generator/README.md

Tool Categories

Analysis & Information Tools

• Code Map Generator (map-codebase): Scans a codebase to extract semantic information (classes, functions, comments) and generates either a human-readable Markdown map with Mermaid diagrams or a structured JSON representation with absolute file paths for imports and enhanced class property information.
• Research Manager (research-manager): Performs deep research on technical topics using Perplexity Sonar, providing summaries and sources.

Planning & Documentation Tools

• Rules Generator (generate-rules): Creates project-specific development rules and guidelines.
• PRD Generator (generate-prd): Generates comprehensive product requirements documents.
• User Stories Generator (generate-user-stories): Creates detailed user stories with acceptance criteria.
• Task List Generator (generate-task-list): Builds structured development task lists with dependencies.

Project Scaffolding Tool

• Fullstack Starter Kit Generator (generate-fullstack-starter-kit): Creates customized project starter kits with specified frontend/backend technologies, including basic setup scripts and configuration.

Workflow & Orchestration

• Workflow Runner (run-workflow): Executes predefined sequences of tool calls for common development tasks.

Generated File Storage

By default, outputs from the generator tools are stored for historical reference in the VibeCoderOutput/ directory within the project. This location can be overridden by setting the VIBE_CODER_OUTPUT_DIR environment variable in your .env file or AI assistant configuration.

Security Boundaries for Read and Write Operations

For security reasons, the Vibe Coder MCP tools maintain separate security boundaries for read and write operations:

• Read Operations: Tools like the code-map-generator only read from directories explicitly authorized through the CODE_MAP_ALLOWED_DIR environment variable. This creates a clear security boundary and prevents unauthorized access to files outside the allowed directory.
• Write Operations: All output files are written to the VIBE_CODER_OUTPUT_DIR directory (or its subdirectories). This separation ensures that tools can only write to designated output locations, protecting your source code from accidental modifications.

Example structure (default location):

Usage Examples

Interact with the tools via your connected AI assistant:

• Research: Research modern JavaScript frameworks
• Generate Rules: Create development rules for a mobile banking application
• Generate PRD: Generate a PRD for a task management application
• Generate User Stories: Generate user stories for an e-commerce website
• Generate Task List: Create a task list for a weather app based on [user stories]
• Sequential Thinking: Think through the architecture for a microservices-based e-commerce platform
• Fullstack Starter Kit: Create a starter kit for a React/Node.js blog application with user authentication
• Run Workflow: Run workflow newProjectSetup with input { "projectName": "my-new-app", "description": "A simple task manager" }
• Map Codebase: Generate a code map for the current project, map-codebase path="./src", or Generate a JSON representation of the codebase structure with output_format="json"

Running Locally (Optional)

While the primary use is integration with an AI assistant (using stdio), you can run the server directly for testing:

Running Modes

• Production Mode (Stdio):
  npm start
  • Logs go to stderr (mimics AI assistant launch)
  • Use NODE_ENV=production
• Development Mode (Stdio, Pretty Logs):
  npm run dev
  • Logs go to stdout with pretty formatting
  • Requires nodemon and pino-pretty
  • Use NODE_ENV=development
• SSE Mode (HTTP Interface):
  # Production mode over HTTP npm run start:sse # Development mode over HTTP npm run dev:sse
  • Uses HTTP instead of stdio
  • Configured via PORT in .env (default: 3000)
  • Access at http://localhost:3000

Detailed Troubleshooting

Connection Issues

MCP Server Not Detected in AI Assistant

1. Check Configuration Path:
   • Verify the absolute path in the args array is correct
   • Ensure all slashes are forward slashes / even on Windows
   • Run node <path-to-build/index.js> directly to test if Node can find it
2. Check Configuration Format:
   • Make sure JSON is valid without syntax errors
   • Check that commas between properties are correct
   • Verify that the mcpServers object contains your server
3. Restart the Assistant:
   • Completely close (not just minimize) the application
   • Reopen and try again

Server Starts But Tools Don't Work

1. Check Disabled Flag:
   • Ensure "disabled": false is set
   • Remove any // comments as JSON doesn't support them
2. Verify autoApprove Array:
   • Check that tool names in the autoApprove array match exactly
   • Try adding "process-request" to the array if using hybrid routing

API Key Issues

1. OpenRouter Key Problems:
   • Double-check that the key is correctly copied
   • Verify the key is active in your OpenRouter dashboard
   • Check if you have sufficient credits
2. Environment Variable Issues:
   • Verify the key is correct in both:
     • The .env file (for local runs)
     • Your AI assistant's configuration env block

Path & Permission Issues

1. Build Directory Not Found:
   • Run npm run build to ensure the build directory exists
   • Check if build output is going to a different directory (check tsconfig.json)
2. File Permission Errors:
   • Ensure your user has write access to the workflow-agent-files directory
   • On Unix systems, check if build/index.js has execute permission

Log Debugging

1. For Local Runs:
   • Check the console output for error messages
   • Try running with LOG_LEVEL=debug in your .env file
2. For AI Assistant Runs:
   • Set "NODE_ENV": "production" in the env configuration
   • Check if the assistant has a logging console or output window

Tool-Specific Issues

1. Semantic Routing Not Working:
   • First run may download embedding model - check for download messages
   • Try a more explicit request that mentions the tool name