AI Agent Template MCP Server

# MCP Context Manager > **Force AI to TRULY understand your codebase before generating code** ## 🎯 What This MCP Server Actually Does MCP Context Manager is NOT just a template provider. It's a quality enforcement system that: 1. **Forces Deep Analysis** - Makes AI read EVERY source file before creating documentation 2. **Requires Evidence** - Every pattern claimed must have 5+ code examples 3. **Prevents Hallucination** - Templates must be based on actual code, not assumptions 4. **Automates Context Loading** - Sets up auto-loading so AI always uses the context ## ⚠️ CRITICAL WARNING ### This MCP Only Works If You Use The Tools! **✅ CORRECT (enforces analysis):** ``` Please use MCP Context Manager's complete_setup_workflow tool ``` **❌ WRONG (bypasses all quality checks):** ``` Please analyze this codebase and create context files... Please create agent-context/PROJECT-TEMPLATE.md... Read every source file and... ``` The MCP provides **tools** with quality gates. Manual instructions bypass everything! ``` MCP Tools: analyze_codebase_deeply → MUST complete → create_project_template Manual: "create files" → AI creates generic files → No analysis! ``` ## 🤔 Claude Desktop vs Cursor | Feature | Claude Desktop | Cursor | |---------|---------------|---------| | **MCP Support** | ✅ Native | ✅ Native (v1.0+) | | **Config Location** | `claude_desktop_config.json` | `.cursor/mcp.json` | | **Setup Method** | Edit config & restart | Settings → MCP Tools | | **Tool Invocation** | Direct | Via Composer Agent | | **Auto-loading** | @ mentions | Project rules | | **Tool Limit** | No limit | 40 max | | **OAuth Support** | Varies | ✅ Built-in | ## 📦 Version Requirements **IMPORTANT**: You need version 1.4.0 or later for the `complete_setup_workflow` tool. Check your version: ```bash npm list mcp-context-manager ``` Update if needed: ```bash npm install -g mcp-context-manager@latest # or for npx users, it auto-updates ``` ## 🚀 The ONLY Correct Way to Use This ### For Both Claude Desktop and Cursor #### Step 1: Configure MCP **Claude Desktop:** Add to `claude_desktop_config.json`: ```json { "mcpServers": { "context-manager": { "command": "npx", "args": ["mcp-context-manager", "serve"] } } } ``` **Cursor:** Create `.cursor/mcp.json`: ```json { "context-manager": { "command": "npx", "args": ["mcp-context-manager", "serve"] } } ``` #### Step 2: Use MCP Tools (NOT Manual Instructions!) **✅ CORRECT - Uses MCP tools with enforcement:** ``` Please use MCP Context Manager's complete_setup_workflow tool to analyze, document, and setup automatic context loading for this project. ``` **✅ ALTERNATIVE - If `complete_setup_workflow` tool not found:** ``` I see the MCP Context Manager tools. Please run them in this sequence: 1. First use the analyze_codebase_deeply tool 2. Then create_shared_tech_stack 3. Then create_project_template with analysis_complete: true 4. Then create_codebase_context 5. Then create_initial_adrs 6. Then create_directory_readme for main directories 7. Finally setup_auto_context_loading ``` **❌ WRONG - Manual instructions bypass quality checks:** ``` Please analyze this codebase and create context files... Create agent-context/PROJECT-TEMPLATE.md with... ``` #### Step 3: What Actually Happens 1. **analyze_codebase_deeply** runs first (2-3 minutes) - AI reads EVERY .ts/.tsx/.js/.jsx file - Collects evidence for patterns - Builds understanding of architecture 2. **create_project_template** checks analysis was done - Rejects if analysis incomplete - Forces evidence-based content 3. **All other tools** create files with real data - ADRs with actual patterns found - Context with code examples - Auto-loading rules configured ## 📁 What Gets Created (WITH Evidence!) After the MCP tools complete their analysis: ``` agent-context/ ├── PROJECT-TEMPLATE.md # Filled with YOUR project data ├── CODEBASE-CONTEXT.md # YOUR actual patterns with code examples ├── .context7.yaml # Prevents hallucination with YOUR versions ├── shared/ │ └── tech-stack.yaml # YOUR actual dependencies + usage info ├── adr/ # 6 architecture files based on YOUR code │ ├── 001-frontend-framework.md # "React because found in 47 components" │ ├── 002-state-management.md # "Zustand found in src/stores/" │ └── ... # All with evidence! └── directories/ # READMEs for YOUR actual directories .cursor/rules/ # Auto-loading for Cursor .mcp.json # Auto-start config ``` ## ✅ How to Verify It Worked **Good Result:** - Files contain specific examples: "PascalCase found in Button.tsx, UserCard.tsx..." - Patterns reference actual files: "Error handling pattern from userService.ts:45" - Tech choices show usage: "React 18.3 (hooks in 47 files, Suspense in 3)" **Bad Result (means AI skipped analysis):** - Generic content: "Project uses React hooks" - No file references: "Follow naming conventions" - Empty or missing files ## 🔧 How MCP Tools Enforce Quality The MCP server provides tools with built-in enforcement: ### The Workflow Tools 1. **`complete_setup_workflow`** - Runs the entire process in correct order 2. **`analyze_codebase_deeply`** - FORCES AI to read every source file 3. **`create_project_template`** - REQUIRES `analysis_complete: true` parameter ### Quality Gates - ❌ Can't create templates without analysis - ❌ Can't claim patterns without evidence - ❌ Can't use generic descriptions - ✅ Must reference specific files - ✅ Must show code examples - ✅ Must count occurrences ### What Each Tool Does - **`create_codebase_context`** - Documents YOUR patterns with evidence - **`create_initial_adrs`** - Creates 6 architecture decisions from analysis - **`create_shared_tech_stack`** - Lists versions WITH usage statistics - **`setup_auto_context_loading`** - Configures auto-loading for both tools ## 📋 PROJECT-TEMPLATE.md Contents The template includes: - Project metadata (name, version, stage) - Tech stack details - Architecture patterns - Development setup - AI assistance configuration - Testing strategy - Deployment process - Common commands - Troubleshooting guides ## 🎯 Why This Matters **Without MCP Enforcement:** - AI creates generic templates - Patterns don't match your code - Documentation becomes outdated - Generated code is inconsistent **With MCP Enforcement:** - Templates reflect YOUR actual code - Every pattern has evidence - AI understands YOUR architecture - Generated code matches YOUR style ## 🤖 How It Works 1. **MCP Server** provides templates and tools 2. **AI Agent** analyzes your codebase 3. **AI fills templates** with your actual data 4. **Every generation** uses this context 5. **Consistent code** matching your patterns ## 💡 Best Practices 1. **Update Regularly** - Keep context files current as project evolves 2. **Be Specific** - Fill templates with detailed, specific information 3. **Document Patterns** - Clear examples help AI follow them 4. **Review Generated Files** - Ensure AI captured patterns correctly ## 🔍 Commands ```bash # Initialize in project mcp-context init # Start server mcp-context serve # Check if context files exist mcp-context check ``` ## 💡 Pro Tips 1. **First Time Setup**: The magic prompt will take 2-3 minutes as the AI reads every file 2. **Update Regularly**: Re-run the analysis when you add major features 3. **Check File Creation**: If folders are empty, remind the AI to create the actual files 4. **Use Context**: Always reference agent-context files before generating code ## 🛠️ Troubleshooting ### "Tool not found" errors **Problem**: Getting "I don't see a complete_setup_workflow tool" or similar **Causes & Solutions**: 1. **Old version**: Update to v1.4.0+ with `npm install -g mcp-context-manager@latest` 2. **Cache issue**: Restart Claude/Cursor after updating 3. **Connection issue**: Check MCP server is running (green indicator) 4. **Alternative**: Use the manual sequence of tools shown in Step 2 ### Version checking ```bash # Check installed version npm list -g mcp-context-manager # Update to latest npm install -g mcp-context-manager@latest # For npx users (auto-updates) npx mcp-context-manager@latest serve ``` ### "AI created files without analyzing code" **Problem**: You used manual instructions instead of MCP tools **Solution**: Use the exact command: ``` Please use MCP Context Manager's complete_setup_workflow tool to analyze, document, and setup automatic context loading for this project. ``` ### "MCP server not found" **Claude**: Check configuration in settings, restart Claude Desktop **Cursor**: Check Settings → MCP Tools for green indicator ### "Empty folders (adr/, scripts/)" The AI skipped the analysis. Tell it: ``` The folders are empty. Please run the analyze_codebase_deeply tool first, then create the files with evidence from your analysis. ``` ### "Generic content without examples" The AI didn't actually analyze. Check if: - Files contain specific file references (Button.tsx, userService.ts) - Patterns show counts (found in 23 files) - Examples include line numbers (pattern at line 45) ### MCP server connection issues **Symptoms**: Red indicator in Cursor, tools not loading **Solutions**: 1. Check MCP server is running: `ps aux | grep mcp-context` 2. Verify config path: `.cursor/mcp.json` exists 3. Try manual start: `npx mcp-context-manager serve` 4. Check logs: Cursor → Developer → Toggle Developer Tools ## 📊 What You'll Get After running the magic prompt, your project will have: | File | Purpose | Impact | |------|---------|--------| | PROJECT-TEMPLATE.md | Complete project documentation | 📈 30% fewer tokens needed | | CODEBASE-CONTEXT.md | Your specific patterns & rules | 🎯 84% fewer bugs | | 6 ADR files | Architecture decisions | 🏗️ Consistent code | | tech-stack.yaml | Version management | 🔒 No version conflicts | | Directory READMEs | API documentation | 📚 Clear interfaces | ## 🎯 Real Example: What Actually Happens ``` You: Please use MCP Context Manager's complete_setup_workflow tool to analyze, document, and setup automatic context loading for this project. Claude: I'll run the complete MCP Context Manager workflow. Let me start with the deep analysis. [Running analyze_codebase_deeply tool] Starting comprehensive codebase analysis... Reading src/components/Button.tsx... Reading src/components/UserCard.tsx... Reading src/components/NavigationBar.tsx... [... reads ALL 147 source files ...] Analysis complete! Found: - Components: PascalCase naming (Button.tsx, UserCard.tsx, NavBar.tsx in src/components/) - State: Zustand stores (authStore.ts, uiStore.ts in src/stores/) - API: Axios with interceptors (found in 23 service files) - Error handling: try-catch with toast notifications (pattern in userService.ts:45) [Running create_project_template tool - requires analysis_complete: true] Creating PROJECT-TEMPLATE.md with evidence from analysis... [Running create_initial_adrs tool] Creating ADR files based on discovered patterns: - 001-frontend-framework.md: "React 18.3 - found hooks in 47 files" - 002-state-management.md: "Zustand v4.5 - stores pattern in src/stores/" [... creates all 6 ADRs with evidence ...] ✅ Created 15 context files with REAL data from YOUR code 📊 Analyzed 147 source files (not just package.json!) 🎯 Documented 23 patterns with 200+ code examples 🤖 Auto-loading configured You: Create a user profile component. Claude: [Reads context automatically] Based on your codebase patterns: - Component naming: UserProfile.tsx (matching Button.tsx, UserCard.tsx pattern) - State: Using authStore from src/stores/authStore.ts - Error handling: try-catch with toast as in userService.ts:45 [Generates code that EXACTLY matches your patterns] ``` ## 🤖 Automatic Context Loading (2025) No more manually reminding AI to read context! The new automation features include: ### For Cursor IDE - **Project Rules**: `.cursor/rules/*.mdc` files auto-attach to every file - **Path Patterns**: Rules apply based on file types and locations - **Quality Gates**: Automatic checks before code generation ### For Claude Desktop - **@ Mentions**: Type `@context` to access all context files - **Project MCP**: `.mcp.json` configures auto-start and resource exposure - **Global Rules**: `~/.claude/CLAUDE.md` for all projects ### Setup Command ``` Please use MCP Context Manager's setup_auto_context_loading tool ``` This creates all necessary configuration files for both tools. ## 🚀 Advanced Usage ### Custom Analysis ``` Please use MCP Context Manager to analyze this codebase, focusing especially on: - Authentication patterns - API error handling - Component composition patterns ``` ### Update Existing Context ``` Please update the agent-context files - we've added GraphQL to the project. ``` ### Generate Specific ADRs ``` Please create an ADR for our new caching strategy using the MCP Context Manager. ``` ## 🤝 Contributing This is a minimal MCP server focused on one thing: ensuring AI agents use project context. PRs welcome! ## 📄 License MIT --- **Remember**: This MCP doesn't just provide templates - it FORCES AI to understand your code first. But only if you use the tools, not manual instructions!