Structured Workflow MCP Server

NOTE: I am not currently working on this or actively maintaining it. I learned a few things about prompting and agents while making this MCP server. It has a lot of very valuable ideas that could be used or improved upon as an MCP server but I'm also looking at ways to incorporate the core ideas into Agents, for example, in Claude. The core idea here is that AI should follow specific, pre-determined steps to solving a problem, just like we as humans might do and there may be other ways to achieve this apart from this MCP server.

An MCP server that enforces disciplined programming practices by requiring AI assistants to audit their work and produce verified outputs at each phase of development.

Why I Built This

TLDR: I got tired of repeating "inventory and audit first" across every AI platform and prompt, so I built an MCP server that automatically enforces this disciplined approach. It forces AI to think systematically and follow structured phases instead of jumping straight into code changes.

So I've built an MCP server that fits into my workflow and thinking process while I'm programming. I made it available via npx and you can download it yourself if you want something local.

In essence I was doing some repeated tasks with AI where I wanted it to complete refactoring work for part of a larger project. I was struggling because it was often missing or glossing over key things: classes or systems that already exist (a preferences service for example), creating duplicates of things, or when correcting mistakes, leaving orphaned unused methods/code around places, and when writing tests it would often pull in the wrong imports or put these together in the wrong way resulting in syntax errors but would jump straight into writing the next test without fixing the first one that was broken.

I sort of stumbled on this idea of the model needing to perform an audit and inventory of the current project (or not even the whole project--just one layer or feature in a project) before moving to any kind of implementation phase and it needed a lint iterate lint phase. I tried this with rules with limited success and then prompting with much better success but I was constantly repeating myself.

So I started noodling on this idea of an MCP server that forced the AI to work through a problem in phases or lanes. So that's what this does. There's a number of different workflow styles and I'm open to any other ideas or improvements.

Feel free to check it out if it helps your use case. It's a work in progress but it has been doing a pretty great job for what I'm using it for now. Happy to share more if you are interested.

Features

Enforced Workflow Phases - AI must complete specific phases in order (setup, audit, analysis, planning, implementation, testing, etc.)

Mandatory Output Artifacts - Each phase requires structured documentation or verified outputs before proceeding

Multiple Workflow Types:

• Refactor workflows for code improvement
• Feature development with integrated testing
• Test-focused workflows for coverage improvement
• Test-driven development (TDD) cycles
• Custom workflows for specialized needs

Output Verification - The server validates that outputs contain meaningful content and proper structure

Session State Management - Tracks progress and prevents skipping phases

How It Works

Here's how the AI moves through a structured workflow:

graph TD
    A[🚀 Start Workflow] --> B[AI Gets Phase Guidance]
    B --> C{Create Phase Output}
    C --> D[Auto-Save with Numbered Naming<br/>00-setup-confirmation-2025-01-07.md]
    D --> E[Phase Validation]
    E --> F{All Phases Done?}
    F -->|No| G[Move to Next Phase]
    G --> B
    F -->|Yes| H[Workflow Complete!]
    
    style A fill:#e1f5fe
    style B fill:#f3e5f5
    style C fill:#fff3e0
    style D fill:#e8f5e8
    style E fill:#fff9c4
    style H fill:#e8f5e8

What happens at each step:

1. Start Workflow - AI calls a workflow tool (refactor_workflow, create_feature_workflow, etc.)
2. AI Gets Phase Guidance - Server provides specific instructions for current phase (audit, analyze, implement, etc.)
3. Create Phase Output - AI works through the phase and creates documentation/artifacts
4. Auto-Save - Files are automatically saved with numbered naming in task directories
5. Phase Validation - Server validates outputs meet requirements before proceeding
6. Next Phase - Process repeats until workflow is complete

One benefit of this breakdown is that the AI agent receive instruction sets that are relevant to the current phase and not the entire workflow. This can help prevent the AI from getting lost in the weeds of the entire workflow and instead focus on the current phase. An interesting article on this can be read here: LLMs Get Lost In Multi-Turn Conversation

Workflow Output

AI-Generated Documentation

The server suggests numbered workflow files as you progress through phases. The AI assistant handles the actual file creation using its own tools:

workflows/
├── your-task-name/
│   ├── 01-audit-inventory-2025-01-04.md
│   ├── 02-compare-analyze-2025-01-04.json
│   ├── 03-question-determine-2025-01-04.md
│   ├── 04-write-or-refactor-2025-01-04.md
│   ├── 05-test-2025-01-04.json
│   ├── 06-lint-2025-01-04.json
│   ├── 07-iterate-2025-01-04.md
│   └── 08-present-2025-01-04.md

Workflow Architecture

File Handling: The server provides suggested paths and formats but does not directly write files. Instead, it instructs the AI assistant to create these files using its own file system access.

Consistent Naming: Files follow a standardized naming convention with phase numbers, names, and timestamps.

Environment Independence: The architecture works across any environment where the AI has appropriate file system permissions.

Graceful Degradation: If the AI is unable to create files, the workflow continues in memory-only mode - your progress isn't interrupted.

Installation

Quick Start (Recommended) - Zero Installation

Add to your AI assistant config - Uses npx automatically:

💡 Note: I recommend using @latest to ensure you always get the newest features and fixes. Without @latest, npx may cache older versions.

VS Code / Cursor / Windsurf - Add to your MCP settings:

{
  "mcp": {
    "servers": {
      "structured-workflow": {
        "command": "npx",
        "args": ["structured-workflow-mcp@latest"],
        "env": {}
      }
    }
  }
}

Claude Desktop - Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "structured-workflow": {
      "command": "npx",
      "args": ["structured-workflow-mcp@latest"],
      "env": {}
    }
  }
}

Global Installation (Optional)

You can install globally on your machine using NPM:

npm install -g structured-workflow-mcp

Then use in your AI assistant config:

{
  "mcp": {
    "servers": {
      "structured-workflow": {
        "command": "structured-workflow-mcp",
        "args": [],
        "env": {}
      }
    }
  }
}

With custom output directory:

{
  "mcp": {
    "servers": {
      "structured-workflow": {
        "command": "structured-workflow-mcp",
        "args": ["--output-dir", "/home/user/workflow-outputs"],
        "env": {}
      }
    }
  }
}

Auto-Install via Smithery

Smithery provides a number of ways to install directly into your apps including this way for Claude Desktop:

npx -y @smithery/cli install structured-workflow-mcp --client claude

Manual Installation

For developers, you can clone the repository and build it locally:

git clone https://github.com/kingdomseed/structured-workflow-mcp
cd structured-workflow-mcp
npm install && npm run build

Usage

Once configured in your AI assistant, start with these workflow tools:

• mcp__structured-workflow__build_custom_workflow - Create custom workflows
• mcp__structured-workflow__refactor_workflow - Structured refactoring
• mcp__structured-workflow__create_feature_workflow - Feature development
• mcp__structured-workflow__test_workflow - Test coverage workflows

Example Output Artifacts

The server enforces that AI produces structured outputs like these:

AUDIT_INVENTORY Phase Output:

{
  "filesAnalyzed": ["lib/auth/user_service.dart", "lib/auth/auth_middleware.dart"],
  "dependencies": {
    "providers": ["userProvider", "authStateProvider"],
    "models": ["User", "AuthToken"]
  },
  "issues": [
    "Single Responsibility Principle violation - handles too many concerns",
    "File approaching 366 lines - recommended to keep widgets smaller"
  ],
  "changesList": [
    {
      "action": "CREATE",
      "file": "lib/auth/components/auth_form.dart",
      "description": "Extract authentication form logic",
      "justification": "Component focused on form validation only"
    }
  ]
}

COMPARE_ANALYZE Phase Output:

{
  "approaches": [
    {
      "name": "Incremental Component Extraction",
      "complexity": "Medium",
      "risk": "Low", 
      "timeEstimate": "30-45 minutes"
    }
  ],
  "recommendation": "Incremental Component Extraction",
  "justification": "Provides best balance of benefits vs. risk",
  "selectedImplementationOrder": [
    "1. Extract form component (lowest risk)",
    "2. Create validation service",
    "3. Refactor main view"
  ]
}

Each phase requires documented analysis and planning before the AI can proceed to implementation.

Tools

Wor

---