Executor State Machine MCP Server (arc-mcp)

  ▄▄▄▄   ▄▄▄▄▄▄▄    ▄▄▄▄▄▄▄       ▄▄▄      ▄▄▄  ▄▄▄▄▄▄▄ ▄▄▄▄▄▄▄   
▄██▀▀██▄ ███▀▀███▄ ███▀▀▀▀▀       ████▄  ▄████ ███▀▀▀▀▀ ███▀▀███▄ 
███  ███ ███▄▄███▀ ███            ███▀████▀███ ███      ███▄▄███▀ 
███▀▀███ ███▀▀██▄  ███      ▀▀▀▀▀ ███  ▀▀  ███ ███      ███▀▀▀▀   
███  ███ ███  ▀███ ▀███████       ███      ███ ▀███████ ███       

An execution state machine implemented as a Model Context Protocol (MCP) server via standard I/O. This server manages step-by-step execution of structured plans defined in YAML, providing a reliable way for AI agents and clients to follow complex, multi-step procedures while tracking execution status and conserving token limits.

🚀 Features

• Plan Initialization: Parse and load sequential execution plans from YAML format.

• State Management: Tracks execution machine states seamlessly (idle, running, halted, completed).

• Token Optimization: Automatically strips out non-executable logging data (e.g., intent) from payload steps when retrieved, keeping context windows lightweight.

• Robust Execution Tracking: Built-in tools for cleanly advancing steps or reporting structured failure modes.

• Standard Protocol: Built on the official @modelcontextprotocol/sdk using the StdioServerTransport.

Related MCP server: Accordo

🧠 The Architect & Plan Generation

This server executes strictly formatted YAML plans. To generate these plans reliably, you should use a highly capable reasoning model (acting as the Architect).

We provide a strict specification and instruction manual for the Architect AI. Simply provide the contents of docs/plan-format.spec.md to your Architect agent (as a system prompt or context document) to ensure it correctly emits YAML plans that this Executor state machine can parse, execute, and mechanically verify.

🛠️ Provided MCP Tools

This server exposes the following tools to the connected MCP client:

1. begin_plan

• Description: Initializes a new plan from a YAML string and sets the state machine to running.

• Parameters:

  • plan_yaml (string, required): The execution plan strictly formatted in YAML.

2. get_next_step

• Description: Retrieves the next executable step. Strips logging data to save tokens.

• Parameters: None

3. mark_step_complete

• Description: Marks the current step as complete, validates the step id, and advances the state machine to the next step.

• Parameters:

  • id (string, required): The ID of the step to mark complete.

4. report_failure

• Description: Halts the plan, logs the failure, and returns a structured diagnostic report meant for the orchestrating Architect. Note: captured output is securely truncated to the last 1000 characters.

• Parameters:

  • id (string, required): The step ID where the failure occurred.

  • reason (string, required): The explanation of the failure.

  • actual (number, optional): Actual outcome/metric.

  • captured (string, optional): Captured stdout, stderr, or context logs.

📦 Prerequisites & Installation

• Node.js v18 or higher (due to underlying MCP SDK requirements)

• npm

1. Clone or download the repository.

2. Install the necessary dependencies:

   npm install

⚙️ Configuration (MCP Client integration)

To integrate this server with your MCP client (e.g., Claude Desktop, custom MCP-enabled IDE), add the following to your MCP settings file.

Note: Be sure to replace YOUR_FULL_PATH_HERE with the actual absolute directory path containing your index.js file:

{
  "mcpServers": {
    "arc-mcp": {
      "command": "node",
      "args": [
        "YOUR_FULL_PATH_HERE/arc-mcp/index.js"
      ],
      "env": {}
    }
  }
}

📜 Dependencies

• @modelcontextprotocol/sdk (^1.29.0)

• js-yaml (^4.2.0)