nicolas-codemate/mcp-tool-creator

MCP Tool Creator

Create a project-scoped MCP server that wraps docker compose exec commands into named tools. The generated server lives in the project and is registered in .mcp.json.

Workflow

Five interactive steps. Ask for user input at each step before proceeding.

Step 1: Discover

<instructions> 1. Read `docker-compose.yml` (or `docker-compose.yaml`, `compose.yml`, `compose.yaml`) from project root 2. Parse the `services:` section to extract service names and images 3. For each service, detect the technology from the image name: - `php`, `php-fpm`, `symfony` -> PHP/Symfony - `node`, `bun`, `deno` -> Node.js - `postgres`, `pgsql` -> PostgreSQL - `mysql`, `mariadb` -> MySQL - `redis` -> Redis - `python`, `django`, `flask` -> Python - `elasticsearch`, `elastic`, `opensearch` -> Elasticsearch - Check `build:` context and Dockerfile for additional hints if image is custom 4. Present detected services with their technologies in a table 5. Ask the user: - Which services to expose as MCP tools (default: all detected) - Runtime preference: Python FastMCP (recommended) or Node.js SDK </instructions> <output> Display a table: | Service | Image | Detected Tech | Include? | Then ask for confirmation and runtime choice. </output>

Step 2: Design

<instructions> 1. Load `references/docker-tool-patterns.md` 2. For each selected service, match patterns from the reference based on detected technology 3. Propose a list of tools with: name, command, parameters, description 4. Include relevant "General Docker" tools (logs, ps, restart) 5. Ask the user to: - Confirm, remove, or add tools - Customize tool names if desired - Add custom tools not in the patterns list 6. For database services, ask for default credentials (user, database name) from the compose env vars </instructions> <output> Numbered list of proposed tools: 1. `php_lint` - Check PHP syntax (service: php) 2. `run_tests` - Run PHPUnit (service: php) ... Ask: "Confirm this list? Add/remove tools? (numbers to remove, or describe custom tools)" </output>

Step 3: Generate

<instructions> 1. Read the appropriate template from `assets/`: - Python: `assets/fastmcp-server-template.py` - Node.js: `assets/nodejs-server-template.js` 2. Copy the template to `mcp-tools/server.py` (or `server.js`) at the project root 3. Replace `{{SERVER_NAME}}` with a descriptive name based on the project (e.g., `myapp-tools`) 4. Generate tool functions for each validated tool from Step 2

Python tool generation pattern:

@mcp.tool()
async def tool_name(param: str = "default") -> str:
    """Tool description.

    Args:
        param: Parameter description.
    """
    return await run_docker_command("service", ["command", param])

Node.js tool generation pattern: Add to TOOLS array:

{
  name: "tool_name",
  description: "Tool description.",
  inputSchema: {
    type: "object",
    properties: {
      param: { type: "string", description: "Parameter description" },
    },
    required: [],
  },
},

Add to HANDLERS map:

tool_name: async (args) => {
  return await runDockerCommand("service", ["command", args.param || "default"]);
},

5. Replace the {{TOOLS_MARKER}} / {{HANDLERS_MARKER}} comments with generated code
6. For database tools with credentials, read defaults from docker-compose.yml env vars
7. Show the generated file to the user for review </instructions>

<constraints> - Keep `run_docker_command` / `runDockerCommand` helper unchanged from template - One tool = one function (Python) or one TOOLS entry + one HANDLERS entry (Node.js) - Always use `-T` flag for docker compose exec (non-interactive) - Tool parameters must have sensible defaults where possible - Never hardcode passwords in tool code; read from environment variables </constraints>

Step 4: Configure

<instructions> 1. Check if `.mcp.json` exists at the project root 2. If it exists, read it and merge the new server entry 3. If it does not exist, create it 4. Add the server configuration:

Python server (system python3):

{
  "mcpServers": {
    "{{SERVER_NAME}}": {
      "command": "python3",
      "args": ["mcp-tools/server.py"]
    }
  }
}

Python server (venv, when system Python is locked):

{
  "mcpServers": {
    "{{SERVER_NAME}}": {
      "command": "mcp-tools/.venv/bin/python",
      "args": ["mcp-tools/server.py"]
    }
  }
}

Node.js server:

{
  "mcpServers": {
    "{{SERVER_NAME}}": {
      "command": "node",
      "args": ["mcp-tools/server.js"]
    }
  }
}

5. Show the resulting .mcp.json to the user </instructions>

<constraints> - Never overwrite existing MCP server entries in `.mcp.json` - Use relative paths for both `command` and `args` — do NOT set `cwd` (Claude Code resolves paths relative to `.mcp.json` location) - Preserve existing JSON formatting and entries </constraints>

Step 5: Validate

<instructions> 1. **Syntax check:** - Python: run `python3 -c "import ast; ast.parse(open('mcp-tools/server.py').read())"` - Node.js: run `node --check mcp-tools/server.js` 2. **Dependencies check:** - Python: verify `mcp` package is available (`python3 -c "from mcp.server.fastmcp import FastMCP"`) - If import fails (PEP 668 / system Python locked), create a venv: `python3 -m venv mcp-tools/.venv && mcp-tools/.venv/bin/pip install mcp` - Then use `mcp-tools/.venv/bin/python` as the command in `.mcp.json` - Node.js: verify `@modelcontextprotocol/sdk` is in package.json or node_modules 3. If dependencies are missing, provide install command: - Python: `pip install mcp`, `uv pip install mcp`, or venv approach above - Node.js: `npm install @modelcontextprotocol/sdk` 4. **Startup test:** attempt to import/parse the server to catch runtime errors 5. Report results and next steps: - Remind user to restart Claude Code to pick up the new MCP server - List the tools that will be available - Suggest adding `mcp-tools/` to `.gitignore` if not tracked </instructions> <output> Summary: - Server: mcp-tools/server.py (or .js) - Tools: N tools for M services - Config: .mcp.json updated - Status: Ready / Issues found

Next steps:

1. Restart Claude Code (or reload MCP servers)
2. The following tools are now available: [list] </output>