Superskills

superskills-mcp is a lightweight, globally installable Model Context Protocol (MCP) gateway. It allows you to effortlessly expose your local bash scripts, python tools, or Node.js skills to any AI assistant (ChatGPT, Claude Desktop, Cursor, etc.) via a unified interface.

Instead of writing a custom MCP server for every small script you create, superskills-mcp acts as a dynamic registry. You simply drop your skill's path into the configuration, and it is instantly available to the AI.

---

✨ Key Features

• 🌍 Global Installation: Run it from anywhere on your system via the superskills-mcp CLI.
• 🤖 Agent Mode (New!): Empower ChatGPT to "think" and "act" by exposing raw primitives (run, read, write) so it can follow a skill's SKILL.md step-by-step.
• 🔌 Dynamic Tool Registration: Add or remove skills dynamically without touching the server code.
• 🔄 Hot Reloading: Apply configuration changes on the fly with zero downtime using superskills-mcp reload.
• 📝 Native Logging: Built-in background daemon management and real-time log tailing.
• 🛡️ Secure Execution: Strict path validations, restricted runner boundaries, and input schema validation via Zod.
• 🌐 Multi-Transport Support: Expose tools over HTTP (for ChatGPT/Ngrok) or Stdio (for local Claude/Cursor).
• 📊 Web Dashboard: An integrated, premium dark-themed UI to monitor tool usage, toggle skills, and manage your local environment.
• 🕳️ Integrated Ngrok: Built-in tunneling support—expose your local tools to the internet with a single config line.
• 📓 Built-in Notes: Native markdown note-taking management tools (list, read, write) included out of the box.

---

🚀 Quick Start

1. Installation

Install the CLI tool globally via npm or pnpm:

npm install -g superskills-mcp

2. Initialization

Generate your global configuration file (this will safely create ~/.superskills/mcp-config.json):

superskills-mcp init

3. Start the Gateway

Launch the MCP server and run it in the background after logging the public URL:

superskills-mcp serve -d

(You can verify it is running by typing superskills-mcp status or by checking curl http://127.0.0.1:8787/health)

---

🤖 Agent Mode (Infrastructure Tools)

Traditional MCP tools are "black boxes"—the AI calls a function and gets a result. However, many complex skills (like baoyu-post-to-wechat) are designed as instruction-based agents.

superskills-mcp now provides Agent Mode, a set of infrastructure tools that give the AI the raw primitives needed to follow a skill's SKILL.md instructions manually.

Available Agent Tools:

Tool	Description
superskills_list_skills	List all registered skills, their directories, and descriptions.
superskills_invoke	Read a skill's SKILL.md (with resolved placeholders) to understand its instructions.
superskills_run	Execute any shell command (e.g., bun, python, bash) and return output.
superskills_read_file	Read the content of any local file.
superskills_write_file	Write or append content to a local file.
superskills_list_dir	List contents of any local directory.
superskills_env	Read credentials from .baoyu-skills/.env files.

How to use: Simply tell ChatGPT: "Invoke the skill 'baoyu_post_to_wechat' and follow its instructions to publish this file." ChatGPT will then read the manual, check your credentials, and run the necessary scripts step-by-step.

🔐 Agent Policy

When exposing Agent Mode over HTTP or ngrok, you should restrict what the model can read, write, and execute.

Add agentPolicy to your config:

{
  "agentPolicy": {
    "enabledTools": {
      "superskills_invoke": true,
      "superskills_run": true,
      "superskills_read_file": true,
      "superskills_write_file": false,
      "superskills_list_dir": true,
      "superskills_env": false
    },
    "allowedPaths": [
      "{notesDir}",
      "~/Documents/superskills-workspace"
    ],
    "allowedCommands": ["bun", "node", "python3"],
    "blockedCommands": ["rm", "sudo", "ssh", "scp", "curl", "wget", "git"],
    "allowEnv": false,
    "allowSkillInvoke": true
  }
}

Skill-level agentPolicy entries inherit and append to the global policy. They can add paths and commands, but they should not weaken global safety rules.

Recommended remote profile

For long-running ChatGPT-facing deployments:

• enable superskills_invoke, superskills_read_file, superskills_list_dir
• disable superskills_write_file unless needed
• disable superskills_env by default
• keep allowedCommands minimal
• use workspace-only or notes-only allowedPaths

---

🛠️ CLI Reference

superskills-mcp comes with a powerful suite of management commands akin to PM2 or Nginx:

Command	Description
superskills-mcp init	Initialize the global config at ~/.superskills/mcp-config.json.
superskills-mcp serve	Start the MCP server using the global config.
superskills-mcp status	Check if the server is actively running in the background.
superskills-mcp stop	Gracefully terminate the running server.
superskills-mcp reload	Smoothly restart the server to apply configuration changes instantly.
superskills-mcp list	Print a formatted list of all currently registered skills.
superskills-mcp add <path>	Auto-parse and add a new local skill directory to your global config.
superskills-mcp remove <name>	Unregister a skill from your global config by its name.
superskills-mcp log	Tail the real-time background logs ([INFO] and [ERROR]).
superskills-mcp update	Self-upgrade the superskills-mcp package to the latest version.

---

📓 Built-in Tools (Notes)

superskills-mcp provides a set of native tools for managing markdown notes. This is perfect for letting AI assistants maintain a local knowledge base or journal for you.

To enable it, add the notes block to your config:

{
  "notes": {
    "dir": "/Users/username/Documents/Notes"
  }
}

This will automatically expose:

• notes_list: List all markdown files in the directory.
• notes_read: Read the content of a specific note.
• notes_write: Create or update a note with given content.

---

🕳️ Integrated Ngrok Tunneling

No more running ngrok in a separate terminal. superskills-mcp can automatically establish a secure tunnel for you.

Simply add your Ngrok token to the server config:

{
  "server": {
    "transport": "http",
    "port": 8787,
    "ngrokToken": "your_ngrok_auth_token",
    "ngrokDomain": "your-optional-custom-domain.ngrok-free.app"
  }
}

When you run superskills-mcp serve, it will log the public URL. Use {url}/mcp as the Action URL in ChatGPT.

---

📊 Web Dashboard

superskills-mcp 现在内置了一个精美的 Web 控制面板。你可以通过它实时监控技能的调用次数，并随时开启或禁用特定的技能。

1. 确保服务器正在运行 (superskills-mcp serve -d)。
2. 在浏览器中打开：http://127.0.0.1:8787/dashboard/

该面板完全集成在 Express 服务器中，无需安装任何额外依赖，即插即用。

---

📦 Managing Skills

Adding Skills Automatically

You can use the CLI to dynamically attach a new local skill:

superskills-mcp add /path/to/your/custom_skill

Note: This command extracts the skill name and description. superskills-mcp will attempt to automatically infer the input schema from the script source code or a .superskills.json sidecar file.

Removing Skills

To detach a skill:

superskills-mcp remove my_custom_skill

Applying Changes

After adding, removing, or manually editing your config, reload the gateway:

superskills-mcp reload

---

🔍 Auto-Discovery (Zero Configuration)

If you have a collection of skills in a folder, you can let superskills-mcp discover them automatically.

Any sub-directory containing a SKILL.md file will be recognized as a tool.

1. Open ~/.superskills/mcp-config.json.
2. Add scanRoots and scanSettings:

{
  "scanRoots": [
    "~/.baoyu-skills/skills"
  ],
  "scanSettings": {
    "watch": true,
    "ignore": ["node_modules", ".git"]
  }
}

• scanRoots: An array of directories to scan for sub-folders containing SKILL.md.
• watch: If true, the server will monitor these directories in real-time. Adding or removing a folder will instantly update the available tools for your AI assistant—no restart required.

---

⚙️ Configuration File Structure

Your global configuration lives at ~/.superskills/mcp-config.json.

{
  "server": {
    "name": "superskills-mcp",
    "version": "0.6.7",
    "transport": "http",
    "host": "127.0.0.1",
    "port": 8787,
    "ngrokToken": "",
    "ngrokDomain": ""
  },
  "notes": {
    "dir": "~/Documents/mcp-notes"
  },
  "defaults": {
    "timeoutMs": 120000,
    "maxOutputBytes": 10485760,
    "runner": {
      "command": "bun",
      "args": ["{serverDir}/scripts/mcp-adapter.ts"]
    }
  },
  "skills": [
    {
      "name": "read_x_to_markdown",
      "description": "Read a given X/Twitter link and return clean Markdown.",
      "skillDir": "/Users/username/my-skills/twitter-scraper",
      "input": {
        "url": {
          "type": "string",
          "format": "uri",
          "description": "X/Twitter post URL to read"
        }
      },
      "env": {
        "KEEP_OUTPUT": "false"
      }
    }
  ]
}

• defaults.runner.args: The {serverDir} placeholder is automatically replaced with this gateway's actual installation path.
• skills[].input: Defines the exact JSON schema that the AI must fulfill when calling your tool.
• skills[].env: Environment variables scoped specifically to that single skill's execution.

---

🔗 Connecting AI Assistants

🌐 Connecting to ChatGPT (HTTP Transport)

For cloud-based AI like ChatGPT, the server must be exposed to the internet.

1. Recommended: Use the Integrated Ngrok Tunneling by running superskills-mcp serve -d.
2. Alternative: Expose the local port via a persistent Ngrok tunnel manually:

ngrok http --domain=your-free-static-domain.ngrok-free.app 8787

3. In your ChatGPT MCP Configuration, provide the proxy URL with the /mcp path: https://your-free-static-domain.ngrok-free.app/mcp

🖥️ Connecting to Claude Desktop / Cursor (Stdio Transport)

Local clients prefer communicating via standard input/output (stdio). You do not need to run serve & manually for these clients; they will spawn the gateway themselves.

Add this block to your local client config (e.g., ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "superskills-mcp": {
      "command": "superskills-mcp",
      "args": ["serve", "--transport", "stdio"]
    }
  }
}

---

🔒 Security Principles

• Isolated Execution: Skills are executed using restricted runners without shell interpolation (shell: false), eliminating command injection risks.
• Directory Verification: The gateway strictly validates that runner scripts exist within the trusted serverDir or the specific skillDir.
• Resource Limits: Configurable timeoutMs and maxOutputBytes prevent runaway scripts from crashing your system.