MCP Progressive AgentSkill

AgentSkill for MCP - Three-layer progressive disclosure validates AgentSkills.io pattern for efficient MCP token usage

English | 繁體中文

---

Design Intent

Proven AgentSkills Pattern

In the AgentSkills.io ecosystem, Progressive Disclosure combined with Script linkage has been validated as effective:

• AI only loads specific information when needed, significantly reducing token usage
• Scripts can be invoked via AI commands, providing high customization potential
• Sharing tools via scripts is far less costly than developing and using complex MCP Servers

The impact of this pattern is significant and substantial.

Current MCP Skill Dilemma

Despite this, many Skills using MCP still face two extreme options:

Option 1: Install MCP directly in Claude Code, Skill only guides AI to call MCP

• MCP server occupies AI context long-term
• Full tool list loaded every conversation
• Token waste, and MCP itself cannot be progressively explored

Option 2: Write custom scripts yourself

• Tests user's programming skills
• High customization but lacks standards
• Each MCP has its own format/specs, high adaptation cost for both open and closed source
• Difficult to maintain and share

This Skill's Goal

Validate whether AgentSkills.io pattern applies to improving MCP usage

This concept validation attempts to port AgentSkills' successful pattern to the MCP domain:

1. Can AgentSkills architecture apply to MCP server management?
2. Is three-layer progressive disclosure effective in MCP scenarios?
3. Is Python scripts + Simple MCP Client architecture more practical than direct MCP usage?

Experimental Nature

This is not a mature product, but an experiment:

• Test AgentSkills pattern applicability in MCP domain
• Explore actual effectiveness of three-layer loading
• Validate pros/cons of Daemon architecture
• Serve as reference prototype for future development

Important Disclaimer

This is a very early, rushed AI-assisted demo version with current goals:

• Validate concept feasibility
• Explore usage patterns
• Collect feedback for improvements

Not recommended for production use. Expect many issues and optimization opportunities. If you find any problems or have suggestions, please open an Issue or contribute a PR.

---

Core Concepts

Three-Layer Progressive Disclosure

Like AgentSkills, you don't need to load everything at once:

Layer 1: Know which servers are available

Load only basic info (name, version, status)
Usage: ~50-100 tokens
Use case: Check availability, choose server

Layer 2: Know what tools this server provides

Load tool list (names + brief descriptions)
Usage: ~200-400 tokens
Use case: Browse available tools, decide what to use

Layer 3: Load only the tools you need

Load complete input format for specific tool
Usage: ~300-500 tokens/tool
Use case: Before calling tool

Why This Approach

Assume an MCP server has 20 tools, you only need 2:

Loading Method	Token Usage	Description
Load All	6,000	20 tools × 300 tokens
Three-Layer Progressive	850	Metadata(50) + List(200) + 2 tools(600)
Savings	86%	Only load what you need

---

Development Status

Future Plans

Based on concept validation results, future directions include:

Short-term Goals

• More convenient MCP Servers management (UI, auto-discovery, one-click install)
• Implement Auth features (API Key management, permission control)

Mid-term Goals

• Enhance MCP Server tool calling experience (better error messages, parameter validation, result formatting)
• Intercept MCP Server output with customizable Script data processing, avoid massive messy data entering conversation memory

This project's direction depends on:

• Concept validation results
• Community feedback
• Actual usage needs

Feedback and suggestions welcome.

---

Quick Start

Prerequisites

• Node.js >= 18.0.0
• Python >= 3.8
• npm

1. Install

python scripts/setup.py

This command automatically checks environment, installs dependencies, and compiles the daemon.

2. Configure

Edit mcp-servers.json:

{
  "servers": {
    "playwright": {
      "transportType": "stdio",
      "command": "npx",
      "args": ["@playwright/mcp@latest", "--isolated"]
    }
  }
}

3. Start Daemon

python scripts/daemon_start.py --no-follow

4. Test Connection

# Layer 1: Check server status
python scripts/mcp_metadata.py --server playwright

# Layer 2: List available tools
python scripts/mcp_list_tools.py --server playwright

# Layer 3: View specific tool format
python scripts/mcp_tool_schema.py --server playwright --tool browser_navigate

5. Call Tool

python scripts/mcp_call.py \
  --server playwright \
  --tool browser_navigate \
  --params '{"url": "https://example.com"}'

---

Usage Examples

Web Automation

Automate browser operations using Playwright MCP server:

# 1. Navigate to website
python scripts/mcp_call.py \
  --server playwright \
  --tool browser_navigate \
  --params '{"url": "https://www.apple.com/tw"}'

# 2. Take screenshot
python scripts/mcp_call.py \
  --server playwright \
  --tool browser_take_screenshot

# 3. Click element
python scripts/mcp_call.py \
  --server playwright \
  --tool browser_click \
  --params '{"element": "Mac link", "ref": "e19"}'

Multi-Tool Batch Execution

Create session.json to execute a series of operations:

[
  {
    "tool": "browser_navigate",
    "params": {"url": "https://example.com"},
    "desc": "Navigate to example.com"
  },
  {
    "tool": "browser_take_screenshot",
    "params": {},
    "desc": "Take screenshot"
  },
  {
    "tool": "browser_click",
    "params": {"element": "Submit", "ref": "e42"},
    "desc": "Click submit button"
  }
]

Execute:

python scripts/mcp_session.py --server playwright --script session.json

Hot Reload Configuration

Reload after modifying mcp-servers.json without restarting daemon:

python scripts/daemon_reload.py

Response example:

{
  "success": true,
  "reloaded": true,
  "oldServers": ["playwright_global"],
  "newServers": ["playwright_global", "filesystem_global"],
  "servers": ["playwright", "filesystem"]
}

---

Architecture

System Architecture

+-----------------------------+
|     AI / Skill Layer        |
|  (Python scripts call API)   |
|  - mcp_metadata.py           |
|  - mcp_list_tools.py         |
|  - mcp_call.py               |
+-----------+-----------------+
            | HTTP (13579)
            v
+-----------------------------+
|   MCP Daemon (Long-Running)  |
|  - Maintain persistent MCP   |
|    connections               |
|  - Provide HTTP API          |
|  - Manage shared sessions    |
|  - Support Hot Reload        |
+-----------+-----------------+
            | MCP Protocol
            v
+-----------------------------+
|        MCP Servers           |
|  - playwright (browser)      |
|  - filesystem (files)        |
|  - github (Git)              |
|  - custom servers            |
+-----------------------------+

Session Management

Session Type	Purpose	Lifecycle
Global Session	Pre-connected, shared by all requests	Daemon start → shutdown
Dynamic Session	Independent connection, specific use	On-demand → manual close

---

Configuration

Environment Variables

Variable	Default	Description
MCP_DAEMON_PORT	13579	Daemon HTTP port

Transport Types

Type	Description	Use Case
stdio	Standard input/output	Local MCP server (default)
http-streamable	HTTP streaming	Remote MCP server
sse	Server-Sent Events	Event-driven MCP server

MCP Servers Configuration

Edit mcp-servers.json:

{
  "servers": {
    "playwright": {
      "transportType": "stdio",
      "command": "npx",
      "args": ["@playwright/mcp@latest", "--isolated"]
    }
  }
}

---

Resources

• SKILL.md - Complete usage guide
• MCP Specification
• MCP TypeScript SDK
• AgentSkills.io

---

This is a concept validation project, feedback welcome