Agentic MCP server
Commit Oct 22, 2026Lets an AI client discover and call MCP servers on demand through three-layer lazy loading, reducing token usage without pre-installing full tool lists.
Also known as cablate/Agentic-MCP-Skill
- Stars
- 41
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Oct 22, 2026
- License
- —
What is Agentic MCP?
Agentic lets an AI client discover and call MCP servers on demand through three-layer lazy loading, reducing token usage without pre-installing full tool lists.
Summary generated from the project's README
About Agentic MCP server
Developers looking to optimize AI context window usage can use this tool to manage Model Context Protocol servers through a progressive disclosure strategy. Instead of loading every available tool into an AI client at once, which consumes significant tokens, this system uses a three-layer approach to fetch only the necessary information. It starts with basic server metadata, moves to tool lists, and finally loads specific input schemas only when a tool is ready to be called.
This experimental project operates via a socket-based daemon architecture that maintains persistent connections to various MCP servers like Playwright or filesystem tools. Users configure their environment through a JSON file and interact with the system using a command-line interface. By implementing this lazy loading pattern, developers can achieve substantial token savings, potentially reducing the cost of managing a large toolset by over 80 percent compared to traditional methods.
What you can do with it
- 1Reduce AI token consumptionAn AI client can browse available tools and only load the full input schema for the specific tool it intends to use.
- 2Manage multiple MCP serversA developer can configure and run various MCP servers like Playwright or GitHub through a single daemon.
- 3Automate web browser tasksYou can instruct an AI to navigate websites, take screenshots, or click elements using the Playwright integration.
- 4Update server configurations liveYou can modify your server list in the configuration file and apply the changes without stopping the daemon.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablecablate/agentic-mcp-skill is public and reachablegithub.com/cablate/agentic-mcp-skill ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)No licenseFinding
- Commit or release in the last yearLast commit or release 2026-10-22github.com/cablate/agentic-mcp-skill ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubAgentic-MCP with Skill
AgentSkill for MCP - Three-layer progressive disclosure validates AgentSkills.io pattern for efficient MCP token usage
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:
- Can AgentSkills architecture apply to MCP server management?
- Is three-layer progressive disclosure effective in MCP scenarios?
- Is Socket-based daemon 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
- npm
1. Install from npm (Recommended)
# Install globally
npm install -g @cablate/agentic-mcp
# Verify installation
agentic-mcp --version
# Start daemon
agentic-mcp daemon start
2. Configure
Edit mcp-servers.json in the project root:
{
"servers": {
"playwright": {
"description": "Browser automation tool for web navigation, screenshots, clicks, form filling, and more",
"type": "stdio",
"command": "npx",
"args": ["@playwright/mcp@latest", "--isolated"]
}
}
}
3. Start Daemon
node dist/cli/index.js daemon start --config <config-path>
4. Test Connection
# Check daemon health
agentic-mcp daemon health
# Layer 1: Check server status
agentic-mcp metadata playwright
# Layer 2: List available tools
agentic-mcp list playwright
# Layer 3: View specific tool format
agentic-mcp schema playwright browser_navigate
5. Call Tool
agentic-mcp call playwright browser_navigate --params '{"url": "https://example.com"}'
Testing
This project uses Vitest for testing with the following test suites:
Running Tests
# Run tests in watch mode
npm test
# Run tests once
npm run test:run
# Run with coverage report
npm run test:coverage
Test Structure
tests/
├── unit/
│ ├── client.test.ts # ProgressiveMCPClient tests (12 tests)
│ ├── socket-client.test.ts # SocketClient tests (7 tests)
│ └── cli/
│ └── commands.test.ts # CLI command tests (6 tests)
Current Coverage
- Statements: 61.71%
- Branches: 47.29%
- Functions: 65.85%
- Lines: 61.53%
Test coverage focuses on core modules (client.ts, socket-client.ts). Daemon integration tests are deferred to future work.
Test Philosophy
This project follows a pragmatic progressive approach:
- Focus on happy path testing first
- Use real MCP server (
@modelcontextprotocol/server-filesystem) for integration - Mock only where necessary (e.g.,
netmodule for SocketClient) - Defer complex scenarios (daemon socket tests) to keep initial implementation simple
Usage Examples
Web Automation
Automate browser operations using Playwright MCP server:
# 1. Navigate to website
agentic-mcp call playwright browser_navigate --params '{"url": "https://www.apple.com/tw"}'
# 2. Take screenshot
agentic-mcp call playwright browser_take_screenshot
# 3. Click element
agentic-mcp call playwright browser_click --params '{"element": "Mac link", "ref": "e19"}'
Hot Reload Configuration
Reload after modifying mcp-servers.json without restarting daemon:
agentic-mcp daemon reload
Response example:
✓ Configuration reloaded
Old servers: playwright_global
New servers: playwright_global, filesystem_global
Architecture
System Architecture
+-----------------------------+
| AI / CLI Layer |
| (CLI commands) |
| - agentic-mcp metadata |
| - agentic-mcp list |
| - agentic-mcp schema |
| - agentic-mcp call |
+-----------+-----------------+
| Socket (newline-delimited JSON)
v
+-----------------------------+
| MCP Daemon (Long-Running) |
| - Maintain persistent MCP |
| connections |
| - Socket communication |
| - Manage shared sessions |
| - Support Hot Reload |
+-----------+-----------------+
| MCP Protocol
v
+-----------------------------+
| MCP Servers |
| - playwright (browser) |
| - filesystem (files) |
| - github (Git) |
| - custom servers |
+-----------------------------+
Configuration
MCP Servers Configuration
Edit mcp-servers.json:
{
"servers": {
"playwright": {
"description": "Browser automation tool for web navigation, screenshots, clicks, and form filling",
"transportType": "stdio",
"command": "npx",
"args": ["@playwright/mcp@latest", "--isolated"]
}
}
}
Configuration Notes:
description(optional): Server description for AI to understand the MCP server's purpose. If not provided, Layer 1 metadata will not include this field.
Resources
- SKILL.md - Complete usage guide
- docs/AGENT_BROWSER_DESIGN_PATTERNS.md - Design patterns learned from agent-browser
- MCP Specification
- MCP TypeScript SDK
- AgentSkills.io
This is a concept validation project, feedback welcome
Agentic MCP questions
How does Agentic-MCP reduce MCP token usage?
It loads MCP information in three layers: server metadata, then tool lists, then full schemas only for the tools you actually call. The README claims this saves about 86% of tokens when a server has many tools but you need few.
What commands does agentic-mcp provide?
It exposes CLI commands for metadata, list, schema, and call, plus daemon start, health, and reload. These map to the three disclosure layers and tool invocation.
How do I install and start Agentic-MCP?
Install globally from npm, then start the daemon with the daemon start command using a config file. Node.js 18 or newer is required.
How are MCP servers configured?
Servers are defined in mcp-servers.json with a description, transport type, command, and args. You can reload the daemon after editing so new servers are picked up without restarting.