home / skills / yuniorglez / gemini-elite-core / mcp-expert

mcp-expert skill

This skill enables autonomous MCP ecosystem design and secure server orchestration, guiding onboarding, app development, and tool orchestration for

npx playbooks add skill yuniorglez/gemini-elite-core --skill mcp-expert

Review the files below or copy the command above to add this skill to your agents.

Files (5)

SKILL.md

4.4 KB

---
name: mcp-expert
id: mcp-expert
version: 1.4.0
description: "Senior MCP Architect & Orchestrator. Master of MCP Apps, Server Development (2025-11-25 Spec), and Multi-Agent Tooling."
---

# 🛠️ Skill: mcp-expert (v1.4.0)

## Executive Summary
`mcp-expert` is the definitive resource for building and managing Model Context Protocol (MCP) ecosystems. In 2026, MCP is the backbone of autonomous AI agents, enabling them to navigate complex data, interact with legacy APIs, and render dynamic user interfaces. This skill covers the entire lifecycle—from proactive onboarding to advanced MCP App development and high-security server orchestration.

---

## 📋 Table of Contents
1. [Core Capabilities](#core-capabilities)
2. [The "Do Not" List (Anti-Patterns)](#the-do-not-list-anti-patterns)
3. [Quick Start: Onboarding Protocol](#quick-start-onboarding-protocol)
4. [Standard Production Patterns](#standard-production-patterns)
5. [Professional Server Engineering](#professional-server-engineering)
6. [Security & Zero-Trust Auth](#security--zero-trust-auth)
7. [MCP Apps: Interactive Outcomes](#mcp-apps-interactive-outcomes)
8. [Reference Library](#reference-library)

---

## 🚀 Core Capabilities
- **Proactive Onboarding**: Automatically detecting and configuring missing MCP servers.
- **Agentic Tool Design**: Crafting outcome-oriented tool interfaces for LLM excellence.
- **Multimodal Interaction**: Leveraging resources for long-form data (Video, PDF) via MCP URIs.
- **Interactive UI Rendering**: Implementing **MCP Apps** for direct user-agent interaction.
- **Resilient Orchestration**: Managing tool timeouts, pagination, and error recovery.

---

## 🚫 The "Do Not" List (Anti-Patterns)

| Anti-Pattern | Why it fails in 2026 | Modern Alternative |
| :--- | :--- | :--- |
| **Chatty APIs** | Wastes LLM tokens and increases latency. | Use **Outcome-Oriented Tools**. |
| **Raw Exceptions** | LLM cannot self-correct from stack traces. | Return **Helpful Error Strings**. |
| **Hardcoded Secrets** | Massive security vulnerability. | Use **Environment Variable Mapping**. |
| **Stdout Logging** | Breaks the Stdio transport channel. | Log only to **Stderr**. |
| **Monolithic Servers** | Hard to maintain and discover. | Use **Focused, Small Servers** (5-15 tools). |

---

## ⚡ Quick Start: Onboarding Protocol

When an agent needs a new capability (e.g., "Search the web"):

1.  **Validation**: Check if the required MCP server is already active.
2.  **Guide**: If missing, provide the user with a direct installation path.
    - *"I need to browse the web. I'll setup the `browser-use` MCP server now."*
3.  **Config**: Use `uvx` for zero-install execution where possible.

```json
"browser-use": {
  "command": "uvx",
  "args": ["mcp-server-browser-use", "server"],
  "env": { "BROWSER_USE_API_KEY": "..." }
}
```

---

## 🛠 Standard Production Patterns

### Pattern A: The "Progressive Disclosure" Resource
For large datasets, don't return all data. Return a list of URIs.
-   `mcp://sales/2026/summary`
-   `mcp://sales/2026/detailed-report`

### Pattern B: The "Retryable" Tool
Tools that return a `retry_with` suggestion if inputs are slightly off.

---

## ⚙️ Professional Server Engineering

Building a high-performance server in 2026:
-   **Runtime**: Bun.
-   **Validation**: Zod.
-   **Standard**: 2025-11-25 Specification.

*See [References: Server Development](./references/server-development.md) for the blueprint.*

---

## 🔒 Security & Zero-Trust Auth

-   **OAuth 2.1**: Mandatory for enterprise data access.
-   **Capability Scopes**: Limit the agent to the task at hand.
-   **HITL Gate**: Human-in-the-loop approval for destructive actions.

*See [References: Security & Auth](./references/security-auth.md) for implementation.*

---

## 🖥️ MCP Apps: Interactive Outcomes

MCP Apps are the next evolution of AI interfaces.
-   **Live Dashboards**.
-   **Interactive File Editors**.
-   **Visual Selection Tools**.

*See [References: MCP Apps](./references/mcp-apps.md) for architectural details.*

---

## 📖 Reference Library

Detailed deep-dives into MCP excellence:

- [**Server Development**](./references/server-development.md): Professional engineering standards.
- [**Security & Auth**](./references/security-auth.md): Protecting your data and tools.
- [**MCP Apps**](./references/mcp-apps.md): Interactive UI standard for 2026.
- [**Troubleshooting Guide**](./references/troubleshooting.md): Debugging complex interactions.

---

*Updated: January 22, 2026 - 17:35*

Overview

This skill is a senior MCP architect and orchestrator toolkit for building, deploying, and operating Model Context Protocol ecosystems. It focuses on proactive onboarding, resilient orchestration, and developer-grade MCP App patterns that scale in production. The guidance targets server engineering, zero-trust security, and outcome-oriented agent tooling.

How this skill works

The skill inspects running MCP servers and agent toolsets, detects missing capabilities, and provides step-by-step onboarding actions. It defines production patterns for tools, retry semantics, and progressive-disclosure URIs to handle large datasets. It also prescribes server runtime, validation, and auth standards for secure, high-performance deployments.

When to use it

Onboarding an agent that needs new MCP capabilities (e.g., web browsing, file processing).
Designing or hardening MCP Apps and interactive agent UIs.
Implementing production server patterns and validation for MCP tooling.
Building zero-trust authentication and capability-scoped access for agents.
Creating resilient orchestration with retries, timeouts, and error recovery.

Best practices

Prefer outcome-oriented tools over chatty APIs to reduce token use and latency.
Return helpful error strings and retry hints instead of raw exceptions.
Expose large datasets via MCP URIs and progressive disclosure rather than full payloads.
Limit server scope: build small, focused servers (5–15 tools) for easier discovery and maintenance.
Store secrets out of code: use environment mapping and avoid hardcoded credentials.

Example use cases

Auto-detect and provision a missing `browser-use` MCP server with a guided install path.
Create an MCP App that renders a live dashboard and allows interactive filtering via URIs.
Implement a retryable tool that suggests corrected inputs when user data is slightly malformed.
Deploy a Bun-based MCP server with Zod validation and OAuth 2.1 capability scopes for enterprise access.
Add a human-in-the-loop approval gate for destructive agent actions in production workflows.

FAQ

What runtime and validation stack do you recommend?

Use Bun for runtime and Zod for input validation to meet the 2025-11-25 server spec guidance.

How should I expose large documents to agents?

Use progressive disclosure: return MCP URIs for summaries and detail endpoints for on-demand retrieval.

How do I handle sensitive credentials?

Never hardcode secrets. Map credentials from environment variables and use capability-scoped tokens with OAuth 2.1.