Hermes Agent: MCP Protocol

Q: How many MCP servers can be connected simultaneously?

No hard limit. 5-10 servers is perfectly normal in practice. Each server runs as an independent process.

Q: Is MCP communication secure?

stdio mode communication is entirely local. SSE mode supports TLS. Authentication is passed via environment variables, not exposed in the protocol.

Q: How to debug MCP connection issues?

hermes mcp test <server> tests connection, --debug shows communication details.

Last updated: 2026-08-31

MCP is the USB port for AI tool interconnection — through Model Context Protocol, Hermes Agent can discover and invoke any MCP-compatible tool server, infinitely expanding capability boundaries.

💡 Tip: MCP (Model Context Protocol) is an open protocol proposed by Anthropic that standardizes interconnection between AI applications and tool servers. Hermes natively supports MCP and can connect to any MCP server.

📋 Prerequisites: Lesson 8 Tools and Toolsets, Lesson 15 Plugins

1. What You Will Learn

#	Content
❶	MCP protocol concepts and principles
❷	Hermes' MCP integration
❸	MCP server configuration
❹	Tool discovery and invocation
❺	Custom MCP server development

2. Story

(1) Pain Point: Every AI Tool Has a Different Interface

Bob uses 5 AI tools, each with its own API format, authentication method, and tool definitions. Integrating a new tool means developing a new adapter.

(2) Solution: MCP Unified Protocol, Plug and Play

Alice connects Hermes to various tool servers via MCP:

BASH

# One line of config to connect an MCP server
hermes mcp add filesystem --command "npx @anthropic/mcp-filesystem /home/alice"
hermes mcp add github --command "npx @anthropic/mcp-github"

# Hermes automatically discovers all available tools
hermes mcp tools
# Discovered 15 new tools: read_file, write_file, search_files, create_issue...

3. MCP Protocol Concepts

(1) What is MCP?

MCP (Model Context Protocol) is an open protocol that defines the communication standard between AI applications (clients) and tool servers:

graph LR
    A[Hermes Agent<br/>MCP Client] -->|Standard Protocol| B[MCP Server 1<br/>Filesystem]
    A -->|Standard Protocol| C[MCP Server 2<br/>GitHub]
    A -->|Standard Protocol| D[MCP Server 3<br/>Database]
    A -->|Standard Protocol| E[MCP Server N<br/>Custom]

(2) MCP Core Concepts

Concept	Description	Analogy
MCP Client	AI application making requests	USB Host
MCP Server	Server providing tools	USB Device
Tool	Specific operation provided by Server	Device function
Resource	Data source provided by Server	Device storage
Prompt	Prompt template provided by Server	Device preset

(3) MCP vs Traditional API

Dimension	Traditional API	MCP
Interface format	Custom per service	Unified standard
Tool discovery	Manual documentation	Auto-discovery
Authentication	Custom implementation	Unified approach
Connection method	HTTP/SDK	stdio/SSE
Ecosystem	Siloed	Interconnected

4. Hermes' MCP Integration

(1) Configure MCP

YAML

# ~/.hermes/config.yaml
mcp:
  enabled: true
  
  # MCP server list
  servers:
    filesystem:
      command: "npx"
      args: ["@anthropic/mcp-filesystem", "/home/alice/projects"]
      
    github:
      command: "npx"
      args: ["@anthropic/mcp-github"]
      env:
        GITHUB_TOKEN: "${GITHUB_TOKEN}"
        
    database:
      command: "python"
      args: ["-m", "mcp_server_postgres"]
      env:
        DATABASE_URL: "${DATABASE_URL}"
        
    brave-search:
      command: "npx"
      args: ["@anthropic/mcp-brave-search"]
      env:
        BRAVE_API_KEY: "${BRAVE_API_KEY}"

(2) Connection Management

BASH

# Add MCP server
hermes mcp add filesystem --command "npx @anthropic/mcp-filesystem /home/alice"

# List configured servers
hermes mcp list

# Test connection
hermes mcp test filesystem

# View tools provided by server
hermes mcp tools filesystem

# Remove server
hermes mcp remove filesystem

5. Tool Discovery and Invocation

(1) Auto-Discovery

After connecting to an MCP server, Hermes automatically discovers all available tools:

BASH

hermes mcp tools

# Output example:
# ┌─────────────────────┬──────────────┬────────────────────────┐
# │ Tool                │ Server       │ Description            │
# ├─────────────────────┼──────────────┼────────────────────────┤
# │ read_file           │ filesystem   │ Read file contents     │
# │ write_file          │ filesystem   │ Write to file          │
# │ search_files        │ filesystem   │ Search in files        │
# │ create_issue        │ github       │ Create GitHub issue    │
# │ list_prs            │ github       │ List pull requests     │
# │ query               │ database     │ Execute SQL query      │
# │ search              │ brave-search │ Web search             │
# └─────────────────────┴──────────────┴────────────────────────┘

(2) Use in Conversation

BASH

me: Read the contents of README.md
Agent: [MCP: filesystem/read_file]
       # My Project...

me: Create an Issue on GitHub
Agent: [MCP: github/create_issue]
       Title? → "Bug: Login fails on mobile"
       ✅ Issue #42 created

me: Search for the latest AI news
Agent: [MCP: brave-search/search]
       Found 5 results...

(3) Tool Routing

When multiple MCP servers provide similar tools, Hermes auto-routes:

YAML

mcp:
  routing:
    # Prefer MCP tools
    prefer_mcp: true
    
    # Priority for same-named tools
    priority:
      filesystem.read_file: 100     # MCP version preferred
      fs_read: 50                   # Built-in version secondary
      
    # Conflict resolution
    conflict_resolution: "priority"  # priority / server_order / ask_user

6. Custom MCP Server Development

(1) Python MCP Server

PYTHON

# mcp_server_custom.py
from mcp.server import Server, Tool

server = Server("custom-tools")

@server.tool("get_weather")
async def get_weather(city: str) -> dict:
    """Get current weather for a city"""
    # Custom logic
    import httpx
    response = httpx.get(f"https://api.weather.com/{city}")
    return response.json()

@server.tool("send_notification")
async def send_notification(message: str, channel: str = "default") -> dict:
    """Send a notification to a channel"""
    # Custom notification logic
    return {"status": "sent", "channel": channel}

if __name__ == "__main__":
    server.run()

(2) TypeScript MCP Server

TYPESCRIPT

// mcp-server-custom.ts
import { Server, Tool } from "@anthropic/mcp";

const server = new Server("custom-tools");

server.tool("query_database", {
  description: "Query the database",
  parameters: {
    sql: { type: "string", required: true },
    limit: { type: "number", default: 100 }
  },
  async execute({ sql, limit }) {
    // Custom database query
    const results = await db.query(sql, { limit });
    return { rows: results };
  }
});

server.run();

(3) Register Custom Server

BASH

# Register custom MCP server
hermes mcp add custom-tools \
  --command "python mcp_server_custom.py" \
  --working-dir /path/to/server

# Or configure in config.yaml

YAML

mcp:
  servers:
    custom-tools:
      command: "python"
      args: ["mcp_server_custom.py"]
      working_dir: "/path/to/server"
      env:
        API_KEY: "${CUSTOM_API_KEY}"

7. MCP Ecosystem

(1) Official MCP Servers

Server	Function	Install
filesystem	File system operations	`npx @anthropic/mcp-filesystem`
github	GitHub operations	`npx @anthropic/mcp-github`
gitlab	GitLab operations	`npx @anthropic/mcp-gitlab`
postgres	PostgreSQL queries	`npx @anthropic/mcp-postgres`
sqlite	SQLite operations	`npx @anthropic/mcp-sqlite`
brave-search	Web search	`npx @anthropic/mcp-brave-search`
google-drive	Google Drive	`npx @anthropic/mcp-google-drive`
slack	Slack operations	`npx @anthropic/mcp-slack`

(2) Community MCP Servers

The community maintains 100+ MCP servers covering:

Project management: Jira, Linear, Asana
Design tools: Figma, Canva
Cloud services: AWS, GCP, Azure
Data analysis: Pandas, Jupyter
Communication: Twilio, SendGrid

❓ FAQ

Q What's the difference between MCP and plugins?

A Plugins are Hermes-specific extensions with deep integration. MCP is an open protocol any AI application can use. MCP is more universal, plugins are more deeply integrated.

Q Do MCP servers affect performance?

A MCP servers are independent processes with zero consumption when idle. Invocation uses stdio/SSE communication, typically <50ms latency.

Q How many MCP servers can be connected simultaneously?

A No hard limit. 5-10 servers is perfectly normal in practice. Each server runs as an independent process.

Q Is MCP communication secure?

A stdio mode communication is entirely local. SSE mode supports TLS. Authentication is passed via environment variables, not exposed in the protocol.

Q How to debug MCP connection issues?

A hermes mcp test <server> tests connection, --debug shows communication details.

Q Does MCP support streaming responses?

A Yes. SSE mode supports streaming returns, suitable for long-running tools.

📖 Summary

MCP is the open protocol for AI tool interconnection, like a USB port
Hermes natively supports MCP Client, connecting to any MCP server
Auto-discovery of tools with unified invocation interface
Supports Python/TypeScript custom MCP server development
Official + community 100+ MCP servers covering various scenarios

📝 Exercises

Basic (⭐): Configure an MCP server (e.g., filesystem), verify tool discovery and invocation.
Intermediate (⭐⭐): Configure 3 MCP servers simultaneously, use tools from different servers in conversation.
Advanced (⭐⭐⭐): Develop a custom MCP server wrapping your commonly-used API, register it with Hermes and verify the complete workflow.

Web-Tutorial Tech Team

A team of developers maintaining programming tutorials. Each tutorial is written and reviewed by developers with expertise in that field. We work to keep our content accurate and reliable — if you spot an issue, please let us know.

About Us

Hermes Agent: MCP Protocol

1. What You Will Learn

2. Story

(1) Pain Point: Every AI Tool Has a Different Interface

(2) Solution: MCP Unified Protocol, Plug and Play

3. MCP Protocol Concepts

(1) What is MCP?

(2) MCP Core Concepts

(3) MCP vs Traditional API

4. Hermes' MCP Integration

(1) Configure MCP

(2) Connection Management

5. Tool Discovery and Invocation

(1) Auto-Discovery

(2) Use in Conversation

(3) Tool Routing

6. Custom MCP Server Development

(1) Python MCP Server

(2) TypeScript MCP Server

(3) Register Custom Server

7. MCP Ecosystem

(1) Official MCP Servers

(2) Community MCP Servers

❓ FAQ

📖 Summary

📝 Exercises

🙏 帮我们做得更好