Writing an MCP Server (Python)

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("my-tools")

@mcp.tool()
def greet(name: str) -> str:
    """Say hello to someone."""
    return f"Hello, {name}!"

@mcp.tool()
def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b

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

Install: pip install mcp

Wrapping a CLI Tool as MCP

import subprocess
from mcp.server.fastmcp import FastMCP

mcp = FastMCP("git-tools")

@mcp.tool()
def git_log(n: int = 5) -> str:
    """Show recent git commits."""
    return subprocess.run(
        ["git", "log", f"-{n}", "--oneline"],
        capture_output=True, text=True
    ).stdout

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

Any CLI program → MCP tool in 5 lines.

Gemini CLI: Setup

# Install
npm install -g @google/gemini-cli
# or
npx @google/gemini-cli

# Authenticate (uses Google account)
gemini

• Open source (Apache 2.0)
• Uses Gemini 2.5 Pro by default
• 1 million token context window

Gemini CLI: Creating an Agent

Create GEMINI.md in your project root:

# Agent Instructions

You are a code review assistant.

## Rules
- Always check for security issues first
- Suggest fixes, don't just point out problems
- Be concise

## Tools Available
- You can read and write files
- You can run shell commands
- You can use git

GEMINI.md = the system prompt for your agent

Gemini CLI: Enabling MONCEF

Add the prompt to GEMINI.md and configure in .gemini/settings.json:

{
  "customAgents": [{
    "name": "moncef",
    "configPath": "moncef.json"
  }]
}

Then invoke it:

gemini --agent moncef

Gemini CLI: Adding MCP Tools

Create .gemini/settings.json:

{
  "mcpServers": {
    "my-tools": {
      "command": "python3",
      "args": ["mcp_server.py"]
    }
  }
}

Now the agent can call greet and add from our MCP server.

Gemini CLI: Built-in Extensions

• Automatic file read/write and shell execution
• Google Search integration
• Customizable safety thresholds
• Sandbox mode for untrusted operations

# Run with sandbox
gemini --sandbox

# Use a specific model
gemini --model gemini-2.5-flash

Kiro CLI: Setup

# Install
npm install -g kiro-cli

# Start
kiro-cli chat

• Uses Claude (Anthropic) as the LLM
• Built-in code intelligence (LSP integration)
• Subagent delegation for parallel tasks

Kiro CLI: Creating an Agent

Create .kiro/settings/agent.md:

# Agent Instructions

You are a code review assistant.

## Rules
- Always check for security issues first
- Suggest fixes, don't just point out problems
- Be concise

Same idea as GEMINI.md — a markdown file defines the agent's behavior.

Kiro CLI: Enabling MONCEF

Drop the config into .kiro/agents/:

cp moncef.json .kiro/agents/moncef.json

Then invoke it:

kiro-cli chat --agent moncef

allowedTools restricts which built-in tools the agent can access.

Kiro CLI: Adding MCP Tools

Create .kiro/settings/mcp.json:

{
  "mcpServers": {
    "my-tools": {
      "command": "python3",
      "args": ["mcp_server.py"]
    }
  }
}

Same MCP server, different config location. That's the beauty of MCP.

Kiro CLI: Unique Features

• Code Intelligence — LSP-powered symbol search, go-to-definition, find references
• Subagents — delegate tasks to parallel workers
• Planner Mode — Shift+Tab to switch to read-only planning
• AWS Integration — native AWS CLI tool access

Goose: Setup

# Install
brew install block/goose/goose
# or
curl -fsSL https://github.com/block/goose/releases/latest/download/install.sh | sh

# Start
goose session

• Open source (Apache 2.0)
• Supports multiple LLMs (OpenAI, Anthropic, Google…)
• Everything is an MCP extension

Goose: Enabling MONCEF

Drop the config into .goose/recipes/:

cp moncef.json .goose/recipes/moncef.json

Then invoke it:

goose session --recipe .goose/recipes/moncef.json

Same JSON format as Kiro and Gemini — just a different path.

Goose: Adding MCP Tools

Create .goose/recipes/mcp.json:

{
  "mcpServers": {
    "my-tools": {
      "command": "python3",
      "args": ["mcp_server.py"]
    }
  }
}

Same MCP server, same JSON format — Goose treats it as just another extension.

Real Example: Trace Analysis Agent

Real Example: User Interaction

User says:

"Load /tmp/kernel-trace and find CPU anomalies"

Agent does:

1. ensure_server → starts Trace Compass
2. create_experiment → loads the trace
3. detect_anomalies → runs isolation forest
4. Summarizes: "3 CPU spikes found at t=1.2s, 3.9s, 7.1s"