MCP: 260 Tools for AI Models

AI models are brilliant at reasoning but helpless at acting. Model Context Protocol (MCP) changes that - giving models the ability to read files, call APIs, execute code, and interact with the world.

What is MCP?

MCP is a standardized protocol for AI-tool communication:

{
    "tool": "read_file",
    "arguments": {
        "path": "/src/main.py"
    }
}

Any model that speaks MCP can use any MCP tool. Build once, use everywhere.

The Tool Explosion

We've built 260+ MCP tools across categories:

File System (25 tools)

read_file, write_file, list_directory
search_files, glob, grep
file_stats, permissions, watch

Code (40 tools)

run_python, run_javascript, run_bash
linter, formatter, type_checker
git operations, test runners
debugger, profiler

Web (35 tools)

fetch_url, scrape_page
search_google, search_bing
api_call, graphql_query
screenshot, pdf_to_text

Data (30 tools)

sql_query, mongodb_query
pandas_operation, json_transform
csv_read, excel_parse
vector_search, embedding

Communication (20 tools)

send_email, send_slack
calendar_event, reminder
notification, webhook

Blockchain (40 tools)

token_balance, token_transfer
contract_call, contract_deploy
nft_mint, swap_tokens
sign_message, verify_signature

AI/ML (25 tools)

image_generate, image_edit
speech_to_text, text_to_speech
translate, summarize
embedding_compute

System (45 tools)

process_list, process_kill
network_info, disk_usage
env_vars, system_info
docker operations, kubernetes

Usage Examples

Research Assistant

Human: Find recent papers on transformer efficiency and summarize them

Agent uses:
1. search_google("transformer efficiency papers 2024")
2. fetch_url(paper_links)
3. pdf_to_text(papers)
4. summarize(combined_text)
5. write_file("summary.md", result)

Code Review Bot

Human: Review the changes in this PR

Agent uses:
1. git_diff("main", "feature-branch")
2. read_file(changed_files)
3. linter(files)
4. type_checker(files)
5. send_comment(pr_number, review)

Data Analysis

Human: Analyze our sales data and create a report

Agent uses:
1. sql_query("SELECT * FROM sales")
2. pandas_operation("group by month")
3. chart_create(data, "bar")
4. write_file("report.html", combined)
5. send_email(stakeholders, report)

Tool Development

Creating new tools is straightforward:

from mcp import Tool, Param

@Tool(
    name="weather",
    description="Get current weather for a location",
    params=[
        Param("location", "string", "City name or coordinates"),
        Param("units", "string", "celsius or fahrenheit", optional=True)
    ]
)
async def weather(location: str, units: str = "celsius"):
    # Implementation
    data = await weather_api.get(location)
    return format_weather(data, units)

Security Model

Tools have permission scopes:

tools:
  read_file:
    scope: filesystem
    paths: ["/project/**"]  # Limited to project

  run_bash:
    scope: execution
    allowed_commands: ["npm", "python", "git"]

  send_email:
    scope: communication
    requires_confirmation: true

Users control what tools can access.

MCP Registry

Tools are discoverable through our registry:

# Search for tools
mcp search "database"

# Install a tool
mcp install sql-tools

# List installed
mcp list

Integration

MCP works with major AI systems:

Claude: Native MCP support
GPT-4: Via function calling adapter
Llama: Via tool use adapter
Zen: Native support

from hanzo import Agent
from mcp import MCPToolkit

# Load all installed tools
toolkit = MCPToolkit.load()

# Create agent with tools
agent = Agent(
    model="zen-coder-34b",
    tools=toolkit.all()
)

# Agent can now use any tool
response = agent.run("Deploy the app to production")

What's Next

Tool composition: Combine tools into workflows
Learning: Models learn new tools from examples
Verification: Prove tools do what they claim
Marketplace: Share and sell custom tools

MCP is how AI becomes useful - not through better prompting, but through better tools.

This post is part of our retrospective series exploring the technical foundations of Hanzo.