pyATS MCP Server

This project implements a Model Context Protocol (MCP) Server that wraps Cisco pyATS and Genie functionality. It enables structured, model-driven interaction with network devices over STDIO using the JSON-RPC 2.0 protocol.

🚨 This server does not use HTTP or SSE. All communication is done via STDIN/STDOUT (standard input/output), making it ideal for secure, embedded, containerized, or LangGraph-based tool integrations.

🔧 What It Does

Connects to Cisco IOS/NX-OS devices defined in a pyATS testbed

Supports safe execution of validated CLI commands (show, ping)

Allows controlled configuration changes

Returns structured (parsed) or raw output

Exposes a set of well-defined tools via tools/discover and tools/call

Operates entirely via STDIO for minimal surface area and maximum portability

🚀 Usage

1. Set your testbed path

export PYATS_TESTBED_PATH=/absolute/path/to/testbed.yaml

2. Run the server

Continuous STDIO Mode (default)

python3 pyats_mcp_server.py

Launches a long-running process that reads JSON-RPC requests from stdin and writes responses to stdout.

One-Shot Mode

echo '{"jsonrpc": "2.0", "id": 1, "method": "tools/discover"}' | python3 pyats_mcp_server.py --oneshot

Processes a single JSON-RPC request and exits.

📦 Docker Support

Build the container

docker build -t pyats-mcp-server .

Run the container (STDIO Mode)

docker run -i --rm \
  -e PYATS_TESTBED_PATH=/app/testbed.yaml \
  -v /your/testbed/folder:/app \
  pyats-mcp-server

🧠 Available MCP Tools

Tool Description

run_show_command Executes show commands safely with optional parsing

run_ping_command Executes ping tests and returns parsed or raw results

apply_configuration Applies safe configuration commands (multi-line supported)

learn_config Fetches running config (show run brief)

learn_logging Fetches system logs (show logging last 250)

All inputs are validated using Pydantic schemas for safety and consistency.

🤖 LangGraph Integration

Add the MCP server as a tool node in your LangGraph pipeline like so:

("pyats-mcp", ["python3", "pyats_mcp_server.py", "--oneshot"], "tools/discover", "tools/call")

Name: pyats-mcp

Command: python3 pyats_mcp_server.py --oneshot

Discover Method: tools/discover

Call Method: tools/call

STDIO-based communication ensures tight integration with LangGraph’s tool invocation model without opening HTTP ports or exposing REST endpoints.

📜 Example Requests

Discover Tools

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/discover"
}

Run Show Command

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "run_show_command",
    "arguments": {
      "device_name": "router1",
      "command": "show ip interface brief"
    }
  }
}

🔒 Security Features

Input validation using Pydantic

Blocks unsafe commands like erase, reload, write

Prevents pipe/redirect abuse (e.g., | include, >, copy, etc.)

Gracefully handles parsing fallbacks and errors

📁 Project Structure

.
├── pyats_mcp_server.py     # MCP server with JSON-RPC and pyATS integration
├── Dockerfile              # Docker container definition
├── testbed.yaml            # pyATS testbed (user-provided)
└── README.md               # This file

✍️ Author

John Capobianco

Product Marketing Evangelist, Selector AI

Author, Automate Your Network

Let me know if you’d like to add:

A sample LangGraph graph config

Companion client script

CI/CD integration (e.g., GitHub Actions)

Happy to help!

The testbed.yaml file works with the Cisco DevNet Cisco Modeling Labs (CML) Sandbox!