🛠️ RADKit NetOps Agent

A multi-tool network housekeeping AI agent use case for your Cisco RADKit service.

---

📘 About

This repository provides a reference implementation for running MCP servers and using them as tools for an AI agent specialized in network operations.

Included components:

Component	Purpose
🌐 RADKit MCP server	Inventory access and network operations
🧩 GitHub MCP server	Repository and issue management workflows
🗺️ DrawIO MCP server	Diagram authoring workflows

These three MCP servers are implemented as services in Docker Compose.

This project includes instructions on how to set up the environment, run containers that serve these MCP servers, and onboard them into the open-source LibreChat platform to create an AI agent with a specific persona.

⚙️ The official RADKit MCP server

RADKit includes an MCP server that can connect to your RADKit service and use the Client API to enable AI-driven operations across your device inventory.

This MCP server exposes a single tool: radkit_eval. Its purpose is to generate and execute Python code based on the Client API, which optimizes how your agent interacts with your RADKit service. Instead of exposing one tool per activity (retrieve inventory, execute commands, and so on), the MCP server can generate a Python snippet that dynamically uses Client API features such as filters and many-to-many command execution across devices.

Testing your RADKit MCP server

If you want to deploy the AI Agent right away, skip to the Running the AI Agent section.

To deploy and test your RADKit MCP server stand-alone (no Docker), follow these steps:

Ensure the following are installed on your host system:

• Python 3.12+
• uv

From repository root:

cd radkit-mcp

Create and activate a virtual environment:

uv venv -p 3.13
source .venv/bin/activate

Install the dependencies:

uv sync

Create your local service config from the example:

cp service.config.direct.example service.config.direct

Then edit your service.config.direct file with your RADKit connection values:

{
  "type": "direct",
  "username": "your_user@your_domain.com",
  "password_base64": "your_E2EE_password_from_the_webui_in_base64",
  "host": "your_radkit_host",
  "rpc_port": your_radkit_rpc_port
}

The parameter type supports four connection types:

type	method	Auth mechanism	Notes
direct	—	username + password (base64)	Connects directly to a local RADKit service (host:rpc_port)
cloud	api_key	Encrypted access token file	Token stored in access_token_path; passphrase via prompt or keyring
cloud	certificate	Client certificate	Certificate must already be provisioned
cloud	sso	Single Sign-On (browser)	Browser-based SSO flow

For this demo, we use direct.

⚠️ When choosing direct, your RADKit service needs to be enabled with RPC and have a RPC port assigned. This can be checked by navigating in your service's Web UI to Settings > service.connectivity.port_direct_rpc. Default is 8181

⚠️ The E2EE password for your remote user can be found in the user details in the Web UI: Remote Users > <your user> > E2EE validation token. It is NOT encoded in base64 by default. Encode it before adding it to the file.

Once the file details are populated, start the RADKit MCP server with the following command:

uv run --no-sync radkit-llm mcp-server \
  --port 8082 \
  --sandbox subprocess \
  --service-config service.config.direct \
  --profile radkit-mcp-profile.py \
  --prompt radkit-mcp-prompt.txt \
  --auto-approve \
  --mcp-log-dir logs/

The following are the available options and parameters for this command:

Parameter	Type	Required	Options / Default	Description
--port	integer	No	any port (e.g. 8082)	TCP port the MCP server listens on
--sandbox	choice	Yes	subprocess | bubblewrap	subprocess: simple, cross-platform, no isolation. bubblewrap: Linux-only, network- and filesystem-isolated sandbox.
--service-config	path	No	—	Path to a JSON service connection config. Omit to allow the MCP client to connect to any service interactively (elicitation flow).
--auto-approve	flag	No	off by default	Automatically approve every RADKit interaction. Required for LibreChat: its MCP client cannot complete the elicitation flow.
--profile	file	No	—	Python script to expand how it processes the data before it interacts with the target RADKit service.
--prompt	file	No	—	Text/Markdown file whose content is appended to the built-in LLM system prompt.
--override-prompt	file	No	—	Text/Markdown file that replaces the entire LLM system prompt (overrides --prompt and all built-in content).
--mcp-log-dir	directory	No	logging disabled	Directory where MCP session logs (.log.md files) are written.

After running this command, your MCP server should be available at http://localhost:8082/mcp. To onboard it in LibreChat, go to the Running the AI Agent section.

About the --profile and --prompt MCP server options

The official RADKit MCP server supports capability expansion in both request processing and LLM behavior through profiles and prompts.

A profile is a Python script executed at sandbox startup (as documented by radkit-llm mcp-server --help).

Its purpose is to inject custom Python functions and policy controls into the runtime scope used by radkit_eval.

In this repository, radkit-mcp-profile.py is used to enforce command governance:

• It wraps execute_cli_commands with a prefix-based allow list.
• It enforces behavior based on RADKIT_ACCESS_MODE (show, full, conf, debug).
• It blocks disallowed commands before they are sent to target devices.

A prompt file (--prompt) appends additional instruction text to the MCP server's built-in system prompt.

In this repository, radkit-mcp-prompt.txt provides execution guidance for radkit_eval, including:

• Safety pre-checks (call get_access_mode() before command execution).
• Preferred helper usage patterns (single execute_cli_commands(...) call with full device/command sets).
• Network-specific selection guidance (for example, IOS_XE filtering) and HTTP helper usage for external APIs.
• Output-formatting expectations (clear formatting and summarization for large outputs).

🤖 Running the AI Agent

This workflow runs RADKit MCP, GitHub MCP, and DrawIO MCP concurrently as Docker containers, and then onboards each server into LibreChat.

Step 1: Prepare environment

# From repository root
cp .env.example .env

Update .env with at least the following:

• GITHUB_PAT (required for GitHub MCP)
• RADKit settings if defaults were modified (RADKIT_SERVICE_CONFIG, RADKIT_PROFILE, RADKIT_PROMPT_FILE)

If RADKit files remain at their default repository paths, setting GITHUB_PAT is typically sufficient.

Default ports:

• RADKit MCP: 8082
• GitHub MCP: 8083
• DrawIO MCP: 8084

Step 2: Start all servers

docker compose up -d --build
docker compose ps

Expected result: three running services (radkit-mcp, github-mcp, drawio-mcp) available in the following URLs:

MCP Server	Endpoint
🌐 RADKit MCP	http://localhost:8082/mcp
🧩 GitHub MCP	http://localhost:8083/mcp
🗺️ DrawIO MCP	http://localhost:8084/mcp

Step 3: Onboard MCP Servers in LibreChat

In LibreChat, register each MCP server using Streamable HTTP transport.

⚠️ If LibreChat runs in a container with a different network namespace, use host.docker.internal instead of localhost.

Step 4: Create the AI Agent

In LibreChat, navigate to the left menu and click Agent Builder. Then populate the required information:

In the Instructions section, copy and paste the contents of AGENT.md. This defines the agent persona and guides how it interacts with MCP servers, users, and incoming requests.

In the section MCP server tools, select the servers that you just onboarded:

Step 5: Test!

Click on the New Chat icon and in the drop-down list of Select Model, look for your agent:

🧪 Usage examples

💬 "Please show me the devices in my inventory"

---

💬 "Please create a new loopback interface 300 name OOB address 2.2.2.2/32 on device iosv-l2"

The AI Agent validates all intents against current configurations before proceeding. If there are conflicts, these are highlighted. If not, explicit approval is requested before proceeding to commit.

---

💬 "Please check compliance of all my IOSXE devices against the best practices of this hardening guide: https://sec.cloudapps.cisco.com/security/center/resources/IOS_XE_hardening"

---

💬 "Please create a detailed report with these findings and save it in our repository"

The AI agent uses the GitHub MCP server for committing this report with the findings of the compliance audit.

---

💬 "Please create an issue per device"

This and this issues are generated on GitHub with an action plan to address the findings of the compliance audit.

---

💬 "Please generate a diagram showing the L2 connectivity between the devices in my inventory. Display the interfaces name on it"

The requested diagram is generated using the DrawIO MCP, and by default the AI Agent displays the Mermaid code in the chat, which is compatible with LibreChat, therefore the diagram is rendered as part of the conversation.

---

Made with ☕️ by Poncho Sandoval - Developer Advocate 🥑 @ DevNet - Cisco Systems 🇵🇹