NetworkOps

AI-Powered Network Automation Platform

Natural language commands to manage multi-vendor infrastructure

NetworkOps is a network automation platform built on the Model Context Protocol (MCP). It enables AI-assisted management of Cisco, Juniper, Nokia, Arista, and Linux devices through natural language — no scripting required.

---

Screenshots

Topology Dashboard

Interactive force-directed network topology with device health indicators, legend, and network stats.

Device Details

Click any device to view its properties, interface status, and remediation controls.

OSPF Adjacencies Overlay

Visualize OSPF adjacencies with area labels directly on topology links.

BGP Sessions Overlay

BGP session status with AS number annotations on peering links.

DMVPN Fabric Status

DMVPN hub-and-spoke overlay with peer state table showing tunnel IPs, NBMA addresses, and uptime.

Ping Sweep with Latency Heatmap

Bulk ping results visualized as a color-coded latency heatmap across the topology.

Command Terminal

Execute show commands on any device with full output display.

AI Documentation Assistant

RAG-powered chat with live network queries — ask questions in natural language and get answers with source citations.

Real-time Telemetry (MDT)

Live CPU, memory, interface traffic, and counters streamed via gRPC at 5-second intervals.

Impact Analysis

Predict the blast radius of a configuration change before executing it.

Intent Drift Engine

YAML-defined expected state validation with per-item compliance checks and violation alerts.

Subnet Calculator

Built-in subnet calculator with VLSM split and quick reference tables.

---

Quick Start

Demo Mode (no network devices needed)

Try NetworkOps without any hardware. Demo mode simulates a multi-device network with routers, switches, and an FRR edge device.

git clone https://github.com/E-Conners-Lab/NetworkOps_Platform.git && cd NetworkOps_Platform
./scripts/quickstart.sh --demo
# Follow the prompts, then open http://localhost:5001 — Login: admin/admin

Or manually:

git clone https://github.com/E-Conners-Lab/NetworkOps_Platform.git && cd NetworkOps_Platform
uv venv --python 3.11                  # downloads Python 3.11 if needed
source .venv/bin/activate
uv pip install -r requirements.txt
cd dashboard && npm install && npm run build && cd ..
JWT_SECRET=$(python3 -c "import secrets; print(secrets.token_hex(32))")
DEMO_MODE=true JWT_SECRET=$JWT_SECRET python dashboard/api_server.py

Don't have uv? Install it with curl -LsSf https://astral.sh/uv/install.sh | sh. You can also use python3.11 -m venv .venv and pip install -r requirements.txt instead.

With Real Devices

git clone https://github.com/E-Conners-Lab/NetworkOps_Platform.git && cd NetworkOps_Platform
./scripts/quickstart.sh
# Edit .env with your device IPs and credentials
# Edit config/devices.py with your device inventory
python dashboard/api_server.py

Docker

git clone https://github.com/E-Conners-Lab/NetworkOps_Platform.git && cd NetworkOps_Platform
./scripts/docker-init.sh
# Edit .env with your credentials (or add DEMO_MODE=true for demo)
docker compose up

---

Demo Mode

Demo mode provides a fully functional dashboard experience using simulated network data. No real devices, SSH connections, or lab infrastructure required.

What Works in Demo Mode

Feature	Status	Details
Topology visualization	Full	7-node network with routers, switches, and FRR edge device
BGP overlay	Full	Simulated iBGP (AS 65000) and eBGP (AS 65100) peering
OSPF overlay	Full	Area 0 adjacencies across all routers
DMVPN overlay	Full	Hub-and-spoke topology with 3 spokes
Switch fabric overlay	Full	2 switches with uplink status and port counts
Ping sweep	Full	Simulated 100% reachability with realistic latencies
Interface list	Full	Per-device interface display with status
Show commands	Partial	show ip interface brief, show version, show ip route return simulated output; other commands return a placeholder message
Login / RBAC	Full	JWT auth, role-based permissions, all dashboard features unlocked
Health endpoints	Full	/healthz and /health/detailed work without Redis

What Does Not Work in Demo Mode

Feature	Reason
Configuration changes	No real devices to push config to
NETCONF/YANG queries	Requires live device connections
pyATS baselines/diffs	Requires SSH to real devices
RAG chatbot	Requires ANTHROPIC_API_KEY and ChromaDB index
Streaming telemetry	Requires gRPC streams from real devices
Terminal (arbitrary commands)	Most commands return placeholder output
NetBox integration	Requires a running NetBox instance
MCP tools via Claude	Tools that SSH to devices will fail; read-only tools work

Demo Topology

        R1 ──── R2
        │        │
   Switch-R1  Switch-R2
        │
        R3 ──── R4
        │
      edge1 (FRR, AS 65100)

• R1–R4: Cisco C8000V routers (IOS-XE) — OSPF Area 0, iBGP AS 65000, DMVPN hub/spoke
• Switch-R1, Switch-R2: Cisco Cat9kv switches
• edge1: FRRouting device — eBGP peering with R3

---

Features

React Dashboard

• Interactive force-directed topology with drag-and-drop positioning
• Real-time device health monitoring (healthy / degraded / critical)
• Role-based access control (Admin / Operator / Viewer)
• Command terminal with output display
• AI chat panel (RAG-powered, requires Anthropic API key)
• Device detail panels:
  • Cisco devices: Interface list with admin/line protocol status, remediation buttons
  • Linux hosts: Uptime, memory, disk, gateway reachability
  • Containerlab devices: Container status, uptime, memory
• Visualization overlays:
  • BGP session status with AS numbers
  • OSPF adjacencies with area mapping
  • DMVPN hub/spoke fabric
  • Switch fabric with EIGRP status
  • Ping sweep with latency heatmap
• Intent Drift Engine:
  • YAML-defined expected state validation
  • Force-directed dependency graph
  • Forward/backward/blast-radius impact analysis
  • Cross-subsystem event correlation
• Hierarchical site view (Region → Site → Rack → Device)
• Position persistence via localStorage
• Change management workflow

MCP Tools (178 tools, 24 categories)

Integrate with Claude Code or Claude Desktop for natural language network management.

Health & Monitoring

Tool	Description
health_check	Check single device connectivity
health_check_all	Parallel health check across all devices
full_network_test	Validate OSPF, BGP, DMVPN, and reachability
linux_health_check	Linux host status (uptime, memory, disk)
get_cpu_memory	CPU and memory utilization
ping_sweep	Bulk ping from a device to all targets

Device Operations

Tool	Description
get_devices	List available devices
send_command	Execute show commands
send_config	Send configuration commands
discover_topology	Discover topology via CDP/LLDP
get_interface_status	Detailed interface status
remediate_interface	Fix interfaces (no shutdown, bounce)
bulk_command	Execute command on multiple devices in parallel

Configuration Management

Tool	Description
backup_config	Save running config to timestamped file
compare_configs	Diff two configs or backup files
list_backups	List saved configuration backups

Routing & Troubleshooting

Tool	Description
get_routing_table	Parsed routing table with protocol filtering
traceroute	Hop-by-hop path analysis
get_arp_table	Parsed ARP table with VRF support
get_mac_table	MAC address table from switches
get_neighbors	Structured CDP/LLDP neighbor data
get_logs	Syslog buffer with severity filtering

NETCONF

Tool	Description
get_interfaces_netconf	Interface info via NETCONF/YANG
get_netconf_capabilities	Device NETCONF capabilities
get_bgp_neighbors_netconf	BGP neighbors via NETCONF

pyATS Integration

Tool	Description
pyats_learn_feature	Learn device feature state
pyats_snapshot_state	Capture device state baseline
pyats_diff_state	Compare against baseline
pyats_check_compliance	Check config against golden templates
pyats_list_baselines	List saved baselines
pyats_list_templates	List golden config templates

NetBox Integration

Tool	Description
netbox_get_devices	List devices from NetBox
netbox_get_interfaces	Device interfaces from NetBox
netbox_get_prefixes	IP prefixes from NetBox
netbox_generate_configs	Generate FRR configs from NetBox via Jinja2
netbox_generate_iosxe_config	Generate IOS-XE configs from NetBox
netbox_allocate_ip	Allocate next available IP from a prefix

Memory & Context System

Tool	Description
memory_search	Semantic search across conversation history
memory_save	Save notes for later retrieval
memory_recall_device	Recent events for a device
memory_stats	Memory system statistics

Event Logging & Cache

Tool	Description
get_event_log	Audit log events (filterable)
clear_event_log	Clear audit log
cache_status	Redis cache statistics

Streaming Telemetry (MDT)

• Real-time CPU, memory, and interface counters via gRPC (5-second intervals)
• WebSocket push to dashboard via Socket.IO
• Configured on C8000V routers and Cat9kv switches

RAG Documentation Chatbot

• Indexed CCIE/Cisco documentation with semantic search
• ChromaDB + sentence-transformers embeddings
• Live network queries via SSH during chat
• Claude-powered responses with source citations

Discord Bot (ChatOps)

• Commands: !devices, !health, !bgp, !ping, !events, !topology, !run
• RBAC: Admin commands require "Network Admin" role

Enterprise Features

Category	Feature
Authentication	JWT tokens, MFA/TOTP with recovery codes
Authorization	Role-based access control (Admin/Operator/Viewer)
Caching	Redis-backed response caching
Rate Limiting	Configurable per-endpoint rate limits
Async Jobs	Celery task queue with PostgreSQL persistence
Health Probes	/healthz, /readyz, /health/detailed, /metrics
TLS/HTTPS	Nginx reverse proxy with TLS 1.2/1.3
Logging	Structured JSON logs for aggregation
SIEM Forwarding	Splunk, Elasticsearch, Syslog, Webhook
Kubernetes	Full manifests with HPA, PDB, rolling deployments
Graceful Shutdown	SIGTERM handling with request draining

---

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                        NetworkOps Platform                          │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│   ┌──────────────┐    ┌──────────────┐    ┌────────────────────┐   │
│   │   Frontend   │    │  API Server  │    │    MCP Server      │   │
│   │   (React)    │◄──►│   (Flask)    │    │  (Claude Tools)    │   │
│   │  Port 3000   │    │  Port 5001   │    │                    │   │
│   │              │    │              │    │  178 tools across  │   │
│   │  • Topology  │    │  • REST API  │    │  24 categories     │   │
│   │  • Terminal  │    │  • WebSocket │    │  • NETCONF/YANG    │   │
│   │  • RAG Chat  │    │  • Auth/RBAC │    │  • pyATS           │   │
│   └──────────────┘    └──────┬───────┘    └──────────┬─────────┘   │
│                              │                       │              │
│                 ┌────────────┴───────────────────────┴──────────┐  │
│                 │             Shared Data Layer                  │  │
│                 │  • networkops.db (Alembic-managed)             │  │
│                 │  • chromadb/ (RAG)    • baselines/ (pyATS)     │  │
│                 └───────────────────────┬───────────────────────┘  │
│                                         │                          │
├─────────────────────────────────────────┼──────────────────────────┤
│                                         │                          │
│              Network Devices (SSH / NETCONF)                       │
│   ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌───────┐ │
│   │ Cisco   │  │ Juniper │  │ Nokia   │  │ Arista  │  │ Linux │ │
│   │ IOS-XE  │  │ Junos   │  │ SR Linux│  │  EOS    │  │ FRR   │ │
│   └─────────┘  └─────────┘  └─────────┘  └─────────┘  └───────┘ │
└─────────────────────────────────────────────────────────────────────┘

---

Setting Up Your Own Lab

Use the quickstart script to configure NetworkOps for your own network devices:

# Interactive setup (recommended)
./scripts/quickstart.sh

# Demo mode (no devices needed)
./scripts/quickstart.sh --demo

# Non-interactive (CI/CD, Docker)
./scripts/quickstart.sh --headless

# Setup and auto-start
./scripts/quickstart.sh --headless --auto-start

The script detects uv automatically for faster dependency installation.

Deploying on a Linux VM (Proxmox, bare metal, etc.)

You can run the entire platform and containerlab on a single Linux host — no Multipass needed.

1. Install prerequisites:

# Docker
curl -fsSL https://get.docker.com | sh
sudo usermod -aG docker $USER && newgrp docker

# Containerlab
sudo bash -c "$(curl -sL https://get.containerlab.dev)"

# uv (Python version manager + package installer)
curl -LsSf https://astral.sh/uv/install.sh | sh
source ~/.bashrc

# Node.js (for building the React frontend)
curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash -
sudo apt install -y nodejs

2. Clone and set up the project:

git clone https://github.com/E-Conners-Lab/NetworkOps_Platform.git && cd NetworkOps_Platform
uv venv --python 3.11
source .venv/bin/activate
uv pip install -r requirements.txt

3. Configure .env:

cp .env.example .env

Set these values in .env:

CONTAINERLAB_LOCAL=true
CONTAINERLAB_ONLY=true
CONTAINERLAB_TOPOLOGY_PATH=/path/to/NetworkOps_Platform/containerlab/datacenter.clab.yml
DEMO_MODE=false
JWT_SECRET=<generate with: python -c "import secrets; print(secrets.token_hex(32))">

4. Pull images and deploy the lab:

docker pull frrouting/frr:v8.4.1
docker pull ghcr.io/nokia/srlinux:24.10.1
docker pull alpine:3.19

cd containerlab
sudo containerlab deploy -t datacenter.clab.yml
cd ..

5. Build the frontend, run migrations, and start:

cd dashboard && npm install && npm run build && cd ..
alembic upgrade head
python dashboard/api_server.py

6. Change the default password:

TOKEN=$(curl -s http://localhost:5001/api/auth/login -H "Content-Type: application/json" -d '{"username":"admin","password":"admin"}' | python3 -c "import sys,json;print(json.load(sys.stdin)['token'])")

curl -X POST http://localhost:5001/api/auth/change-password -H "Content-Type: application/json" -H "Authorization: Bearer $TOKEN" -d '{"old_password":"admin","new_password":"YourNewPassword"}'

7. Access the dashboard at http://<your-vm-ip>:5001 and log in with admin and your new password.

Note: CONTAINERLAB_LOCAL=true runs Docker commands directly instead of routing through Multipass. CONTAINERLAB_ONLY=true hides EVE-NG/physical devices that aren't present, showing only containerlab devices in the dashboard.

Containerlab via Multipass (macOS)

If you're on macOS and prefer to use Multipass to host the containerlab VM, leave CONTAINERLAB_LOCAL=false (the default) and set CONTAINERLAB_VM to your VM name. See core/containerlab.py for details.

Supported Device Types

Platform	device_type	Features
Cisco IOS-XE	cisco_xe	SSH, NETCONF, pyATS
Cisco IOS	cisco_ios	SSH only
Linux	linux	SSH
FRRouting	containerlab_frr	SSH (via Containerlab)
Nokia SR Linux	containerlab_srlinux	SSH (via Containerlab)

Example Device Entry

DEVICES = {
    "core-router": {
        "host": "192.168.1.10",
        "device_type": "cisco_xe",
        "username": "admin",
        "password": "secret",
        "netconf_port": 830,
    },
    "linux-server": {
        "host": "192.168.1.100",
        "device_type": "linux",
        "username": "root",
        "password": "password",
    },
}

---

Example Claude Commands

"Show me the interfaces on R1"
"What's the OSPF neighbor status across all routers?"
"Run a health check on all devices"
"Take a baseline snapshot of R2 and check for drift"
"Ping from R1 to all router loopbacks"
"Compare the running config of R1 and R2"
"Show me the PIM neighbors and RP mapping"

---

Project Structure

.
├── network_mcp_async.py     # MCP server entry point
├── mcp_tools/               # 178 MCP tools (24 categories)
├── config/
│   └── devices.py           # Device inventory (static or NetBox)
├── core/                    # Core engines
│   ├── intent_engine.py     # Intent-based validation
│   ├── impact_analyzer.py   # Impact analysis
│   ├── demo/                # Demo mode fixtures
│   └── unified_db.py        # Unified SQLite database
├── dashboard/               # React dashboard + Flask API
│   ├── api_server.py        # Flask API backend (port 5001)
│   ├── routes/              # Modular route blueprints
│   ├── auth/                # JWT, MFA/TOTP, RBAC
│   ├── src/                 # React frontend
│   ├── Dockerfile.api       # API container
│   └── Dockerfile.frontend  # React + nginx container
├── templates/               # Jinja2 config templates (FRR, IOS-XE)
├── alembic/                 # Database migrations
├── ansible/                 # Playbooks and inventory
├── containerlab/            # Topology and device configs
├── docker/                  # Production Dockerfiles
├── helm/                    # Helm chart for Kubernetes
├── k8s/                     # Kubernetes manifests
├── nginx/                   # TLS reverse proxy config
├── netbox/                  # NetBox integration
├── scripts/                 # Quickstart, benchmarks, utilities
├── services/                # Discord bot, background services
├── tests/                   # Test suite
└── docs/                    # Documentation

---

Requirements

• Python 3.11+ (recommended; pyATS doesn't support 3.14 yet)
• Node.js 18+ (for dashboard)
• Network devices accessible via SSH (for non-demo usage)
• Claude Code or Claude Desktop with MCP support (for AI tools)

Optional

• Redis (caching and rate limiting)
• PostgreSQL (job persistence)
• Anthropic API key (RAG chatbot)
• NetBox instance (IPAM integration)

---

Using a Local LLM

The RAG chatbot uses Claude by default, but you can swap in a local LLM if you prefer to keep everything on-prem.

What already runs locally

• Embeddings — sentence-transformers (all-MiniLM-L6-v2) runs locally for all vector search and document indexing. No external API calls.
• MCP tools — The 178 MCP tools use the standard MCP protocol over SSH/NETCONF. Any MCP-compatible client can consume them, and the HTTP proxy (mcp_http_proxy.py) exposes them as REST endpoints for any LLM or script.

What requires changes for a local LLM

The RAG chat panel (rag/query.py) is wired to the Anthropic SDK. To point it at a local model (Ollama, vLLM, LM Studio, llama.cpp, etc.), you need to modify three files:

1. rag/query.py — Replace the Anthropic client with an OpenAI-compatible client:

# Before (Anthropic)
import anthropic
client = anthropic.Anthropic(api_key=api_key)
response = client.messages.create(
    model=model,
    max_tokens=1024,
    system=self.SYSTEM_PROMPT,
    messages=[{"role": "user", "content": user_message}]
)

# After (OpenAI-compatible — works with Ollama, vLLM, LM Studio)
from openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1", api_key="unused")
response = client.chat.completions.create(
    model="llama model here",
    max_tokens=1024,
    messages=[
        {"role": "system", "content": self.SYSTEM_PROMPT},
        {"role": "user", "content": user_message}
    ]
)

2. rag/sanitizer.py — Update the model allowlist (lines 15–22) to include your local model names, or remove the allowlist check entirely.

3. dashboard/routes/chat.py — Change the default model from claude-sonnet-4-20250514 to your local model name.

Tool-calling (live device queries)

The RAG chat can execute live SSH commands against your devices mid-conversation (e.g., "run a health check"). This uses Anthropic's tool_use block format. If you want this working with a local LLM:

• The tool-calling loop in rag/query.py (around line 406) needs to be adapted to the OpenAI function-calling schema
• You'll need a model with solid function-calling support — Llama 3.1 70B+, Qwen 2.5 72B, or Mistral Large work well
• Smaller models (7B/8B) will struggle with multi-step tool chains

If you only need RAG Q&A (ask questions about indexed documentation) without live device queries, you can skip the tool-calling refactor entirely. Any local model that handles basic chat completions will work.

Recommended local setups

Setup	Model	VRAM Required	Tool-calling
Ollama	llama3.1:70b	~40 GB	Yes
Ollama	qwen2.5:32b	~20 GB	Yes
Ollama	llama3.1:8b	~5 GB	RAG only (no tools)
vLLM	Any HuggingFace model	Varies	Yes (with OpenAI server)
LM Studio	Any GGUF model	Varies	Yes (OpenAI-compatible API)

---

Performance

Tested on MacBook Pro M4, Gunicorn with 4 workers:

Concurrency	Requests/sec	Avg Latency
100	5,376	24ms
500	5,100+	slight increase
1,000	4,800+	near saturation

Zero errors at all concurrency levels. See scripts/wrk_benchmark.sh for full benchmark suite.

---

Security Notes

• Never commit .env — it contains credentials and secrets
• Generate a unique JWT_SECRET for every deployment
• The dashboard uses RBAC to limit command access by role
• Dangerous commands (reload, write erase, etc.) are blocked
• Review commands before execution in production

---

License

MIT License