Midnight LLM Monitor

Midnight LLM Monitor is a lightweight hardware intelligence daemon for machines running local LLM servers such as Ollama, llama.cpp, and vLLM-adjacent tooling.

It exposes live metrics over REST and WebSocket so Midnight Coder can make decisions from real hardware state instead of guesses.

Current focus

• GPU and VRAM visibility in the overview
• Ollama running model CPU/GPU split, context usage, and model metadata
• llama.cpp process discovery and model path/context hints
• low-overhead polling with cached expensive collectors
• plugin discovery for third-party collectors

Install

Requires Node.js 20 or newer.

npm install -g midnight-monitor

Run without installing:

npx midnight-monitor

Access

Start the daemon and open the dashboard in a browser:

midnight-monitor start

Then visit:

• http://127.0.0.1:9898/ when the monitor runs on your own machine
• http://<server-ip>:9898/ when the monitor runs on another host

The default HTTP port is 9898. If you change it in midnight-monitor.config.json, use that port instead.

CLI

midnight-monitor
midnight-monitor start
midnight-monitor stop
midnight-monitor status
midnight-monitor doctor
midnight-monitor benchmark

midnight-monitor without arguments prints help. Use midnight-monitor start to start the daemon.

Release

Releases are published from GitHub Actions, not from a local npm publish.

Prerequisites:

• GitHub repository secret NPM_TOKEN must contain an npm automation token with publish access to midnight-monitor.
• Commits merged to main must use Conventional Commits so semantic-release can decide the next version.

Publish flow:

npm run typecheck
npm run lint
npm test
npm run build
git push origin main

The Release workflow runs npx semantic-release, creates the GitHub release, and publishes the package to npm.

Configuration

Create midnight-monitor.config.json in the working directory.

{
  "server": {
    "host": "127.0.0.1",
    "port": 9898,
    "wsPath": "/ws"
  },
  "collectors": {
    "enabled": [
      "cpu",
      "ram",
      "swap",
      "gpu",
      "disk",
      "network",
      "temperatures",
      "processes",
      "ollama",
      "llamacpp",
      "history"
    ],
    "modules": []
  },
  "intervals": {
    "cpu": 1000,
    "ram": 1000,
    "swap": 1000,
    "gpu": 1000,
    "disk": 10000,
    "temperatures": 5000,
    "network": 1000,
    "processes": 1000,
    "ollama": 2000,
    "llamacpp": 2000,
    "history": 1000
  }
}

REST API

• GET / dashboard web
• GET /health
• GET /metrics
• GET /cpu
• GET /memory
• GET /swap
• GET /gpu
• GET /disk
• GET /network
• GET /ollama
• GET /history
• GET /processes

Example:

curl http://127.0.0.1:9898/metrics

WebSocket

Connect to /ws for live updates.

const socket = new WebSocket("ws://127.0.0.1:9898/ws");
socket.onmessage = (event) => {
  console.log(JSON.parse(event.data));
};

Web Monitor

Open http://127.0.0.1:9898/ to see the visual monitor with:

• a GPU/VRAM-first overview for quick capacity checks
• live CPU, RAM, swap, disk, network, and temperature cards
• Ollama running models with CPU %, GPU %, context, context limit, and size
• installed Ollama models with architecture and quantization
• llama.cpp process detection with model path and context hints
• top resource-consuming processes
• history charts for recent pressure and speed changes
• a gear menu for appearance settings
• restore defaults for theme and layout
• draggable and resizable widgets
• warning, critical, and info analysis badges

Example payload

{
  "timestamp": "2026-07-17T12:00:00.000Z",
  "cpu": { "usage": 32, "cores": 8, "threads": 16, "frequencyMhz": 4300 },
  "ram": {
    "usedBytes": 123456789,
    "totalBytes": 17179869184,
    "usagePercent": 41.2
  },
  "swap": { "usedBytes": 0, "totalBytes": 34359738368, "usagePercent": 0 },
  "gpu": {
    "vendor": "AMD",
    "model": "Radeon RX 580",
    "usagePercent": 91,
    "temperatureCelsius": 72,
    "vram": {
      "usedBytes": 7488270336,
      "totalBytes": 8589934592,
      "freeBytes": 1101664256
    }
  },
  "ollama": {
    "running": [
      {
        "name": "midnightcoderagent/MidnightCoder-30B:latest",
        "cpuPercent": 59,
        "gpuPercent": 41,
        "context": 32768,
        "contextLength": 262144,
        "quantization": "IQ2_M",
        "architecture": "qwen3moe"
      }
    ],
    "installed": []
  },
  "llamacpp": { "running": [] },
  "analysis": [
    {
      "severity": "warning",
      "source": "gpu.vram",
      "message": "VRAM almost full. Model may spill into RAM."
    }
  ]
}

JSON Schemas

CpuMetrics

{
  "usage": "number",
  "cores": "number",
  "threads": "number",
  "loadAverage": "[number, number, number]",
  "frequencyMhz": "number",
  "uptimeSeconds": "number"
}

GpuMetrics

{
  "vendor": "string",
  "model": "string",
  "usagePercent": "number | null",
  "temperatureCelsius": "number | null",
  "vram": {
    "totalBytes": "number | null",
    "usedBytes": "number | null",
    "freeBytes": "number | null"
  }
}

Architecture

flowchart LR
  C[Collectors] --> M[Monitor Scheduler]
  M --> H[History Store]
  M --> A[Analysis Engine]
  M --> R[REST API]
  M --> W[WebSocket Stream]
  P[Plugin Packages] --> C

Collector plugins

Collectors are discovered from the built-in src/collectors/ directory and can also be loaded from installed npm packages.

Each collector implements:

initialize(context);
collect(context);
health();
dispose();

Third-party packages can export a createCollector-style factory and be enabled in midnight-monitor.config.json.

Contributing

1. Fork or branch.
2. Run npm install.
3. Use npm run typecheck, npm run lint, and npm test.
4. Keep collectors isolated and failure-tolerant.
5. Include tests for parsing and analysis changes.
6. If a change needs a new dependency, add it to package.json and update the lockfile before opening a PR.