OllamaWatch — Catch Ollama Crashes Before Your Users Do (Telegram Alerts)

Stop babysitting your Ollama server. Get a Telegram alert the moment a model crashes, the GPU runs out of memory, or the API stops responding.

Built for homelabbers and self-hosters running Ollama 24/7 who don't want to pay for SaaS monitoring or refresh dashboards manually.

Features:

• DOWN alerts (🔴): instant notification when the API is unreachable, hung, or returning errors
• DEGRADED alerts (⚠️): CUDA OOM, runner crashes, log errors, GPU memory above 95%
• RECOVERED alerts (✅): service is back, includes latency and currently loaded models
• Smart alerting: fires only on state transitions, no spam
• Built-in fix hints: every alert includes a suggested remediation
• NVIDIA GPU monitoring via nvidia-smi (gracefully skipped on AMD/Intel/CPU-only setups)
• Cross-platform: Linux, macOS, Windows (auto-detects log paths)
• Single-file Python script: no Docker, no Node.js, no SaaS subscription
• MIT licensed

Install: pip install -r requirements.txt python3 ollamawatch.py check # one-shot health check python3 ollamawatch.py watch # continuous monitoring

Requires: Python 3.8+, Ollama (local or remote), free Telegram bot (5-min setup).

Use cases:

• Monitor a remote Ollama box you cannot see (headless homelab, VPS)
• Get woken up if an overnight batch job hangs the GPU
• Catch memory leaks before they take down your inference stack
• Detect when a model pull has stalled mid-download
• Replace expensive SaaS monitoring ($0/month vs $30+/month)