model-watchdog
model-watchdog is a zero-dependency Python tool that monitors an AI agent's health endpoint and automatically rolls back configuration changes if the agent starts failing, then restarts the service. It solves the problem of agent downtime caused by bad config changes, and is for developers running AI agents or services with HTTP health checks and config files.
✨ Key features
- Auto-rollback on repeated health check failures
- Saves last known good config backup
- Restarts service after rollback
- Alerts via Telegram, Slack, Discord, or webhook
- Zero dependencies, single Python file
- Supports JSON and optional YAML config
🎯 Use cases
- Monitor and auto-recover AI agent config changes
- Prevent prolonged downtime from bad model upgrades
- Integrate health checks into CI with one-shot mode
- Run as a systemd service for 24/7 monitoring
📦 Installation
🧰 Requirements: Python 3.8+; no external dependencies required, optional pyyaml for YAML config.
# Clone the repository
git clone https://github.com/feralghost/model-watchdog.git
cd model-watchdog
# No installation needed; run directly with Python 3.8+
🚀 Usage
python3 watchdog.pypython3 watchdog.py --config watchdog.yamlpython3 watchdog.py --check-once❓ FAQ
What happens when the agent fails health checks?
After K failures within M minutes, it rolls back the config to the last known good backup and restarts the service.
How does it decide what is a 'good' config?
When the agent is healthy after a config change, it saves the current config as the 'good backup'.
Can I use it with services other than OpenClaw?
Yes, it works with any AI agent or service that has an HTTP health endpoint and a config file plus restart command.
Does it require any external dependencies?
No, it uses only the Python standard library. Optionally, you can install pyyaml for YAML config support.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.