ForgedWatchdog
A filesystem watchdog built for AI agent environments. Monitors write activity, classifies risk, enforces path boundaries, and auto-pauses runaway processes — without being tied to any specific framework or toolchain.
The Problem
AI agents write files. That's fine — until they don't stop. Runaway writes, out-of-bounds file creation, and cascading side effects are real failure modes in multi-agent systems. Most monitoring tools aren't built for this.
ForgedWatchdog was built out of a real need: monitoring a local AI infrastructure stack without being locked into a specific framework. It watches the filesystem, classifies what it sees, and reacts — automatically.
Features
Real-Time Detection
Watches configured paths with sub-second latency. Create, modify, delete, and move events are captured and queued for classification.
Smart Classification
Events are scored against a risk model — routine writes score low, out-of-bounds operations score high. Configurable thresholds, no hardcoded rules.
Auto-Pause
When high-risk activity is detected, ForgedWatchdog can pause the responsible process, fire a webhook, or write a kill file — your choice.
Boundary Enforcement
Define read-only zones, output-only zones, and fully restricted paths. The watchdog enforces them and logs every violation.
Structured Audit Logs
Every event written to JSONL. Timestamp, path, event type, risk score, agent ID (if configured), and action taken. Grep-friendly.
Works Outside Any Framework
Designed to be generic — not tied to any specific AI framework, orchestration system, or toolchain. Point it at paths and go.
Quick Start
pip install forgedwatchdog# watchdog.yml
watch_paths:
- ~/ai-workspace
- /var/agents/outputs
readonly_paths:
- ~/ai-workspace/config
alert:
webhook: https://your-endpoint/alert
kill_file: /tmp/.agent-pauseforgedwatchdog --config watchdog.ymlTech Stack
Development Status
Read the Blog Post
How ForgedWatchdog was designed — the problem, the architecture, the decisions.
Read the Story →Need Something Custom?
ForgedWatchdog is a starting point. If you need tailored monitoring for your systems, let's talk.
Work With Me →