OpenAlerts - Monitoring layer for Agentic frameworks

#opensource #github #programming #ai

If you're building with agentic frameworks, you've probably been through this.

Everything looks fine. Your agent is running. Users are chatting. Then someone messages you: "Hey, your bot hasn't replied in 20 minutes."

You check the logs. Your LLM provider started throwing 401s an hour ago. Or a tool call failed mid-session and your agent just started making things up. Or your gateway stopped working and no one noticed.

No alert. No error. Nothing.

The Problem

Tools like Datadog, PagerDuty, and Grafana are great for monitoring servers and infrastructure. But they don't catch agent-specific issues - things like expired API keys, sessions stuck in loops, tool calls failing quietly while the agent hallucinates, or message queues backing up with no warning.

These aren't server problems. They're agent problems. And without something watching for them, you'll only find out when your users complain.

What We Built

I'm Dev, co-founder of Steadwing. We kept running into silent failures with our own agents in production. We looked for a simple alerting tool made for agent frameworks and couldn't find one.

So we built OpenAlerts - an open-source alerting layer that watches your agents in real-time and tells you when something breaks.

Just install it, connect a messaging channel, and you're good to go. No complicated setup.

Here's what it does:

8 real-time alert rules - covers LLM errors, infra failures, gateway downtime, stuck sessions, high error rates, queue buildup, tool failures, and heartbeat issues. You can adjust all thresholds and cooldowns.
Multi-channel alerts - get notified on Telegram, Discord, Slack, WhatsApp, or Signal.
LLM-enriched alerts - optionally uses your model to add a plain-English summary and a suggested fix to each alert. If the LLM call fails, it just sends the regular alert.
Built-in dashboard - a live view of events, logs, and system health. Runs locally, nothing extra to set up.
Zero-token chat commands - /health, /alerts, and /dashboard work in any connected channel and don't use any LLM tokens.
Framework support - works with OpenClaw, OpenManus, and Nanobot today, with more coming.