Everything your agent needs to debug RabbitMQ.
Qarote detects incidents, explains root cause, and exposes all of it to your AI coding agent over MCP. Here's the whole product — starting with the part nothing else has.
An MCP surface your agent already speaks.
Most RabbitMQ tooling stops at metrics. Qarote ships a deterministic incident engine and managed root-cause analysis behind one MCP surface — so you debug by asking, in the editor you're already in.
{
"mcpServers": {
"qarote": {
"command": "npx",
"args": ["-y", "@qarote/mcp"]
}
}
}A deterministic rules engine, not a guess.
Live, rules-based detection fires the moment broker state crosses a threshold — dropped consumers, depth anomalies, stalled acks. No baseline period, no dashboards to build. Open source, MIT.
orders.incoming — consumers droppedRoot cause, written for you on demand.
One explain_incident call turns an incident plus the surrounding broker state into a plain-language root cause. Bring your own key, run it managed, or keep it fully local on Ollama. Plan-gated — never deploy-gated.
The 3 consumers disconnected at 15:06; depth climbs ~8k/min with zero drain — likely a crashed worker downstream. Check deploys around 15:05. ↳ co-firing: events.payment also lost consumers.
Catch the anti-patterns before they page you.
A structural scan flags the RabbitMQ mistakes that become 3 a.m. incidents — missing dead-letter exchanges, idle no-consumer queues, default-exchange overuse — each one something your agent can read back to you.
no DLX on orders.*Rejected messages are dropped silently across 5 queues.
idle no-consumer queueemails.outbound has had zero consumers for 9 days.
default-exchange overuse82% of publishes route through the default exchange.
The live broker map, with drill-down.
See exchanges, queues, and bindings as a graph. Click any node to inspect depth, consumers, and bindings — the old object-list "Browse" dissolved into a map you can actually reason about. Charts keep 30 days of metric history.
Route what matters, where it matters.
Findings reach you by email, Slack, or webhook, with per-severity routing so criticals page and the rest digest. Email is free; Slack and webhooks unlock on Developer.
Two minutes to your first finding.
Connect over the Management HTTP API on TLS — no exporter, no agent on the box. One broker or many; RabbitMQ 3.x & 4.x; single-node or clustered; CloudAMQP and Amazon MQ included.
# point Qarote at the Management API host: aws-rabbit.prod.internal port: 15672 tls: on user: qarote-readonly # → 14 queues · 6 exchanges · 3 findings
Your network, your model, your rules.
Run Qarote cloud-hosted, or self-host with Docker, Compose, Dokku, or a standalone binary. Self-hosted validates its license offline — no phone-home — and AI Explain can run on local Ollama, so broker data never leaves your network.
Detection core is MIT and open source. Self-hosted license validation is offline. Local Ollama keeps every byte of broker state inside your perimeter.
See it on your own broker.
Connect in under two minutes, then ask your agent what's wrong.