Stop babysitting dashboards. Ask your cluster.
Gregor connects your AI agent to Apache Kafka and returns answers and diagnoses in plain language — grounded in real signals, explainable, and read-only by design. No new dashboard to learn.
No card. Self-host or managed cloud. We email you when your slot opens.
consumer-group orders-billing lag is climbing on payments. why, and is it the brokers?
finding Partition payments-3 lag 1,243,902 and rising at ~8.4k msg/s.
cause One consumer left during an incremental rebalance (KIP-848). payments-3 reassigned to pod/orders-7, processing 4× slower (p99 1,840ms vs 460ms).
verdict Not a broker issue. ISR healthy, no under-replicated partitions. Throughput-bound consumer.
Representative output. Every root cause is synthesized from real signals — never invented.
/ the problem
Another dashboard won't tell you what's wrong.
Control Center, Conduktor, AKHQ, Redpanda Console, Grafana stacks — they render the metrics beautifully and leave the hard part to you. Diagnosing lag spikes, under-replicated partitions and rebalances still means staring at graphs and leaning on tribal knowledge that lives in one engineer's head.
If you run Kafka without a dedicated SRE or observability team, that engineer is you — usually at 2 a.m.
dashboard-centric tools
You ask the question. You read ten panels. You correlate them in your head. You guess.
tribal knowledge
"Check ISR, then the rebalance logs, then the slow consumer" — undocumented, unowned, unscalable.
heavy footprint
Capture infra, agents, sidecars, retention — another system you now have to operate.
gregor
You ask in plain language. You get a root cause, the evidence, and a verdict.
/ how it works
From zero to grounded answers.
Stand Gregor up next to your cluster, connect the agent you already use, and start asking — or let us run the core for you.
- 1
Drop in the agent
Deploy one lightweight agent next to your Kafka cluster — Docker Compose, a single binary, or Helm — and give it a read-only Kafka principal. It connects outbound only, so there are no inbound ports to open. Self-hosted, your data never leaves your network.
- 2
Connect your AI agent
Point the agent you already use — Claude, Cursor, or an internal agent — at Gregor's MCP (Model Context Protocol) server. One config block, and there's no new dashboard to learn.
- 3
Ask in plain language
Ask the way you'd ask a teammate. Your agent calls Gregor's read-only tools and answers — grounded in your cluster's real, live state.
› why is the orders consumer group lagging?
› are any partitions under-replicated?
// prefer managed?
Use Gregor Cloud
Don't want to operate the core yourself? Deploy only the lightweight collector in your network — still read-only, still outbound-only — and we host, run and upgrade everything else. Point your agent at your hosted Gregor endpoint and ask away. Same workflow, nothing to run or upgrade beyond the collector.
Transparency In managed mode, metrics and cluster metadata are sent to Gregor Cloud over an encrypted, outbound-only connection. To keep all data inside your network, choose self-hosted.
/ why you can trust it
Grounded, not guessing.
A deterministic engine detects incidents and the signals behind them; the LLM synthesizes the root cause from that real evidence — it never invents one. Detection is explainable, and the model only ever reasons over signals Gregor actually observed.
Read the facts
Consumer lag, throughput, topology and ISR health, config anti-patterns — pulled live and continuously over a read-only principal.
Surface the incident
Deterministic rules flag the incident and the signals behind it, continuously — consumer-lag, under-replication, offline partitions, rebalance fault. Reproducible, not a model's opinion.
Name the root cause
The LLM synthesizes the root cause from that evidence and explains it — grounded in what Gregor observed, never fabricated.
/ capabilities
Four things to do with your cluster, in plain language.
Root cause, not just symptoms
The deterministic Incident Diagnosis Engine finds the cause of consumer-lag, detects under-replicated and offline partitions, and analyses rebalances — KIP-848 aware.
The whole picture, on demand
Consumer lag per group, topic and partition. Throughput, cluster topology, ISR health and historical trends — ask for exactly the slice you need.
Read the log directly
On-demand message inspection and replay straight from the log. No capture pipeline, no sidecars, no extra retention to operate.
Agent-native, no dashboard
An MCP server your agent talks to. Ask natural-language questions from the tools you already work in — there's no console to babysit.
/ security
Secure and sovereign by design.
Gregor is built for teams that won't hand their cluster to a black box. It connects with the least privilege it can, never writes, and — self-hosted — keeps every byte inside your network.
- read-only principal
- It can describe and consume. It cannot produce, alter or delete. Full stop.
- least-privilege ACLs
- Scoped to exactly the topics and operations it needs — nothing more.
- outbound-only agent
- No inbound ports to open. The agent dials out; nothing dials in.
- TLS / SASL
- Encrypted and authenticated against your brokers, the way you already run them.
# least-privilege, read-only
kafka-acls --add \ --allow-principal User:gregor \ --operation Read \ --operation Describe \ --topic '*' \ --group '*'
/ deploy your way
Run it wherever you already run things.
A single lightweight Go agent plus the core service — Kafka 4.x / KRaft native, no ZooKeeper assumptions. Self-host it and your data never leaves your network, or let us run the managed cloud.
// self-hosted, three ways
Try it in one command
Postgres included. Point it at your cluster with read-only credentials and you're observing.
Bare VMs, no runtime
One static Go binary for air-gapped environments. Bring your own Postgres.
Native to your cluster
One chart for Kubernetes — runs happily alongside Strimzi and other operators.
We run it for you — without holding your data.
Offload the service to us. Your collector stays inside your network and is outbound-only — nothing dials in, and your messages never leave.
/ works with your agents
Speaks MCP. Bring your own agent.
Gregor exposes a Model Context Protocol server, so it plugs into the agents your team already uses — no bespoke integration, no lock-in. Point any MCP client at your endpoint and start asking.
mcp client
{
"mcpServers": {
"gregor": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://gregor.your-net.internal/mcp"]
}
}
} Toggle the endpoint above · self-hosted stays inside your network, managed cloud is a hosted URL · stdio clients use mcp-remote
/ why now
Three things changed at once.
- 01
Kafka 4.0 dropped ZooKeeper
KRaft-only changes the operational model. New tooling should be born native to it — not retrofitted.
- 02
KIP-848 changed rebalances
The new consumer-group protocol shifts how lag and reassignment behave. The diagnosis rules have to know it.
- 03
Agents and MCP matured
Capable agents and a real interop protocol make agent-native ops practical — not a demo.
/ faq
Straight answers.
What is Gregor? + ×
Gregor is an AI-agent-native observability and incident-diagnosis tool for Apache Kafka. You connect the AI agent you already use to your cluster over the Model Context Protocol (MCP) and ask questions in plain language — instead of reading another dashboard.
How is Gregor different from Control Center, Conduktor, AKHQ, Redpanda Console or Grafana? + ×
Those tools are dashboard-centric: they render the metrics and leave the diagnosis to you. Gregor is agent-native — it surfaces the incident and the signals behind it, then answers in plain language, so you don't have to correlate ten panels and lean on tribal knowledge.
Does the LLM guess the diagnosis? + ×
No. A deterministic engine detects incidents and the signals behind them; the LLM only synthesizes the explanation from that real evidence — it reasons over facts Gregor actually observed and never invents a cause.
What access does Gregor need? Is my data safe? + ×
Gregor connects with a read-only Kafka principal and least-privilege ACLs. The agent is outbound-only — no inbound ports to open — and traffic is encrypted over TLS/SASL. It can describe and consume; it cannot produce, alter or delete.
Self-hosted or managed — where does my data live? + ×
Self-hosted keeps all data inside your network. With Gregor Cloud you deploy only the lightweight collector — still read-only, still outbound-only — and metrics and cluster metadata are sent to Gregor Cloud over an encrypted, outbound-only connection. Choose self-hosted to keep everything in-network.
Which AI agents and Kafka versions are supported? + ×
Gregor speaks MCP, so it works with Claude, Cursor, Cline, GitHub Copilot, Codex, Windsurf, opencode, and any MCP client. It's built for Kafka 4.x / KRaft — no ZooKeeper assumptions.
Is Gregor available now? + ×
Gregor is in early access. Join the waitlist and we'll reach out when a slot opens — self-hosted or managed, your call.
Get your cluster on the waitlist.
Gregor is in early access. Join the waitlist and we'll reach out when a slot opens — self-hosted or managed, your call.
One email when you're in. No spam, no sharing.