Ask why it broke. Get the evidence.
Root Cause Console is an open-source AI SRE agent that runs on your own servers. It watches Prometheus, Loki, and Kubernetes, investigates incidents with read-only tools, and checks every answer against your data before you see it.
$sh deploy/setup-env.sh && docker compose up -d --buildWorks with the monitoring you already run
- Prometheus
- Loki
- Kubernetes
- Amazon CloudWatch
- Azure Monitor
- Alertmanager
An on-call engineer that looks at everything and touches nothing.
At 3am, the hard part isn't fixing the problem. It's finding it across dashboards, logs, and pods. Root Cause Console does that search for you: it keeps a live list of what's wrong, and when you ask why, an AI model works through your systems the way an SRE would.
It's an agent with a person in the loop. It can read every source you connect. People decide what to change.
What it does
- Reads metrics, logs, and cluster state every few seconds
- Turns them into a short list of problems: firing alerts, crash loops, out-of-memory kills, failed rollouts
- Investigates with up to 10 read-only queries and writes the root cause with evidence
- Alerts your team on Slack, Teams, Google Chat, or email
What it never does
- Write to your systems: there's no code that can
- Run commands: runbooks are steps for people
- Put action buttons in alert messages
- Read Kubernetes Secrets
See it from the inside
This is the real console, not screenshots. Click around: a database upgrade has just broken payments at Kestrelle, a sample online shop. Open the investigation, check the evidence, or ask your own question and watch the agent work.
Sample data recorded from a real installation. Nothing you do in the demo is saved or sent anywhere.
How an investigation runs
Ask in plain words. The model chooses what to check, reads the results, and decides the next step, and you watch every call as it happens.
Your question goes with the live context
The model starts with what's wrong right now: problems with their diagnosis, firing alerts, down targets, golden signals, and recent deployments.
The console gathers the first evidence
Before the model starts, it runs the obvious queries itself: a crashing pod's error lines, its Kubernetes description, the chart of a signal out of range.
The model investigates
It picks read-only tools, up to 4 per round and 10 in total, and never runs the same call twice.
It answers with a cause and a confidence
A short answer or a full report, with evidence charts and a PDF for the incident review.
The console checks the answer
Every number it quotes is compared with your data, and every query it cites is run again.
Everything an SRE team needs around the agent
Golden signals at a glance
Latency, traffic, errors, and saturation for every service, including Spring, NGINX ingress, JVM, container, database, and queue metrics when they exist.
One root, many symptoms
Problems that start together are grouped with their likely cause, so you fix the database instead of chasing five alerts from the services that depend on it.
Runbook workflows
Step-by-step fixes shown as a pipeline and checked off one at a time. AI drafts need a review. Nothing runs them.
Alerts that arrive with findings
Critical alerts can start an investigation automatically, so the email or Slack message already says what it found.
Projects and PDF reports
Group the sources behind one system and see how often each problem returns, how long it lasts, and what tends to happen after a deployment.
Less alert noise
Mark false alarms and expected behavior once, choose how long that label lasts, and stop being paged for it.
Cost you can see
Tokens, cost, and time for every investigation, a monthly budget, and a token limit per investigation.
Works without internet
Install on servers with no internet access, with a local model through Ollama and your own certificates.
Built for teams
Email invitations, four roles, private investigations you can choose to share, single sign-on with OIDC, security policies, and an exportable audit log.
Your server, your keys
Each installation belongs to one company, with its own database and secrets. Run the model on your own hardware with Ollama and nothing leaves your network.
Every answer is checked before you trust it
AI models are confident even when they're wrong. So after the model answers, the console checks its work against your real systems and tells you what didn't hold up.
Checks catch made-up numbers and failed queries. They can't prove a diagnosis is right, which is why the evidence is always shown next to the answer.
- Cited numbers are compared with your dataA latency of 840 ms in the answer has to be 840 ms in Prometheus.
- Cited queries are run againA query that fails or returns nothing is flagged.
- A competing explanation is requiredThe full report has to say what else could explain it, and why it's less likely.
- “Unknown” is an accepted answerWhen the evidence isn't there, the model is allowed to say so, and you're warned when an answer ignores an active problem.
Connect what you already have
Tools are offered to the model only for what's connected. Logs or a cluster work on their own, without any metric source.
Metrics
- Prometheus
- Amazon Managed Prometheus
- Azure Managed Prometheus
- Google Managed Prometheus
- Amazon CloudWatch
- Azure Monitor
Logs and cluster
- Loki
- Kubernetes
- Alertmanager
- CI/CD webhooks
AI models
- Ollama
- Claude
- Gemini
Alerts
- Slack
- Microsoft Teams
- Google Chat
Sign-in and storage
- OIDC single sign-on
- PostgreSQL
- SQLite
Running in five minutes
You need a server with Docker. The setup script asks a few questions, writes your secrets, and prints your first password.
Want to see it find a real failure? Apply the shop simulation to any cluster and ask why payment is crash-looping.
git clone https://github.com/AymaneK24/SRE-AI-Agent.git
cd SRE-AI-Agent
sh deploy/setup-env.sh # writes .env and prints your password
docker compose up -d --build
docker compose ps # console, postgres, caddy
kubectl apply -f deploy/k8s/reader-rbac.yaml # namespace and read-only account
kubectl apply -f deploy/k8s/postgres.yaml
kubectl apply -f deploy/k8s/console.yaml
kubectl apply -f deploy/k8s/caddy.yaml
kubectl apply -f deploy/k8s/networkpolicy.yaml
kubectl apply -f deploy/examples/shop-simulation.yaml
kubectl -n shop get pods -w
# after 1 to 3 minutes, open Investigate and ask:
# "Why is payment crash-looping?"
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt
RCC_ALLOW_LOOPBACK_TARGETS=true python -m app
ruff check . && pytest
Four roles, clear limits
Investigations are private until you share them. Admins see what everything costs. Owners control sign-in and security policies.
| Viewer | Sees status, saved charts, shared reports, and the question and answer of shared investigations. |
|---|---|
| Engineer | Also investigates, runs queries, edits panels and projects, generates reports, records verdicts, and follows and writes workflows. |
| Admin | Also manages connections, AI settings, alerts, and members, reads all investigations and the audit log, and reviews workflows. |
| Owner | Also manages security policies and single sign-on. |
Open source, in the open
Root Cause Console is free to use, change, and run in production, for any company. Read the code that touches your monitoring, and help shape where it goes next.
Questions
Is it free?
Yes. It's open source under the Apache License 2.0, including for commercial use. The only costs are your server and, if you choose an API model like Claude or Gemini, that provider's usage.
Can it change or break my systems?
No. The console has no code that writes to your systems. The model only gets read-only tools for metrics, logs, and Kubernetes, the Kubernetes account can't read Secrets, and runbooks are instructions for people.
Which AI models can I use?
Ollama on your own server, or Claude or Gemini with an API key. Choose per installation, and set a monthly budget and a token limit per investigation.
Where does my data go?
It stays on your server, in your database. With Ollama, nothing leaves your network. With an API model, the context of each investigation is sent to that provider, and you can see the exact text sent.
Do I need Prometheus?
No. Loki logs or a Kubernetes cluster work on their own. Each source you add gives the model more evidence and more tools.
How accurate is it?
That depends on the model and on what your sources can show. No accuracy score ships with the software. Instead, it includes a benchmark of known failures so you can measure a model against your own incidents before trusting it.
Is the live demo the real product?
Yes. It's the console's own interface, running in your browser against data recorded from a real installation watching a simulated incident. Investigations replay what the real agent did, including its tool calls and the answer checks.