Detect. Diagnose. Remedy. Verify.
It fixes problems inside the rules you set, records every action, and runs on your own AI model.
How it works
Five steps, re-run every cycle. Closed-loop is the default, not a roadmap milestone.
Install
One Helm command, on any Kubernetes 1.27 or newer.
Detect
Checks for 21 kinds of Kubernetes problems, plus 10 kinds each for AWS, GCP and Azure.
Remediate
Five fixers, each with strict limits. Fixes suggested by AI need a person to approve them.
Report
Sends findings to Slack, Alertmanager, Jira and more. You can also read them with kubectl.
Verify
Checks again after every fix. No “it probably worked”. It confirms.
Try it offline
Run srenix diagnose --snapshot ./export/ on a saved copy of your cluster state. Nothing to install and no access needed. About 30 seconds.
Run it live
A Helm chart installs a watcher and scheduled checks. It reacts within seconds of a change.
Very small
A 13 MB image that runs without admin rights. It uses under 100m CPU and 100 MB memory, and accepts no inbound traffic.
It reads your stack, not a copy of it
It runs 21 checks on Kubernetes and 10 groups of checks each on AWS, GCP and Azure. Then 21 analyzers look for drift, log errors and workload problems. This step only reads. It never changes anything.
One failed check does not raise an alert. Only a second failure in a row does. This cuts false alarms.
A hypothesis, checked against evidence
For every critical finding, SreNix looks deeper: DNS, web responses, certificates and recent events. Then it writes a one-line cause.
Bounded fixers, never an open shell
Each fixer does one kind of thing, in one place, up to a set limit. If a problem is not on your approved list, SreNix does not guess. It gives a person the full diagnosis and a fix to approve.
“Resolved” is a measurement
SreNix runs the same check that found the problem and confirms it is gone. If not, it records the fix as failed. It never assumes a fix worked.
You can read every guardrail
Read all of it before you install anything.
- Predictable core: no AI model ever touches your cluster
- More freedom only for fixes that keep working, with a safety switch
- A signed record you can export with one command
- No data leaves your network
# fixers/k8s-limit-raise.policy.yaml
scope:
namespaces: [payments, checkout]
resources: [deployment]
bounds:
memory_multiplier_max: 2.0
requires_verification: true
rollback_on_regression: true
approval:
in_envelope: auto
out_of_envelope: signed_url
signer_roles: [sre-oncall]
identity: oidc://srenix-fixer
Lessons are shared. Your data is not.
Each customer’s data stays with that customer. But when the same type of failure keeps appearing, it becomes a new built-in fix. Everyone gets it in the next release.
Bring your own LLM
AI features are optional. Use OpenAI, Anthropic, or a model you host. AI-suggested fixes need a person’s approval. Automatic approval happens only when every safety check passes.
On-call should be quieter every week
Helm install in 5 minutes. No telemetry exfiltration. No per-investigation surprises.