An on-call agent for your own systems
By the time you're paged, the evidence is already in Slack.
Rootsight investigates every alert the moment it fires, read-only: metrics, logs, dependencies, infrastructure, recent deploys and their diff. It names the cause, or hands the incident to a person when it doesn't recognise it, and posts the evidence to Slack. It changes nothing in your systems.
See it work
The same incident, two ways
A real incident recorded on our test stack: a deploy broke the sentiment service. Press Replay to watch who does what.
What changes for the team
An SRE plus Rootsight, against an SRE alone
What does what
Nine parts, one job each
Pick a part to see what it reads, what it does, what it never does, and where you see it.
alert → investigator → novelty guard → explanation → Slack → your review → retraining
Measured, not promised
How well it diagnoses
950 simulated incidents it never trained on: 800 from the 16 known failure types, and 150 from failure types it has never seen, which it should hand to a person.
It starts before you do
The investigation runs when the alert fires, not when someone opens a laptop. The page in Slack arrives with evidence attached.
Attention goes where it's needed
Incidents it recognises come with a cause and a fix to review. Ones it doesn't are handed over with the reason, so nobody trusts a guess.
Nothing to undo
Every check is a read. Changes to infrastructure are drafts until approved with the admin password, and they leave as a pull request.
Knowledge stays
Every reviewed incident and every fix confirmed by checking again becomes training data. A new model replaces the old one only if it scores at least as well.
No secrets kept
Tokens live in your CI's secret store or this machine's keychain, never in a file. They're redacted from logs, Slack and anything sent to an LLM.
Set up
Install it from anywhere
One npm command installs Rootsight wherever Node and Python run; then run setup in your repository. Setup reads your code and monitoring first, asks only what it can't find, and writes nothing until you agree.
npm install -g rootsightrootsight init- Look around. Where you run, your services, your CI, your Prometheus and Loki.
- Ask. Features, Slack, the LLM, the admin password; only what it couldn't find.
- Fit to your stack. Writes queries for the metric and log conventions you actually use.
- Review. Every file, package, download and secret, before anything happens.
- Set up. Progress bars with time left; files written whole or not at all.
Where it stands today
- Trained on simulated incidents. On 5 real incidents recorded on our test stack and kept aside, the model in use handed all 5 to a person: none wrong, none solved.
- Cloud discovery is tested against stand-ins shaped like the AWS, Google Cloud and Azure APIs, not real accounts.
- The fact check verifies numbers, names and sources in explanations; it can't judge statements like "within the normal range".