
I built Calyx because my monitoring never told me why production broke
When a clinic dashboard went down, Alertmanager only said a probe was slow. Calyx connects live logs and GitHub so the next alert includes impact, a likely cause, and what to do next.
I run a clinic dashboard, a fairly critical medical app. Front-desk staff log in through it to actually do their work. When it goes down, it is not a “we’ll look at it in the morning” situation.
A while back it went down and I had no idea. The login page was still up. Cloudflare was still up. The origin was dead: a 502 on
frontdesk.prohumanhealth.com . I found out from a message, not from my own stack.That bothered me more than the outage itself, because I had monitoring. Uptime checks, Alertmanager, Slack, the usual setup. The alert I got looked like this:
1 alert for alertname=EventioHighLatency severity=warninginstance = https://alerts.arnabbhowmik.in/-/healthy job = eventio-synthetics description = probe duration > 2s summary = Eventio endpoint slow
So I knew something was slow. I did not know if it was a deploy, a timeout, a bad query, a crashed process, or something I had merged an hour earlier. I opened Grafana, CloudWatch, GitHub, and Slack, and spent a while lining timestamps up by hand. By the time I understood what happened, the useful window was already gone.
That is the part existing tools are bad at. They are good at telling you that something is wrong. They almost never tell you what caused it.
I built Calyx for that gap.
What Calyx does
Calyx is an incident investigation tool. You send it logs, you connect the GitHub repo, and it watches for abnormal behaviour on its own. You do not have to sit there creating monitors and thresholds.
When something looks off, it investigates using the live logs and the codebase, then sends an alert that actually has three useful fields:
- Impact: what broke and how bad it is
- Root cause: what changed, including the commit / deploy if there was one
- Recommended action: what to do next
A normal Alertmanager mail says “probe duration > 2s.” A Calyx alert looks more like this:
Error spike on checkout-api · Critical · OpenImpact: Checkout 5xx jumped 18× in 4 minutes. Payments are failing for about 12% of sessions.Root cause: Deployapi@7f3a1cat 17:24 UTC coincided with the spike. Timeouts to payments-ledger.Recommended action: Roll back api to the previous release, then inspect the ledger client timeout.
Same “something is on fire” feeling, except you are not starting from zero.
It does not secretly restart production for you. It explains, then suggests. If a change needs to be made, a person still has to approve it.
How you set it up
There are four steps. After that it is running.
Create a workspace. This is the team space. Incidents, channels, and the people who should see alerts all live here.
Create a project. One service you care about. For me that is the clinic dashboard. Call it something you will recognise later, like
frontdesk , not an internal id.Add a log source. This is the important one. Logs are what Calyx actually watches. You can ship them a few ways:
-
calyx-loggerin a Next.js or Node app (withCalyxLoggerinnext.config, orinit()on the server). It picks upconsole.error/warn, window errors, and unhandled rejections. Frontend and backend get separate write tokens.
- CloudWatch Logs, through a small Lambda that forwards the subscription envelope.
- An installer on the machine (
curl | bash) if the app runs under systemd, PM2, Docker, a log file, or Kubernetes.
- Or just
POST /v1/logswith a source token.
You do not configure “alert me if error rate > 2%.” Logs start coming in and detectors watch them.
Connect GitHub. Install the GitHub App on the repo. Now Calyx can see pushes, deploys, and the files that changed, which is usually the missing piece when you are staring at an error spike.
That’s it. Workspace, project, log source, GitHub. After that it is ready.
What happens when something breaks
A detector looks at services that have been logging recently. If the error rate jumps (or, later, if a service goes quiet or gets slow), it opens an incident instead of sending you a new email every minute for the same problem.
Then an investigation runs against:
- the logs around that window
- GitHub commits and deploys in the same window
- whatever we already know about that service
The result is the card above. The point is to join three things you normally keep in three tabs: telemetry, “what did we ship,” and the actual code.
I also did not want the model to sound sure when it is guessing. Claims on an incident are labelled:
- Observed: this showed up in the logs (error rate crossed 12% on
/checkout)
- Correlated: these two things happened close together (deploy finished three minutes earlier)
- Hypothesis: this is the model’s best explanation, marked as such
- Verified: someone actually confirmed it
If there is not enough evidence, it should say that rather than invent a root cause.
Where you use it
The same investigation is available in a few places, because during an incident nobody wants a new dashboard.
The web app is the incident room: timeline, evidence, thread, and you can ask follow-up questions there.
Slack gets the alert card. You can acknowledge, resolve, and keep talking in the thread. Replies stay attached to that incident.
The CLI and MCP expose the same tools, so you (or Cursor / Claude Code / whatever you use) can ask “why is checkout 500ing?” without leaving the editor.
calyx ask , calyx logs query , calyx logs tail , incident search. The tenant comes from the token; the client cannot pick someone else’s data.A note on letting it change production
Explaining an outage is one thing. Letting software roll back a deploy or restart a process is another, especially on a medical app.
So the rule is: Calyx can propose. A human approves. Cloud credentials stay on your side, not in Calyx. Proposals are dry-run first, stored in the database, and an approval can only be claimed once. That path exists so we can get to “restart this” or “roll this back” later without it being a chat message that happened to say yes.
For now the useful part is still the investigation. Knowing why is the thing I did not have.
What this is not
It is not a full APM suite and I am not trying to replace Datadog graph for graph. It does not auto-fix production. You can keep Prometheus. You should not need to set up a pile of static monitors for Calyx to be useful.
The model will be wrong sometimes. That is why the labels exist, and why anything that writes to prod goes through a person.
Try it
Create a workspace, add a project, connect logs, connect GitHub. When the next error spike happens, the alert should tell you what was affected, what likely caused it, and what to do, by reading your logs and your code together, not by repeating that a probe took more than two seconds.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article