Alerts
Get a Slack message or a signed webhook when one of your instances fails, stops or is preempted.
Alerts tell you what the platform did to your instances: one failed, its host stopped reporting, a reservation preempted it, it stopped or it was deleted. They go to a Slack channel or to a webhook of your own, and every notification is kept for 30 days with how each delivery went.
You set them up in three steps, on the Alerts page of the console, through the API or through the MCP server:
- Create a channel, where notifications go.
- Subscribe to the events you want, for the instances you want, on one or more channels.
- Read the history to see what was sent and what the receiver answered.
Reading alerts needs alerts:read, which every role has. Creating and deleting channels and subscriptions needs alerts:write, which owners have.
Channels
A Slack channel takes a Slack incoming webhook URL:
curl https://api.cloud.confidential.ai/v1/notification-channels \
-H "Authorization: Bearer $CLOUD_API_KEY" -H "Api-Version: 2026-10-15" \
-H "Idempotency-Key: $(uuidgen)" -H "Content-Type: application/json" \
-d '{"name": "ops", "type": "slack", "url": "https://hooks.slack.com/services/…"}'A webhook channel takes any https:// URL on port 443 whose host has a public address. Its response carries signingSecret once; keep it to check deliveries. POST /v1/notification-channels/{id}/rotate-secret makes a new one.
A channel's URL and secret are never returned after that: host and urlEnding tell your channels apart. POST /v1/notification-channels/{id}/test sends one test notification now and returns what the receiver answered. An org may have 20 channels.
Subscriptions
curl https://api.cloud.confidential.ai/v1/alert-subscriptions \
-H "Authorization: Bearer $CLOUD_API_KEY" -H "Api-Version: 2026-10-15" \
-H "Idempotency-Key: $(uuidgen)" -H "Content-Type: application/json" \
-d '{"events": ["instance.failed", "instance.preempting"], "channels": ["nc-…"]}'| Event | When |
|---|---|
instance.failed | The instance's status became failed; reason says why |
instance.unknown | Its host has not reported for 60 seconds |
instance.preempting | A reservation needs its GPUs, and its guest was asked to shut down |
instance.stopped | It stopped |
instance.restarting | A boot ended and the next is starting, at most once every ten minutes for each instance, so a guest that keeps rebooting does not send one for every boot |
instance.terminated | It is gone |
instance.maintenance_scheduled | The platform needs it restarted, and restarts it at its maintenance event's deadline unless you restart it first; reason says why |
cluster.unknown | A cluster's control plane stopped answering, so its state is not known |
cluster.unavailable | A cluster whose control plane was up has not been available for 10 minutes; reason says why |
cluster.terminated | A cluster is gone |
Without a filter, a subscription covers every instance and cluster in your org. filter.projects, filter.instances, filter.clusters and filter.labels narrow it, and all that are given must match. filter.instances matches no cluster, and filter.clusters no instance. A change that matches several subscriptions with the same channel goes to it once. An org may have 50 subscriptions.
Webhook deliveries
A webhook receives a POST with the notification as JSON:
{
"id": "ntf-…",
"type": "instance.failed",
"occurredAt": "2026-10-22T23:12:04Z",
"org": "org-…",
"instance": { "id": "i-…", "project": "default", "status": "failed", "reason": "host_lost" },
"message": "The instance's host was lost, and an instance never moves. Delete it, and create a new one."
}Each delivery carries two headers:
Cloud-Notification-Id, the notification's id. A notification can arrive more than once, so drop repeats by it.Cloud-Signature: t=<unix seconds>,v1=<hex>, wherev1is the HMAC-SHA256, with your signing secret, of<t>.<body>. Check it, and refuse an oldt.
import hashlib, hmac, time
def verify(secret: str, header: str, body: bytes) -> bool:
parts = dict(p.split("=", 1) for p in header.split(","))
expected = hmac.new(secret.encode(), f"{parts['t']}.".encode() + body, hashlib.sha256).hexdigest()
return hmac.compare_digest(expected, parts["v1"]) and abs(time.time() - int(parts["t"])) < 300Answer with any 2xx within 10 seconds. Anything else is retried with backoff, from 30 seconds to an hour, for a day. Redirects are not followed.
History
GET /v1/notifications lists what your org was sent in the last 30 days, newest first. Each delivery has its status: pending while it is retried, delivered, or failed. It also has its attempts, the receiver's last HTTP status, its latency, and why it failed. A channel whose deliveries have all failed for a day shows status: failing, and it keeps receiving.