Skip to main content
The /metrics endpoint exposes Prometheus metrics for self-hosted, single-tenant EigenPal deployments.

Authentication

Send your API key in the Authorization header as Bearer YOUR_API_KEY:
Replace eigenpal.example.com with your deployment’s hostname and YOUR_API_KEY with your API key. A successful request returns HTTP 200 and metrics in text format. Use a key with Observability (read-only) or Full access scope, created by an owner or admin of the installation’s organization. To create a dedicated metrics key, go to Settings > Monitoring > Create an observability key.

Metrics reference

All metric names start with eigenpal_ and use base units: seconds and bytes. The tables list labels emitted by EigenPal. Prometheus adds job and instance labels to each series. Worker CPU counters reset when the worker process restarts. Run, human review, and webhook counters and histograms also carry internal: true for organizations listed in EIGENPAL_INTERNAL_ORG_IDS (demo and test organizations), false otherwise. Filter on internal="false" to keep demo traffic out of alerts.
type is workflow or agent.
Join the counters to eigenpal_automation_info to show names. For example, the workflows that used the most tokens in the last 24 hours:
Or the OCR pages each workflow processed per minute:
Counting starts when your deployment is upgraded to the release that added these metrics. For earlier usage, see Settings > Usage > LLM usage by workflow. Active workflows are reported at 0 before their first run. A deleted workflow keeps its counters and is named Deleted (<id>). On very large installations, workflows beyond the first few hundred are reported as automation_id="other".Token counts match the Usage page. On that page, input tokens include cache reads and writes; here, total input is input + cache_read + cache_write. Agent runs report their tokens when the run reports its totals, not after each model call.
These counts refresh every 5 minutes.
Test events sent with Send test are not counted. Series carry no endpoint, URL or organization labels. Use the delivery history on each endpoint page to find which receiver is failing.
Each worker process has a unique worker label that changes on restart. After 45 seconds without a heartbeat, eigenpal_worker_up becomes 0. Unresponsive workers remain listed for 10 minutes; workers that shut down cleanly disappear immediately. Offline workers report only identity, uptime status, heartbeat age, and start time.Event loop lag measures timer delays. High lag can delay lease renewals and heartbeats. Use the maximum to detect individual stalls that the mean or 99th percentile can hide.
Use the up series that Prometheus records for each target to check endpoint health.

Alert rules for webhooks

Importing the dashboard does not install alerts. These rules cover the usual webhook failures; adjust thresholds and the job selector to your setup.