/metrics endpoint exposes Prometheus metrics for self-hosted, single-tenant
EigenPal deployments.
Authentication
Send your API key in theAuthorization header as Bearer YOUR_API_KEY:
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 witheigenpal_ 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.
Queue
Queue
Runs and steps
Runs and steps
type is workflow or agent.LLM and OCR
LLM and OCR
Usage by workflow
Usage by workflow
Join the counters to
eigenpal_automation_info to show names. For example,
the workflows that used the most tokens in the last 24 hours: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.Organizations and users
Organizations and users
These counts refresh every 5 minutes.
Human review
Human review
Webhooks
Webhooks
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.
Worker processes
Worker processes
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.
Endpoint
Endpoint
Use the
up series that Prometheus records for each target to check endpoint
health.Histogram buckets
Histogram buckets
Alert rules for webhooks
Importing the dashboard does not install alerts. These rules cover the usual webhook failures; adjust thresholds and thejob selector to your setup.