How it works

Drop it in.
See everything.

No collectors to operate, no query language to learn. Instrument once, and SentriNode handles ingestion, analysis, prediction, and the fix.

app.sentrinode.com / dashboardLIVE
p5042ms
p95188ms
errors0.4%
llm spend$0.82/min
Latency · last 30 minp50 p95
anomaly · llm-proxy p95 +203% vs baselineFix It
The walkthrough

From a pip install to a root cause.

Five things happen between your first span and a fix. Here is each one. The numbers in the pictures are illustrative.

01 / INSTRUMENT

Two lines, then your calls are visible.

Add the SDK for Python or Node, or point any OpenTelemetry client at the endpoint. Every LLM call starts streaming: model, tokens, latency and cost.

  • No collector to run
  • No prompt or response content is recorded
  • Already emitting GenAI spans? Skip the SDK
# install pip install sentrinode-llm anthropic import os, sentrinode_llm sentrinode_llm.instrument( tenant="your-workspace", api_key=os.environ["SENTRINODE_API_KEY"], ) # use anthropic / openai exactly as before
02 / COLLECT & ANALYZE

One live model of your whole stack.

Spans land at the ingest and feed a real-time engine. It recomputes percentiles, error rates and per-model cost continuously, and watches every service for anomalies.

  • p50 / p95 latency, error rate, throughput
  • Per-model LLM spend and tokens
  • Two-window anomaly detection
Your appSDK, or any OpenTelemetry client
IngestOTLP endpoint, auth, size and rate limits
Live enginepercentiles · errors · cost · anomalies
Youdashboard, alerts, AI analysis
self-hosted: every box runs on your machine
03 / PREDICT

Warned before the page fires.

A trend forecast spots a latency breach coming, names the culprit service and gives an ETA, so you act while it is still a trend instead of an outage.

  • Forecast from the live trend
  • Culprit service and time to breach
  • Replay any incident frame by frame afterwards
illustrative
p95 latency · forecast
breach in ~14 min · culprit: llm-proxy
04 / FIX IT

From anomaly to ticket in one click.

AI writes the root cause from the live anomaly, drafts a filled-in Jira ticket, and lists the exact diagnostic commands to run.

  • Root cause in plain English
  • Ready-to-file ticket with full context
  • Commands, not just advice
illustrative
Root cause

llm-proxy p95 rose 203% after retries to the upstream model began stacking behind a saturated connection pool.

Ticket
Raise llm-proxy pool size and cap retriesHigh
Run this
kubectl top pods -l app=llm-proxy kubectl logs deploy/llm-proxy --since=15m | grep -i retry
05 / CAP SPEND

A budget that actually stops.

Set a monthly LLM budget and get alerted at your threshold. With the hard cap on, the SDK refuses calls once you are over, instead of reporting it after the bill.

  • Alert at a threshold you choose
  • Hard cap enforced in the SDK
  • Per-model breakdown of where it went
illustrative
$82.10of $100 monthly limit
$0alert at 80%hard cap $100
Hard cap on. At 100% the SDK refuses further calls until the next period or until you raise the limit.
Also built in

For when you need to dig further.

Ghost mode

Replay any incident like a recording. Scrub the timeline and watch the latency spike and the failure cascade as it happened.

What-if simulator

Model a traffic spike, extra latency or more capacity against your real baseline, and see the projected impact and cost.

Adaptive telemetry

Recommendations on what to sample or drop, with savings estimates, so you ingest only what is worth paying for.

Works with what you have

Bring your own instrumentation.

SentriNode reads OpenTelemetry GenAI spans, so anything already emitting them works without adopting our SDK. Costs are computed server-side from the model and token counts — you do not have to send a price.

Any instrumentor

LangChain, LlamaIndex, OpenLLMetry, Traceloop, LiteLLM and the Vercel AI SDK all spell these attributes differently. All of them are read.

39 models, 8 providers

Anthropic, OpenAI, Google, Meta Llama, Mistral, Cohere, DeepSeek and xAI — including provider-prefixed ids from OpenRouter and Bedrock. Override any price without a redeploy.

Machines too

Point an OpenTelemetry Collector at the same endpoint for CPU, memory, disk and network. Works on Linux, macOS and Windows.

Setup instructions → Limits & throughput

An unknown model still records tokens and latency — it is reported as unpriced rather than free, so nothing silently reads as zero.

Ready when you are

See your stack in two minutes.

Free to start. No credit card.

Start free →