Open source · OpenTelemetry · TypeScript

Keep your logs.
Spend on the signal.

Score diagnostic value on OpenTelemetry logs before you send them to an expensive reasoning model. Every record stays in your archive. Analysis is recommended only for high-value, uncertain, or protected events.

Read the guide Try a sample

v0.3.0 · MIT · npm install jevlogs

Jev Logs routing log records into a background bin and a signal bin. — Background skips the analysis model. Signal is recommended for deeper analysis. Every record is kept.

Works withOpenTelemetryTypeSafe JevVercel AI SDK

Sample routing

The same four logs as the CLI demo

Health checks go to background (keep, skip the analysis model). Failures go to signal (keep, send to analysis). Drag a record or tap a destination. These are the four sample logs from npx jevlogs.

Ready to score diagnostic value.

Incoming logs

INFO

GET /health → 200api-gateway · routine heartbeat

DEBUG

Cache hit: product:482catalog · expected operation

ERROR

Payment capture failedcheckout · protected by severity

WARN

Connection pool near capacitydatabase · investigate degradation

backgroundkeep · skip LLM0

signalkeep · send to analysis0

4 records. Drag, tap a destination, or run the sample.

How it works

Score first. Analyze second.

Jev Logs wraps the OpenTelemetry log exporter you already use. It attaches a diagnostic-value score, a priority, and a routing recommendation. Your archive exporter is unchanged.

Score every record

Each log gets a 0-100 diagnostic-value score, a priority, and an actionable probability from TypeSafe's Jev.

Keep uncertain and protected records

A record can skip analysis only when value is 25 or below, priority is low, and actionable probability is below retainBelow. Errors, protected records, and timeouts stay on the analysis path.

Route a separate analysis branch

Start in annotate mode so every record is exported with jev.* attributes. Use analysis-only mode on a second processor when you are ready to skip low-value LLM calls.

Model price

Typed routing at TypeSafe launch pricing

$0.042per million input tokens
TypeSafe's published launch price

One request returns choices, rubric scores, and probabilities. A routing decision does not need a generated explanation.

Parallel questionsShared log contextFree outputAt TypeSafe launch pricingUncertain answersStay eligible for analysis

Vercel AI Gateway lists $0.04/M. Billing may differ. Jev is early access. Provider speed numbers are not Jev Logs benchmarks. Typed outputs can still be wrong.TypeSafe announcement

Cost estimate

Estimate analysis spend

This uses estimateSavings() from the package. It is an estimate, not a benchmark. Include errors, uncertain decisions, and failures in the fraction still sent to your LLM.

Estimated monthly savings

$1,05187.6% less analysis spend

Analyze everything$1,200.00

Jev + selected LLM calls$149.40

Includes Jev input and question overhead, plus selected LLM input and output. Excludes storage, ingestion, hosting, retries, discounts, and prompt caching. Assumes equal average log sizes; savings depend on your workload.

Install

Wrap the exporter you already have

Install with npm install jevlogs. npx jevlogs runs the offline demo. Add jevlogs.config.json and run npx jevlogs --live for a local OpenTelemetry receiver. Send OTLP HTTP/JSON, set forwardUrl for your collector, or wrap your existing exporter.

Read the guide

telemetry.ts

import { JevLogExporter } from 'jevlogs';
import { BatchLogRecordProcessor } from
  '@opentelemetry/sdk-logs';

const processor = new BatchLogRecordProcessor({
  exporter: new JevLogExporter({
    exporter: yourExporter
  })
});
// Every log kept. Value + priority attached.

Node.js 22+ · AI SDK 7.0.105 · OTel SDK Logs 0.222.0

MIT licensed

Source is on GitHub

View source

Independent project. Not affiliated with TypeSafe, Vercel, or OpenTelemetry.

Keep your logs.Spend on the signal.