SOC 2 Type II · HIPAA ready · EU region

Your AI is in production.
Now watch it.

Kyvath traces every prompt, tool call and token so you can see latency, cost and quality per feature — and catch a regression on Tuesday instead of a customer complaint on Friday.

npm i @kyvath/otel | pip install kyvath-ai | OpenTelemetry native
Traces Cost Evals Alerts
completion · /v1/chat · prod live
chat.stream
1,842ms
↳ retrieve.context
212ms
↳ guardrail.check
74ms
↳ llm.anthropic.sonnet
1,244ms
↳ tool.sql_query
288ms
p95 latency2.1s▼ 18%
Cost / req$0.0041▼ 7%
Faithfulness96.4%▼ 1.2%
⚠ Regression detected — faithfulness on support.answer dropped 1.2% after prompt v14 · view diff
3.1Bspans / month
<15msSDK overhead
34%avg. cost saved
240+teams onboard
18model providers
01 — Platform

One console for the whole AI stack.

Distributed traces

Follow a request from your API handler through retrieval, guardrails, the model call and back. Nested spans, exact token counts, full prompt and completion attached.

Cost attribution

Slice spend by customer, feature, model or team. Set budgets and get pinged before the invoice, not after.

Online evals

Score live traffic with LLM judges, heuristics or your own graders. Failures get sampled into a review queue automatically.

96.4faithfulness

Prompt version diffing

Compare v13 vs v14 by latency, cost and quality on the same traffic — then roll back in one click if the numbers disagree with the vibes.

- You are a helpful assistant.
+ You are a concise support agent. Cite sources.

Session replay

Replay a full conversation with every tool result inline. Debug the weird one-in-ten-thousand case without asking the user for a screenshot.

Alerts that matter

Threshold, anomaly or eval-based. Route to Slack, PagerDuty or a webhook with the trace attached.

02 — Compare

Why teams switch from log-spelunking.

CapabilityKyvathLogs + dashboardsGeneric APM
Nested LLM spans with prompts✓ nativemanual✗
Token-level cost attribution✓ per request✗✗
Online eval scoring✓ built in✗✗
Prompt version diffing✓✗✗
OpenTelemetry compatible✓partial✓
Self-hosted option✓ Enterprise✓✓
Setup time~4 minutesweeksdays
03 — Install

Two lines of code. Really.

node · server.ts
import { Kyvath } from '@kyvath/otel';

Kyvath.init({ apiKey: process.env.KYVATH_KEY });
// that's it — every LLM + tool call is traced
python · app.py
import kyvath, os

kyvath.init(api_key=os.environ['KYVATH_KEY'])
# auto-instruments openai, anthropic, langchain
An engineer on a night on-call shift watching three monitors glowing with traces and mint-green graphs
02:14 — production The regression alert reached the on-call engineer before the customer did.

“We found a $14k/month prompt leak in the first hour. Kyvath paid for itself before the trial ended.”

Dana Whitfield
Dana WhitfieldVP Engineering · Loopwork

“The eval queue changed our release process. We no longer ship prompt changes on hope.”

Idris Malik
Idris MalikML Lead · Cortex Health
04 — Pricing

Volume pricing. Move the slider.

10M
1M10M50M200M500M2B
Estimated monthly bill $99

Includes unlimited seats, 30-day retention, all evals and every integration.

Developer

$0

50k spans/month, 7-day retention, 1 project. Free forever for side projects.

Start free

Team

from $99

Volume-based, unlimited seats, 90-day retention, online evals, SSO.

Start trial

Enterprise

custom

Self-hosted, unlimited retention, custom graders, DPA, named engineer.

Talk to sales
05 — FAQ

Technical questions.

Does tracing add latency to my requests?

Spans are buffered and flushed asynchronously. Measured overhead is under 15ms p99 at 10k requests/minute, and the SDK can be configured to batch harder if you need more headroom.

Which providers are instrumented out of the box?

OpenAI, Anthropic, Google, Mistral, Cohere, Bedrock, Azure OpenAI, Together, Groq, Fireworks and Ollama — plus LangChain, LlamaIndex and Vercel AI SDK.

Can I redact PII before it leaves my server?

Yes. A pre-flush hook lets you scrub prompts and outputs client-side. Enterprise plans also support full self-hosting with no outbound traffic.

Do you train models on my traces?

Never. Traces are encrypted at rest with per-tenant keys and are deleted on your retention schedule.

See what your AI is actually doing.

Free for 50,000 spans a month. No sales call required.