For teams building with Claude

Know when your AI is degrading—before your users do.

LLM observability and evaluation built for small teams shipping production AI. Anthropic-first by design. Free tier, no card required.

Start free Read the docs

10,000 traces/month free. Signup to first trace in under 5 minutes.

One import. Zero changes to your call sites.

Wrap your existing Anthropic client. Klaria captures traces asynchronously, so your request path stays the same speed.

import Anthropic from '@anthropic-ai/sdk'
import { wrap } from '@klaria/sdk'

const claude = wrap(new Anthropic(), {
  apiKey: process.env.KLARIA_API_KEY,
})

// Use Anthropic exactly as you always have.
// Klaria captures traces in the background.
const response = await claude.messages.create({
  model: 'claude-sonnet-4-6',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Hello, Claude' }],
})

TypeScript SDK first. Python SDK in early access alongside it. OpenAI clients work with the same wrap().

What you get

Trace every call

Request, response, tokens, latency, and cost for every LLM call. Filter by model, error state, or time range. Drill into single traces to debug.

Cost visibility

Per-model and per-day spend. Anthropic prompt cache hit rate and dollar savings broken out. Anomaly flag when daily spend runs 3x above your trailing average.

Evals on real traffic

Write a rubric, pick a judge model, and score your actual production traces — not synthetic test sets. Drill into every failing trace.

Built for teams who chose Claude.

Most observability tools were built OpenAI-first and bolted on Anthropic support later. We inverted that. The result: features that actually understand the Claude API—prompt caching, batch monitoring, computer use, MCP servers, extended thinking—instead of generic LLM tracing pretending to fit.

OpenAI works too. Our wedge is depth where it matters.

See your first trace in 5 minutes

Create an account, grab an API key, wrap your client. That's the whole setup.

Start free