Now in beta — works with Anthropic

Observability for
AI workflows

trace.ai gives you real-time visibility into every LLM call — tokens, cost, latency, anomaly scores, and AI-powered root cause analysis. Two lines of code to get started.

your-app.ts
// Before
const anthropic = new Anthropic()

// After — that's literally it
const tracer = new Tracer({ apiKey })
const anthropic = tracer.wrapAnthropic(new Anthropic())

// Every call is now traced ✦

< 1ms

Overhead per call

4-layer

Anomaly scoring

Real-time

Dashboard updates

1-click

AI root cause analysis

Everything your AI stack needs

From raw token counts to intelligent anomaly detection — trace.ai covers the full observability stack for LLM-powered products.

Real-time traces

Every LLM call captured instantly — tokens, latency, cost, model, prompt, and output. Streamed live to your dashboard.

🔬

4-layer anomaly engine

Detects hard failures, format violations, output shape mismatches, and numeric anomalies. Scored, ranked, and explained.

AI root cause analysis

One click runs claude-sonnet-4-6 over your entire run. It reads every step and tells you exactly what went wrong and why.

📊

Cost & token tracking

Per-step and per-run cost breakdown. Spot regressions before they hit your bill. Track model distribution over time.

🔔

Slack & Sentry alerts

Critical anomalies fire to your Slack channel and your Sentry project — with full context, scores, and step details.

🔑

Project API keys

Isolate projects, teams, and environments. Each key routes to its own dashboard with its own alert configuration.

Anomaly Detection

Catch what logs can't

Hallucinations, malformed JSON, token accounting mismatches, output shape drift — these don't throw exceptions. They silently corrupt your pipeline. trace.ai's scoring engine catches them before your users do.

L1

Hard failures

Status errors, missing output, token ghosts

L2

Format violations

JSON contract broken, unexpected output type

L3

Shape fingerprinting

Output structure doesn't match what the prompt asked for

L4

Numeric anomalies

Latency spikes, cost outliers, token ratio drift

Criticalrun a3f9…200 pts
parse-request75pts
2001 json_contract_violation +50
3011 bracket_imbalance +25
enrich-context20pts
4007 high_latency_low_output +20
generate-response200pts
1001 status_failure +100
1002 error_present +100
AI Analysis

The pipeline failed in generate-response, but the root cause originated in parse-request — malformed JSON output propagated silently through enrich-context and caused a null-reference crash downstream.

Plug in to your existing stack

trace.ai alerts fit into the tools your team already uses. Set them up in under a minute from your project settings.

#

Slack

Webhook integration

Paste your Slack Incoming Webhook URL and get instant alerts when a step errors or your error rate crosses a threshold — right in your team channel.

Individual step error alerts
Error rate threshold alerts (e.g. >25% in last 20 calls)
Configurable per project
Send a test ping from the dashboard

Sentry

DSN integration

Add your project's Sentry DSN and anomalies fire directly into your Sentry issues feed — with full context, condition codes, and scores attached as tags.

Fires to your own Sentry project
Critical only, Warning + critical, or Off
Fingerprinted by step name (one issue per step)
Includes score, layer, and error codes as tags

Both integrations are configured per-project in Settings → Integrations. No code changes needed.

Up in under 5 minutes

No agents, no config files, no infra to manage.

01

Install the SDK

npm install @trace-ai/sdk
02

Wrap your client

const anthropic = tracer.wrapAnthropic(new Anthropic())
03

Watch your dashboard

// Every call now appears in
// trace.ai in real time

Start tracing your
AI pipeline today

Free to get started. No credit card required.

Get started free