anthropic sdk · open beta

Your LLM
is failing.
Silently.

LLMs don't throw exceptions. They hallucinate, return broken JSON, spike costs, and drift — all while your logs show green. trace.ai catches it.

start free →read docs
< 1ms
overhead per call
4 layers
of detection
20 calls
to warm L5

live simulation — click a scenario

1001status_failure·1007token_accounting_mismatch·2001json_contract_violation·4007high_latency_low_output·5001latency_iqr_fence·5002tokens_iqr_fence·2003enum_contract_violation·4005classify_step_token_bloat·5003cost_iqr_fence·1003empty_output_on_success·
1001status_failure·1007token_accounting_mismatch·2001json_contract_violation·4007high_latency_low_output·5001latency_iqr_fence·5002tokens_iqr_fence·2003enum_contract_violation·4005classify_step_token_bloat·5003cost_iqr_fence·1003empty_output_on_success·

Detection engine

L1 · L2 · L4 · L5

Everything included

The full stack.
No config files.

Per-step cost & token tracking
Every call captured: tokens, latency, cost, model. Per-run Gantt waterfall. Monthly budget alerts.
AI root cause analysis
One click runs claude-sonnet over your entire run. Reads every step, every score, tells you what broke and why.
Step identity & profiles
System prompts embedded with all-MiniLM-L6-v2. Each step gets a semantic profile — L5 baselines are per-step, not project-wide.
Trend detection
Per-step health across time: warming / healthy / degrading / critical. Catches latency creep that per-call scoring misses.
Slack integration
Paste a webhook URL. Step errors, rate spikes, anomaly alerts — right in your team channel. No code.
Sentry integration
Every call as a Sentry performance transaction. Anomalies as issues, fingerprinted by step name.

Setup

Three steps.
Under five minutes.

01
Install
typescript
npm install @trace-ai/sdk
02
Wrap your client
typescript
const tracer = new Tracer({ apiKey })
const anthropic = tracer.wrapAnthropic(
  new Anthropic()
)
03
Use normally
typescript
await anthropic.messages.create({
  model: 'claude-haiku-4-5-20251001',
  messages: [...],
  _trace: { stepName: 'classify' },
})

Stop guessing.
Start tracing.

get started free →read the docsno credit card · free to start