Traceview instruments LLM applications in production — traces, evals, cost, and drift, streamed into one dashboard.
tokens / min
0
streaming
p99 latency
0
budget 800ms
hallucination rate
0
grader: faithfulness
live traces
status operational · region us-east · retention 30d · sampling 1:1
instrumenting llm stacks at
[ 01 / product ]
follow every agent step.
Step-by-step traces for every run — prompts, tool calls, retrievals, and completions, span by span.
score quality in production.
Run graders against live traffic and catch hallucinations before your users file them.
know what every feature costs.
Attribute spend to feature, model, and customer — then cap it before it caps you.
| model | tokens | cost |
|---|---|---|
| gpt-4o | 12.4M | $182.40 |
| claude-3-7 | 8.1M | $141.75 |
| llama-3.3-70b | 22.9M | $30.12 |
[ 02 / architecture ]
The SDK wraps your calls. The collector batches and redacts. Storage keeps it queryable. The dashboard tells you what to fix.
Instrument any LLM call.
@traceview/sdk · otlpBatch, sample, redact PII at the edge.
1:1 · head samplingColumnar traces, query-ready.
30-day retentionLive queries, evals, alerts.
p99 in 2 clicks[ 03 / pricing ]
$0/mo
for side projects and first traces
most deployed
$49/mo
billed monthly
custom
volume pricing, invoiced
every tier ships the full sdk. no per-seat fees.
One agent run or request, however many steps. Every prompt, tool call, and completion is a span inside it.
Batching is asynchronous and off the hot path. Typical overhead is under 2ms per call.
OpenAI, Anthropic, and LangChain out of the box; anything else via the OTLP endpoint.
US or EU region, your choice at project creation. Zero-retention mode is available.
Enterprise deploys into your VPC. The dashboard and SDK are identical.
By traces ingested, never by seat. Overage is billed at published per-trace rates.