L0
Answers for the team that has to say yes.
You're the one who checks a new SDK before it touches production. Where we don't have a published answer yet, this page says "to be confirmed" — not a guess.
< 5 ms per trace · sent async
Redact in the SDK
Redact sensitive fields — before anything leaves your machine.
L1
Traces ship off your request path.
Traces are recorded as your code runs and sent asynchronously, off the path that returns a response to your user. The stated overhead is under 5 ms per trace.
L2
Privacy isn't an upgrade.
Traces can carry prompt contents, and prompt contents are sensitive. Redact sensitive fields in the SDK and they never leave your process. Redaction is available on every plan, not held behind a higher tier.
redact — illustrative config, SDK-side
import tracelayer
tracelayer.init(
api_key="tl_live_...",
# redact these fields before a trace leaves your process
redact=["user.email", "messages[].content"],
)import tracelayer
tracelayer.init(
api_key="tl_live_...",
# redact these fields
# before a trace leaves
# your process
redact=[
"user.email",
"messages[].content",
],
)The option names above are illustrative — the exact redaction API is to be confirmed.
L3
What happens to your data.
A trace holds the timing, token counts, cost, and — unless you redact them — the inputs and outputs of each call in a request. Traces go to TraceLayer over the REST API, and you read them in the web dashboard. Redacted fields are removed in the SDK and never leave your process.
L4
Where TraceLayer runs.
If running TraceLayer on your own infrastructure is a hard requirement, ask before you build against it.
L5
Your response doesn't wait on us.
Reliability details are in the table, under L5 · reliability.