Common questions

Do I need to rewrite my agent code to use Niitaka?

No. Niitaka works by patching your LLM provider client at the import level — you call instrument_openai() (or whichever provider you use) once at startup, and every subsequent API call is automatically captured. Your agent code stays exactly as it is. The only addition is wrapping your agent run in start_session(), which is a one-line context manager.

What happens if Niitaka is unavailable — will my agent stop working?

No — your agent will not crash. If the Niitaka backend is unreachable, the SDK catches the error silently and execution continues. However, event logging is currently synchronous, so a network timeout (2 seconds max per call) may add a brief delay per LLM call before the SDK gives up and moves on.

Is my data — including prompts and responses — sent to Niitaka's servers?

Yes. The SDK captures prompt messages and response text as part of each event by default — this is what powers the trace viewer and replay features in the dashboard. We do not share this data with third parties. If you have strict data residency requirements, self-hosted deployment is on the roadmap. Reach out if this is a blocker for your team.

How is Niitaka different from LangSmith, Arize, or Sentry?

Most observability tools help you see what happened. Niitaka is built around helping you decide what to do about it. On top of tracing and session monitoring, Niitaka adds A/B experiments with live traffic splitting, policy evaluation and replay, agent versioning with one-click rollback, and a real-time policy engine that enforces cost limits and guardrails — without any code changes. If your main need is a trace viewer, other tools may suit you. If you need to run experiments, version agent configs, and enforce guardrails in production, that is what Niitaka is designed for.

Does adding instrumentation slow down my agent?

No measurable latency is added. Event IDs are generated locally (UUID) and all network calls — event creation, event completion, session lifecycle, signals, and decision traces — are dispatched on background daemon threads. Your agent code never blocks on Niitaka I/O. If the backend is unreachable, the background threads silently swallow the error and your agent continues unaffected.

Is there a free tier? What are the limits?

Yes — Niitaka has a free tier to get started. See the pricing page for current plan details.