FAQ

Common questions

Everything you need to know about instrumenting LLM apps with Puse.

How is Puse different from Langfuse or LangSmith?
Those tools optimize prompts and evals. Puse optimizes uptime — error rates, cost regressions, and pager-style Slack alerts. Think SRE for AI agents, not prompt engineering.
Do you store prompts?
No. Puse stores metadata only: model, tokens, latency, cost, error messages, and trace IDs. Your completions never land in our database.
Does streaming work?
Yes. The SDK wraps streaming responses and captures token usage as chunks arrive, so cost tracking stays accurate even on long agent runs.
Can I self-host?
Absolutely. Puse runs on a single VPS: Fastify ingest, Next.js dashboard, Postgres. See the self-host guide for Docker Compose instructions.
Which providers are supported?
OpenAI and Anthropic out of the box via one-line SDK wrappers. More providers are on the roadmap — reach out if you need one today.
How do alerts work?
Configure thresholds per project for error rate and cost spikes. When a regression is detected over a rolling window, Puse posts to your Slack channel with context.