Langfuse is genuinely one of those tools you don't realize you needed until you're 3 weeks into debugging a multi-step LLM pipeline at 11pm. The tracing and prompt versioning features are solid, and the self-hosted option via Docker is a big deal if you're working in regulated environments where data can't leave your infra.
That said, it's not perfect for complex agent workloads. When your trace is a 30-minute agent run with 15+ tool calls, the observation list view gets hard to parse fast. It's observation-first by design, which makes sense for prompt-centric apps but feels limiting once your agents grow teeth.
Worth noting: Langfuse got acquired by ClickHouse in early 2026, so there's some uncertainty around the roadmap. For now though, if you need open-source LLM observability and self-host matters to you, it's still the go-to. Just know that complex agent debugging might push you to look elsewhere eventually.
Langfuse
@nolan_vu Thanks a lot for the kind review!
If your table views get noisy, you can filter e.g. for "agents-only" in the filter sidebar and then save this as a filter view. Feel free to open a discussion on github with more details on your specific setup, would love to help improve this.
Langfuse joined ClickHouse to invest heavily in the project, both on the product and infrastructure/team side. Learn more about this here: https://langfuse.com/blog/joining-clickhouse. For Langfuse users, this is a big upside, as ClickHouse is backed by some of the best investors and has significantly more resources than we had before joining them.