Compare

Traza and Langfuse

Both store LLM and agent traces, both are open source, and both accept OpenTelemetry. They are built differently and aimed at different jobs. This page sets out the differences as each project documents them, so you can pick the one that fits.

Last checked October 17, 2026

At a glance

	Traza	Langfuse
What it is	A trace database for agent workloads that acts on what it records: run fuses, replay, a hash-chained ledger	An LLM engineering platform: tracing, prompt management, evaluations, datasets and a playground
License	Apache-2.0, the whole engine	MIT, except the `ee/` directories, which carry their own license (per its LICENSE file)
Self-hosted, you run	One binary and a data directory	Web and worker containers, PostgreSQL, ClickHouse, Redis or Valkey, and S3-compatible storage (per its self-hosting docs)
OpenTelemetry ingest	OTLP/HTTP, protobuf or JSON, at `/v1/traces`	OTLP/HTTP, protobuf or JSON, at `/api/public/otel`
Cloud billing unit	Spans	Units: traces, observations and scores each count
First paid cloud plan	Pro: $49 / mo, 20M spans / month	Core: $29 / mo, 100k units / mo, then $8 / 100k units, lower with volume

Langfuse facts from its pricing page, self-hosting docs, OpenTelemetry docs and LICENSE, read on October 17, 2026.

Choose Traza when

You need to stop runs, not just see them. Run fuses hold each run to cost, step, repeat and failure-streak limits, with a pre-call check, a verdict on every acknowledgement and a kill switch. Guard
You want to re-run agents for free. Replay serves recorded answers through the OpenAI and Anthropic SDKs, refuses the first divergent request, and diffs runs. Replay
Someone will audit what you stored. A hash-chained ledger, evidence packs with one digest, erasure receipts, and a notary on Traza Cloud. Prove
You want the smallest thing to operate. Self-hosting is one process with no external database.
You think in spans. Billing is per span with millions included, never per seat.

Choose Langfuse when

Prompts live in your observability tool. Langfuse manages prompts with versioning, release labels and composition, and a playground to try them.
You want evaluation built in. Online and offline evaluations, datasets and experiments from the SDK or the UI, and annotation queues for human review. Traza records scores and runs against datasets, but leaves the runner and scorers to you.
You need long history on cloud. Langfuse Pro offers three years of data access; Traza Cloud’s longest self-serve retention is 90 days on Team.
You need high availability you run yourself. Langfuse runs on PostgreSQL, ClickHouse and Redis, each of which can be deployed highly available; the Traza engine is single-node, with no replication or failover in this release.

Running both

Both accept OTLP over HTTP, so an OpenTelemetry Collector can fan the same spans out to each while you evaluate. Langfuse authenticates with a Basic header built from its public and secret keys; Traza Cloud with a bearer API key. Nothing about either requires the other to change.

Spotted something out of date? Both products move quickly. The Langfuse facts on this page are as its own documentation stated them on October 17, 2026; follow the links above for the current word.