Tip: select any text on this page to bring up the Ask tooltip and question the Synedrix OS assistant about it.
What Langfuse is
Langfuse describes itself as an open-source AI engineering platform that helps teams collaboratively debug, analyze, and iterate on their LLM applications.
It is a self-hostable platform with four product surfaces built on one data model: observability through tracing, prompt management, evaluation, and metrics. A trace holds nested observations (spans, generations, and events), so a multi-step agent run is inspectable at every level.
Ingestion is OpenTelemetry-based, which is what keeps the layer substitutable: agents emit OTLP rather than importing a vendor tracing SDK.
What it is normally used for
How Langfuse is used generally, before Synedrix OS enters the picture.
- Debugging non-deterministic applications, where behaviour cannot be reproduced by re-running code alone
- Cost and usage attribution: per-generation token and currency cost rolled into dashboards
- Decoupling prompt changes from code deploys through versioned prompt management
- Testing a change before shipping it, using datasets and experiments as a CI gate
- Scoring live production traffic with model-as-judge and code evaluators
What it owns, and what it does not
Every platform in the stack has a non-overlapping responsibility. The boundary is as much a part of the design as the capability.
Owns
- Prompt versioning
- LLM and agent trace capture
- Token and cost attribution
- Model-call latency and error visibility
- Evaluations and quality scoring
- Nested trace debugging
Does not own
- Workflow sequencing, retries, approval routing
- → n8n (Layer 3)
- Stateful, cyclic, multi-agent reasoning
- → LangGraph (Layer 4)
- Policy decisions
- → Policy Guard (os-agt-002)
- Canonical state and the full outcome record
- → Supabase PG17 (Layer 7)
- Lineage and audit reconstruction
- → Neo4j (Layer 7) and os-agt-008
- Behavioural analytics and feature flags
- → PostHog (Layer 10)
- Infrastructure health and SLO alerting
- → Grafana LGTM Stack (Layer 10)
Why Langfuse
It is the reason LangSmith is not in the stack
LangGraph's own documentation recommends pairing it with LangSmith, a proprietary tool.Synedrix OS uses LangGraph without that pairing. Langfuse was chosen instead, as a separate decision recorded in ADR-OS-001.
It runs with no internet access
Langfuse works without an internet connection. By default it makes exactly one outbound call (a version check that reads from a local cache), and even that can be switched off in the open-source build.
Self-hosted is the same product as hosted
Running Langfuse yourself means running the same software Langfuse sells as a managed service, not a limited version of it. Self-hosting does not cap how much it can scale.
It can be swapped out without rewriting agent code
Langfuse receives data through OpenTelemetry, an open, vendor-neutral standard, rather than its own private software library. Because of that, agent code never has to mention Langfuse by name, so a different observability tool could replace it later without touching the agents themselves.
Licence position
Langfuse questions, answered
Short, citable answers drawn from this page.
What role does Langfuse play in Synedrix OS?
Langfuse (LLM engineering) is mapped to Layer 10: Observability in Synedrix OS. One observability plane across orchestration and runtime. Traces are not an audit trail.
What does Langfuse own in Synedrix OS?
Langfuse owns: Prompt versioning; LLM and agent trace capture; Token and cost attribution; Model-call latency and error visibility; Evaluations and quality scoring; Nested trace debugging.
What does Langfuse not own in Synedrix OS?
Langfuse does not own: Workflow sequencing, retries, approval routing (owned by n8n (Layer 3)); Stateful, cyclic, multi-agent reasoning (owned by LangGraph (Layer 4)); Policy decisions (owned by Policy Guard (os-agt-002)); Canonical state and the full outcome record (owned by Supabase PG17 (Layer 7)); Lineage and audit reconstruction (owned by Neo4j (Layer 7) and os-agt-008); Behavioural analytics and feature flags (owned by PostHog (Layer 10)); Infrastructure health and SLO alerting (owned by Grafana LGTM Stack (Layer 10)).
What licence is Langfuse under?
Langfuse is released under MIT Expat outside ee/; ee/ under a commercial Enterprise License. Everything the OS assigns to this layer is genuinely MIT, but the ee/ directories require a licence key and contain data retention policies, audit logs, project-level RBAC, and ingestion masking: the features a regulated deployment asks for first. Self-hosted telemetry is also on by default, so an air-gapped deployment has to switch it off: the open-source build disables it with an environment variable, while Enterprise telemetry is used for licence compliance and cannot be disabled.
The full reference page, with the operating model, deployment detail, and links to every canonical specification this platform touches, lives in the documentation repository, which is private while the platform is being built.