Compliance monitoring
Run evals that check AI outputs against loaded regulations and policies. Catch non-compliant responses across back-office and customer-facing flows before they ship.
Observe and evaluate AI agents across your institution. Give engineering, platform, and risk teams visibility into production behavior, with deployment in Langfuse Cloud or your own infrastructure.
Standardize observability and evals across teams: Bring agents across frameworks, models, and gateways into a shared view of traces and evaluations.
Ship reliable agents with clear deployment quality gates: evaluate model and agent changes against datasets before release.
Retain execution history: keep traces of AI executions as evidence for supervisory and regulatory reviews.
Flag potential policy violations: with evals that score outputs against your policies and regulations.
Self-host or air-gap: deploy with no internet access, lock to internal users via VPN to cater for data sensitivity needs.
Used by 21 of the Fortune 50
110B+ observations/month
100,000+ engineers building on Langfuse
Used by 21 of the Fortune 50
110B+ observations/month
100,000+ engineers building on Langfuse


βGenerative AI will only earn enterprise trust when we can see what's happening under the hood. Langfuse enables us to track every prompt, response, cost, and latency in real time, turning black-box models into auditable, optimizable assets.β
Walid Mehanna, Chief Data & AI Officer at Merck
200+
users, engineers, PMs and domain experts
80+
use cases built on Langfuse
1
central platform team
Customer story: Self-hosted


How a European neobroker runs self-hosted Langfuse in production.
Read story βCustomer story: Spend management
How Ramp auto-improves agents on Langfuse.
Read story βCustomer story: Banking
How DKB, Germany's largest digital bank, automates 20,000 customer conversations daily.
Read story βNever in the inference path
Langfuse observes your model and tool calls; it does not proxy them. Docs β
Deployment where your data must stay
Langfuse Cloud in the EU, US or Japan, self-hosted in your VPC, or fully air-gapped. Docs β
Multiple layers of data redaction
Client-side and server-side PII protections before anything is stored. Docs β
Enterprise access controls
SSO and role-based access control scoped to organizations and projects. Docs β
Audit logs
Record both LLM activity and developer actions for internal and supervisory review. Docs β
Managed cloud in the EU, US or Japan.
Your appβOTel SDKβLangfuse Cloud
Platform and risk teams in financial services are looking for a complete observability and eval suite. Langfuse covers the full surface, and we are open about where the work lives in your pipeline rather than in our UI.
Test before release. Monitor production. Review with your experts.
Golden datasets built from production traces, versioned experiments with baseline comparison, and a release gate that fails the pull request on regression.
Docs βDeterministic sampling of live traffic, LLM-as-judge and code evaluators, and threshold alerts to Slack, webhooks, or GitHub Actions when quality drifts.
Docs βAnnotation queues for subject-matter experts, corrected outputs, and one-click promotion of failures into a permanent regression set.
Docs βScore Analytics measures agreement between human labels and model judges (Cohen's Kappa, F1, Pearson, Spearman) so you can defend the judge to model risk.
Docs βControl changes, protect sensitive data, and manage access.
Immutable versions, staging and production labels, protected labels for separation of duties, and full audit history.
Docs βMasking in the SDK before data leaves your application, with trace structure preserved for debugging.
Docs βOrganizations and projects as the data boundary, project-level RBAC, OIDC SSO with domain enforcement, SCIM provisioning, audit logs, and a metrics API for chargebacks.
Docs βConnect your models and manage deployment in your infrastructure.
OTLP ingest from your AI gateway, judge models pinned to your own Bedrock, Azure OpenAI, Vertex, or OpenAI connection.
Docs βDashboards and widgets managed through the API and CLI, versioned in Git, deployed identically to dev, staging, and prod.
Docs βDocumented self-hosting on AWS with Terraform and Helm, a published release cadence, and autoscaling guidance.
Docs βUse cases
Run evals that check AI outputs against loaded regulations and policies. Catch non-compliant responses across back-office and customer-facing flows before they ship.
Support anomaly detection and investigation workflows β reduce cost of service and risk with better observability of agent/tool behavior.
Observe and improve multi-step advisory flows, score outcomes, and keep an audit trail suitable for model risk review. Evaluate your system on representative cases before it runs in production.
Trace agents that combine identity, fraud and credit-bureau checks into a risk summary. Score decisions, flag drift, and keep the auditable trail that credit and model risk teams require.
Raise autonomy rate of AI support agents: the share of cases resolved without human intervention. Trace edge cases, score good/bad runs, and iterate so agents cover more of the long tail safely.
DKB, Germany's largest digital bank: 20,000 conversations / dayTrace Claude Code, Codex, Cursor, OpenCode, and GitHub Copilot β no proxy required. See cost per developer and model, replay failed sessions, and search across the org.
Start with the steps from prototype to production, then see how teams like Trade Republic run Langfuse in production.
Eight steps regulated teams follow, from tracing and error analysis to production monitors and human review.
Evaluate model and prompt changes against datasets and policy evals before release.
Trace a banking assistant with PII masking, offline evals, and a CI gate on a candidate prompt label.
Trace, score and iterate on a support agent to raise its autonomy rate safely.
How a European neobroker runs self-hosted Langfuse in production. Read or watch on YouTube.
Yes. You can deploy Langfuse in your own cloud, VPC, or on-premises infrastructure. Langfuse supports deployments without public internet access (air-gapped); features such as LLM-as-a-judge evaluations need a model endpoint reachable within your environment. Our team can help you assess the setup for your deployment and security requirements. Self-hosting, networking documentation.
Configure SDK masking to remove or replace sensitive inputs, outputs, and metadata before trace data leaves your application. You can also apply masking centrally through an OpenTelemetry Collector within your infrastructure. Combine these controls with project separation and role-based access to determine which teams can access the data you choose to retain. Masking, access controls.
Langfuse's observability integrations collect and export traces in the background while your application continues to call its model provider or gateway. You do not need to route model requests through Langfuse to use observability. If you also use prompt management, SDK caching and fallback prompts help keep your application resilient to connectivity issues. Background export, prompt availability.
Langfuse integrates with gateways such as Kong and LiteLLM to capture model calls, token usage, cost, and latency. Its OpenTelemetry support lets you add AI tracing and evaluation alongside your existing observability stack. Gateway integrations, LiteLLM, existing OpenTelemetry setups. The Langfuse AI Gateway (beta) brings model access controls together with tracing and cost tracking for OpenAI and Anthropic.
Configure retention policies per project and schedule exports of observations and evaluation scores to your own Amazon S3, Google Cloud Storage, or Azure Blob Storage. This lets you preserve exported records under your institution's storage and retention policies. Set up exports before data expires, as data deleted by retention policies cannot be recovered. Data retention, scheduled exports.
Langfuse Cloud is fully managed, with subscription plans and usage-based pricing. Self-Hosted Enterprise runs in your infrastructure, with enterprise administration features and support under a custom commercial agreement; your team operates the deployment. Both Enterprise options include unlimited users. Talk to us to compare features, support, and total costs for your requirements. Cloud pricing, Self-Hosted Enterprise pricing.
Talk through deployment options, compliance needs, and how teams like Trade Republic use Langfuse in production.
No credit card required, free tier available, self-hosting option