Langfuse v4.4.0 in brief: Langfuse is a real option if you run LLM applications and want self-hosted tracing, evaluation, and prompt management instead of sending trace data to a third-party SaaS. At the August 4, 2026 snapshot, the project is active, widely adopted, and part of ClickHouse since January 2026 per the README, but it is not a lightweight add-on: production deployments assume ClickHouse plus a relational database, recent release notes document Docker publishing failures, and open issues cover migration and storage problems. Treat self-hosting as an infrastructure decision. Verify licensing, image tags, storage, and authentication before production.

Key facts at a glance

This article is a source and documentation review based on the repository, README, release notes, and open issue reports collected on August 4, 2026. No local installation or runtime test was attached to the evidence, so statements about behavior come from the maintainers’ documentation and issue reports, not from a hands-on test.

What Langfuse is and what v4.4.0 changed

Per the repository README, Langfuse is an open source LLM engineering platform for developing, monitoring, evaluating, and debugging AI applications. The maintainers list LLM observability and tracing, prompt management, evaluations, datasets, an LLM playground, and a public API with typed Python and JavaScript/TypeScript SDKs as core features. Their integration list includes OpenTelemetry, LangChain, the OpenAI SDK, LiteLLM, LlamaIndex, Haystack, the Vercel AI SDK, and Mastra, among others.

For self-hosting, the README offers Docker Compose locally or on a single VM, Kubernetes with Helm (called the preferred production deployment), and Terraform templates for AWS, Azure, and GCP. The README also states the codebase ships in Alpine-based Docker containers and is built with ClickHouse. Issue reports from operators reference a Prisma-managed relational database plus ClickHouse URLs, which matches the picture of a multi-component stack rather than a single binary.

Release v4.4.0 was published on August 4, 2026. According to the maintainers’ release notes, it adds features such as ClickHouse writer time distribution metrics, migration handling for experiment instrumentation, and in-app agent execution support. Operator-relevant fixes include rejecting OpenTelemetry spans with missing or malformed IDs with a 400 response, preventing NextAuth callback URLs from causing server errors, safer session observation metadata serialization, and dependency bumps for express-rate-limit, ip-address, undici, fast-uri, and postcss.

Who should consider Langfuse

Consider Langfuse if you ship LLM features with real traffic and need a shared view of prompts, latency, failures, and costs across a team, and if you already operate containers plus database infrastructure. The natural audience for self-hosting is teams with data-residency rules, strict cost-tracking requirements, or a policy against sending prompts to a third-party SaaS.

If your team is small, your trace volume is modest, and nobody wants to run ClickHouse and a relational database, the managed Langfuse Cloud the maintainers also offer is the lower-effort route. For evaluation-stage work, the SDK and integration quality matter more than where the server runs.

What to check before production

License terms

The README states the repository is MIT licensed except for the ee folders, while GitHub’s API metadata for the repository reports NOASSERTION. Before building a product on Langfuse, confirm with the license file which components are open under MIT and which are in the ee area, and what that means for redistribution and commercial use.

Version and image tags

The maintainers’ release notes for v4.2.0 and v4.3.0 state that Docker publishing failed for those versions and advise using the latest release instead (latest, 4.3, or 4.3.1). In practice, pin a specific image tag, verify the digest you pull, and read the release notes before upgrading. Note also that the project maintains a v3 line. v3.225.0 was published on August 3, 2026, so confirm which major line you are on and what the migration path is.

Upgrades and migrations

Open issue #7913 describes a self-hosted v2-to-v3.80.0 upgrade that ended in an Internal Server Error on projects.environmentFilterOptions, with traces present in the database but missing from the UI. Open issues #7995 and #4675 describe a web pod startup failure with “Cannot set property message of ZodError” across versions and Helm chart releases. These are documented incidents rather than guarantees, but they justify staging every upgrade with a backup of the database and ClickHouse, plus UI smoke tests after migration.

Media and object storage

Open issue #6504 reports that automatic media upload failed for a self-hosted Langfuse instance using Google Cloud Storage, while a manual upload through the SDK worked. If multimodal inputs or attachments are part of your product, test the media path against your own storage backend before production.

Authentication

Open issue #12738 notes that the MCP server supports only HTTP Basic authentication with static project-scoped API keys and has no SSO/OIDC support. The documented workaround is a reverse proxy in front of the MCP endpoint, which adds operational complexity and does not give per-user identity inside Langfuse. If per-user MCP audit trails are a requirement, plan custom development or accept the proxy compromise.

Cost and token accounting

Open issue #8020 reports a possible Gemini token and cost mismatch when using the Google SDK, where the UI appeared to count total tokens as completion tokens and inflate the cost. If exact cost reporting matters for non-OpenAI models, verify usage mapping with controlled test payloads rather than trusting the dashboard.

Maintenance signals

The repository is active: it was pushed to on August 4, 2026, and the release dates show v4.2.0 (July 31), v4.3.0, v4.3.1, and v3.225.0 (August 3), and v4.4.0 (August 4). That cadence ships fixes quickly. v4.3.1 addressed a CI Docker builder setup issue, but it also means an upgrade treadmill for self-hosters. The README states Langfuse became part of ClickHouse in January 2026, which is relevant context for project longevity. At snapshot time the repository showed 32,498 stars, 3,492 forks, and 740 open issues; several of those open issues are self-hosting bugs with maintainer assignments, which tells you where the rough edges are.

Limitations

• The stack is heavy: the evidence points to ClickHouse plus a relational database, and the README positions Kubernetes/Helm as the preferred production deployment. A single VM may work for small installs, but this review did not verify that.
• Cost accuracy depends on provider pricing mappings, and the open Gemini issue shows those mappings can be wrong for some providers.
• The MCP integration lacks SSO/OIDC, which limits per-user identity in MCP workflows.
• The v4 migration is still in progress: release notes reference routing legacy trace widgets through v2, preserving legacy trace compatibility, and detecting legacy dataset APIs, so some surfaces are transitional.

Alternatives and evaluation questions

If you already accept a managed platform, Langfuse Cloud is the lowest-effort option and is listed in the README alongside self-hosting. For evaluation-only workflows, integrated tools such as promptfoo — described by the maintainers as an open source LLM testing platform — cover part of the evaluation space, but they do not replace a tracing backend.

Answer these questions with your own data before choosing:

• Do you need tracing only, or also evals, datasets, and prompt management? The operational cost differs.
• Is your framework in the integration list, and can your team instrument it correctly?
• What trace volume and retention do you need, and can ClickHouse handle that on your budget?
• Do you need SSO for the web UI and MCP, or are static API keys acceptable?
• Who owns upgrades, backups, and incident response for the platform?

Verdict

Langfuse v4.4.0 is worth adopting if you treat it as infrastructure: you have an LLM product in production, a team that can instrument SDKs, and operators who can run ClickHouse and a relational database. The documented Docker publishing failures, startup errors, and storage issues do not make it a bad choice; they make it a project you should stage-test and monitor. If you cannot staff that, use Langfuse Cloud or a vendor platform, because the maintenance cost of self-hosting will otherwise be the surprise.

Sources

• langfuse/langfuse repository — README, metadata, and issues
• Langfuse v4.4.0 release notes
• Langfuse v4.3.0 release notes (Docker publishing note)
• Langfuse v3.225.0 release notes
• Langfuse self-hosting documentation
• Issue #7913 — projects.environmentFilterOptions migration error
• Issue #6504 — automatic media upload fails with self-hosted GCS
• Issue #12738 — MCP server SSO/OIDC support request
• Issue #8020 — Gemini token/cost mismatch
• Issue #7995 — ZodError startup failure in web pod

I help with Linux server deployment, troubleshooting, and maintenance for teams evaluating projects like this — including Langfuse rollout, staged upgrades, ClickHouse sizing, and object-storage configuration. If you want a second pair of eyes before you commit, that is the kind of work XP812 does; contact me through the site.