Sutrace vs Langfuse: managed EU-resident agent observability with budget interlock

TL;DR. Langfuse is the genuinely good Apache-2.0 option in the LLM-observability category. If you need self-host, Langfuse is the honest answer — well-documented, actively maintained, the self-hosting docs are the cleanest in the category. Don't migrate off Langfuse for the sake of migrating. The cases where Sutrace makes sense: you don't want to operate your own observability infrastructure, you want hard budget caps that fire synchronously (Langfuse observes; it doesn't enforce), you've hit the MCP gap (Langfuse's MCP support is partial as of early 2026), or you want one dashboard for AI agents alongside hardware, software, and APIs. Langfuse Cloud Pro starts at $50/mo; Sutrace ingest pricing is comparable at low volume and pulls ahead at scale. EU residency by default (europe-west3).

This page is the long version. It covers Langfuse's strengths, the honest gaps, the migration path, and where each tool fits.

What Langfuse gets right

Langfuse is the OSS observability tool that actually works. Most "open source" observability projects are abandonware or thin marketing wrappers around a closed cloud. Langfuse isn't.

1. Apache-2.0, no asterisks. The full server is permissively licensed. You can deploy it commercially without phoning home. The self-hosting docs cover Postgres, ClickHouse, S3-compatible blob storage, OAuth, and SSO. This is the rare OSS project where the self-host path is the supported path, not a second-class citizen.

2. Strong eval tooling. Datasets, LLM-as-judge, custom evaluators, score-based regression tracking. The ZenML comparison of Langfuse vs LangSmith is the honest read — Langfuse's eval surface is competitive with LangSmith's, and the data model is cleaner.

3. Cost transparency. Langfuse computes per-trace cost natively, with model-specific pricing tables that are kept up to date in the OSS repo. The math is auditable.

4. The Cloud tier is reasonable. Langfuse Cloud Pro at $50/mo for the starter tier — generous for the feature set. The Cloud is EU-hostable too.

5. The right philosophy. Langfuse's docs lead with "observe what's actually happening, then run evals against it." Most competitors lead with "buy our prompt management." The Athenic comparison gets at why this ordering matters.

If your hard requirement is self-host and you have an SRE team to operate it, Langfuse is the answer. Stop reading.

When Langfuse still wins

Be honest with yourself. Pick Langfuse if:

• You require self-host. Sutrace is cloud-only. If your data classification mandates everything stay in your VPC and a SCC-bound DPA isn't enough, Langfuse self-host is the answer.
• You have an SRE team. Operating Langfuse production means Postgres, ClickHouse, and a Redis. That's three databases to back up, monitor, and upgrade. If you have a platform team that already does this, Langfuse adds modest operational overhead. If you don't, you'll spend more on operations than you save on subscription.
• You want the Apache-2.0 license. Sutrace's SDK is open source; the cloud platform is not. If license terms are non-negotiable for your buyer, Langfuse wins.
• You're a content / RAG team that lives inside the Langfuse data model. The traces-as-trees abstraction is closer to what RAG pipelines look like than the OTel-spans abstraction.

Where Sutrace pulls ahead

1. Hard budget caps that enforce, not observe

Langfuse computes spend. It doesn't stop spend. When a stuck loop hits — see the RelayPlane runaway $0.80→$47 case — Langfuse will show you the runaway in the dashboard. Sutrace will block the next provider call before it fires.

This is the architectural difference and it matters for production. An observability tool that doesn't intervene is a postmortem tool. Sutrace's budget interlock fires synchronously at the SDK or proxy layer.

2. MCP support that ships now

Langfuse's MCP (Model Context Protocol) integration is partial as of early 2026 — there's an experimental ingestion path but the trace surface for MCP server calls is incomplete. Sutrace tags every MCP server call as a span with the server identity, tool name, input/output, and the calling agent. If you're shipping anything Anthropic-flavoured (Claude Desktop, Claude Code, custom MCP-using agents) this gap shows up fast.

3. On-host prompt redaction

Langfuse stores your prompts. You can self-host to keep them in your VPC, but you still store them. Sutrace's on-host redactor strips PII, keys, and customer-defined patterns before the trace is exported. Your dashboard sees the redacted version. Your audit log is the redacted version. The original never leaves your VPC, even if Sutrace itself is cloud.

For teams with strict customer-data handling (financial services, health, EU public sector) this is the critical difference. You can have Langfuse self-host and still log prompts you shouldn't be logging.

4. Prompt-injection signals natively

Every Sutrace span carries an injection-detection score. Langfuse can be made to do this with a custom evaluator, but it's a build-it-yourself project. Given that the named 2025/2026 CVEs (EchoLeak CVE-2025-32711, CamoLeak CVE-2025-59145, Tenable's 7-vuln ChatGPT disclosure) are no longer hypothetical, having injection telemetry as a built-in attribute matters more in 2026 than it did when Langfuse's roadmap was set.

5. One dashboard, not seven

Langfuse observes LLMs and only LLMs. If you also have hardware (SCADA), software services, web/APIs, or a Datadog bill that's eating you alive (Datadog comparison) — that's a separate tool. Sutrace unifies all of it. For platform teams running portfolios, the unification is worth real money.

6. Managed EU residency

Langfuse Cloud has an EU region. Self-host runs wherever you put it. Sutrace is EU-resident by default (europe-west3, Frankfurt) with no US replication. Same outcome if you use Langfuse Cloud EU; significant operational difference if you'd otherwise self-host.

7. Multi-provider routing visibility

If you use OpenRouter, AWS Bedrock, or any LLM gateway, Sutrace tags every span with both the gateway and the upstream provider that actually served the request. Langfuse can record what your code asked for; the upstream-provider attribution requires custom instrumentation. See the multi-provider routing post.

Side-by-side comparison

Migration playbook

If you're moving off Langfuse, here's the actual sequence.

Step 1. Stand up an OTel collector. If your Langfuse is already receiving via OTel/OpenLLMetry, skip ahead. Most Langfuse deployments use the native SDK; the migration is to switch the SDK to OTel auto-instrumentation.

Step 2. Switch SDK calls. Replace langfuse SDK calls with openinference-instrumentation-* (or our SDK directly). The semantic-convention attributes overlap heavily; for the standard LLM call attributes the translation is 1:1.

Step 3. Export evals. Langfuse datasets export as JSONL. Sutrace ingests them via API. Custom evaluators (Python functions returning a score) port across with a decorator change.

Step 4. Run dual-write for 1–2 weeks. Send to both. Compare a few traces side-by-side. The Langfuse "trace tree" data model collapses cleanly into nested OTel spans.

Step 5. Cut over. Decommission Langfuse self-host (or cancel Cloud) at the end of a billing cycle.

If you're staying on Langfuse self-host but want to add Sutrace for the budget-cap and redaction layer, you can run both. The SDK supports multi-export.

Frequently asked questions

Should I self-host Langfuse or use Sutrace cloud?

If you have an SRE team that already operates Postgres + ClickHouse and your data classification mandates self-host, Langfuse. If you don't want three more databases to operate and a SCC-bound DPA satisfies your DPO, Sutrace.

What about Langfuse Cloud EU?

Reasonable option if you want managed and you want the Langfuse data model. Sutrace's differentiation is then the budget interlock, on-host redaction, prompt-injection signals, MCP tracing, and the broader stack unification — not residency.

How does pricing actually compare?

Langfuse Cloud Pro starts at $50/mo. Sutrace's flat ingest tier is comparable at low volume and pulls ahead at scale because we don't bill per-trace. For a 10M event/month workload, Langfuse Cloud is roughly $300–$600 depending on retention; Sutrace is roughly $200–$400. Self-host Langfuse is "free" in subscription terms but costs you Postgres + ClickHouse infrastructure plus SRE time, which usually lands at $400–$1500/mo for serious workloads.

Can I use Sutrace's SDK with Langfuse?

Both consume OTel. You can fan out from a single OTel collector to both backends during migration. Once cut over, drop one exporter.

Does Sutrace replace Langfuse's prompt management?

No. Most teams move prompts back to their repo (versioned in git) or to a small CMS. Prompt management as a feature is light enough that a dedicated tool isn't load-bearing for most teams.

What about the MCP gap specifically?

MCP server calls show up as spans in Sutrace with the server identity (mcp.server.name), tool name (mcp.tool.name), input/output, and the calling agent. We tag MCP timeouts and errors as their own span events. Langfuse's current MCP support records call shape but the server-identity attribution is incomplete.

Is Langfuse the right choice for evals specifically?

It's competitive with LangSmith and Sutrace. If evals are your primary use case and you don't need the budget-cap / redaction / injection layer, Langfuse is fine. The Hamel Husain commodification post is the honest read on whether any vendor's eval tooling is the right answer at your scale.

Does Sutrace support EU residency?

Default. europe-west3 (Frankfurt). No US replication. DPA here.

Get started

Self-serve. Drop the SDK in your provider client, set a budget cap, and the dashboard fills in within minutes. No sales call. Pricing here. For the broader picture see the AI agent observability use case and the 4-way honest comparison post.