European LLM Observability and Evaluation Tools

#1 LangWatch

Amsterdam, Netherlands Free Developer plan (50k events/month); Growth EUR 29 per core seat per month; Enterprise on request Demo on request

Best for: Teams that want tracing, evals and simulated users in one open-source tool

• Operating company. Reasoning Engines B.V.
• Jurisdiction. EU (Netherlands)
• Where the data sits. Cloud regions EU, US, UK and APAC; or your own infrastructure
• Source code. Closed source
• Replaces. LangSmith, Arize, Helicone

LangWatch puts tracing, evaluators, agent simulations and prompt management into one workspace, which covers the loop most agent teams assemble by hand: see what the agent did, score it, then replay realistic user scenarios before a change ships. The simulations use an LLM to play the user over several turns, with a judge that grades each conversation against criteria written in plain language, and they run inside pytest or vitest so they can gate a pull request.

Reasoning Engines B.V. is the Amsterdam company behind it, and the terms follow the Dutch Civil Code; the vendor does not state a founding year and no customers are named on the pages we reviewed.

The code is published under Apache-2.0 on GitHub, and according to the vendor the cloud runs in EU, US, UK and APAC regions or in your own Docker, Kubernetes or VPC setup. Growth costs EUR 29 per core seat per month, and the free plan keeps data for only 14 days.

Standout feature. LangWatch is the only entry that pairs Apache-2.0 tracing with simulated users that run as ordinary pytest or vitest tests.

#2 Latitude

Barcelona, Spain Free Starter (20K credits/month); Pro $99/month (100K credits); Enterprise on request Demo on request

Best for: Teams running agents in production that want failures grouped and triaged

• Operating company. Latitude Data S.L.
• Jurisdiction. EU (Spain)
• Where the data sits. Hosted data stored and processed on European servers
• Source code. Closed source
• Replaces. LangSmith, Helicone, Arize

Latitude starts where other tools stop: after the traces have arrived. Failing sessions become signals, similar failures are merged into one issue with a trend line and example traces, and alerts reach Slack, email or webhooks. A team that has thousands of daily conversations reads a short list of recurring problems instead of scrolling through single runs, and human annotations on any span feed datasets for regression runs.

Latitude Data S.L. operates it from Barcelona under Spanish law, with disputes heard in Barcelona; no founding year or named customers appear on the pages we reviewed. The code is MIT-licensed on GitHub, ingestion works through the SDK or any OpenTelemetry pipeline, and the vendor says hosted data stays on European servers and that it is SOC 2 Type II certified. Starter gives 20K credits with 30-day retention, and Pro is $99 a month with 90 days.

Standout feature. Latitude clusters production failures into trended issues by default, so triage starts from a pattern instead of a single trace.

#3 Galtea

Barcelona, Spain Free plan (EUR 0, 1,500 credits/month); Enterprise on request Demo on request

Best for: AI teams in banks, insurers and public bodies that must document how they tested a model

• Operating company. Galtea Solutions S.L.
• Jurisdiction. EU (Spain)
• Where the data sits. Cloud, or self-hosted or private tenant on Enterprise
• Source code. Closed source
• Replaces. LangSmith, Arize

Galtea begins with a specification instead of a dataset. You write down what an AI system should and should not do, and the platform generates test cases, simulated multi-turn conversations and red-team attempts from it, then scores the answers with judge evaluators. That order suits a bank or an insurer that has policies on paper but no labelled production traffic yet.

Galtea Solutions S.L. is a Barcelona company, and its about page says it emerged from research at the Barcelona Supercomputing Center; no founding year or named customers are stated. According to the vendor, the site lists ISO 27001 and GDPR data processing agreements, and the pricing page publishes credit costs per operation. The free plan has 1,500 credits, one user and 90-day retention, while private-tenant or self-hosted deployment is Enterprise on request.

Standout feature. Galtea generates a test set from your written requirements, so evaluation can begin before any user has touched the system.

#4 Giskard

France Founded 2021 Open source free; Professional on request Demo on request

Best for: Teams that need to test models and agents before release

• Operating company. GISKARD AI SAS
• Jurisdiction. EU (France)
• Where the data sits. On-premise, private cloud or SaaS with data residency options
• Source code. Closed source
• Replaces. Weights & Biases Weave evaluations, Arize, Deepchecks

Giskard is a testing and red-teaming tool for LLM agents and ML models, and it is the one entry here that is built to run before release rather than watch production. It scores prompt injection, data disclosure, sycophancy, hallucination, contradictions and inappropriate refusals, and the Professional tier lists more than 50 automated adversarial probes, including multi-turn attacks and coverage of risk frameworks such as OWASP.

GISKARD AI SAS is a French company, and its site footer says the product is built in France.

According to the About page, it was founded in 2021 by people with experience at Dataiku and Thales; no customers are named on the pages we reviewed. The library is Apache-2.0, an SDK links it to CI/CD, and the enterprise tier lists on-premise, private cloud or SaaS with data residency. Against LangSmith it is a narrower tool: it does not trace, version datasets or serve models.

Standout feature. Giskard turns security worries such as prompt injection and data disclosure into repeatable adversarial tests that a French vendor runs in your CI pipeline.

#5 Orq.ai

Amsterdam, Netherlands Pay-as-you-go: 4.5% fee on Orq-managed keys; 1M free BYOK requests a month, then 4%; Enterprise on request Demo on request

Best for: Teams that want one governed endpoint for all models with traces included

• Operating company. Orq.AI Holding B.V.
• Jurisdiction. EU (Netherlands)
• Where the data sits. Processed in the EEA per its privacy notice
• Source code. Closed source
• Replaces. Helicone, LangSmith, Portkey-style gateways

Orq.ai approaches observability from the gateway. Every request sent through its OpenAI-compatible endpoint is logged as a trace with cost, tokens, latency and full payloads, so a team gets tracing without instrumenting each service. Budgets, spend limits, guardrail rules and PII redaction are applied to the same request, and the vendor says they can alert or block before the money is spent.

Orq.AI Holding B.V. is registered with the Dutch chamber of commerce with an Amsterdam address, and the service terms are governed by Dutch law; the founding year and customers are not stated on the pages we reviewed.

The privacy notice says data is processed in the EEA but names no cloud region. The pricing page counts each routed request, agent step or evaluator run as a span, with 100k spans a month before EUR 7 per 100k, and Enterprise adds on-prem and VPC options.

Standout feature. Orq.ai applies model routing, spend limits, PII redaction and tracing at a single gateway call.

#6 Rhesis AI

Potsdam, Germany Free cloud allowance and free self-hosted edition; see website for paid cloud plans Demo on request

Best for: Teams where non-engineers know the right answers and must review the agent

• Operating company. Rhesis AI GmbH
• Jurisdiction. EU (Germany)
• Where the data sits. Cloud or self-hosted
• Source code. Closed source
• Replaces. LangSmith, Arize

Rhesis AI is built for the situation where the people who know the right answer are not the people who build the agent. Reviewers open a shared link, chat with the live agent, turn conversations into test cases and leave pass or fail verdicts on whole answers or single turns. Each verdict is attached to the test case and the agent version that produced it, and recurring feedback becomes metrics that run on every change.

Rhesis AI GmbH is the Potsdam company named on the imprint, and the README says the project is made in Potsdam, Germany; no founding year or customers are stated. The code carries an MIT badge plus an enterprise-licensed part, and the documentation lists conversation simulation, adversarial testing, DeepEval-based metrics, tracing, SSO and audit-log streaming. The cloud has a free monthly allowance and the open-source edition is not metered, but paid cloud prices were not verifiable.

Standout feature. Rhesis AI lets domain experts judge a live agent through a shared link and converts each verdict into a regression test.

#7 Coralogix AI Center

Ramat Gan, Israel AI data $1.50 per 1M tokens; logs $0.42 per GB; traces $0.16 per GB; 14-day free trial Demo on request

Best for: Companies that already want one observability platform for infrastructure and AI agents

• Operating company. Coralogix Ltd.
• Jurisdiction. Israel (outside the EU/EEA)
• Where the data sits. EU1 (AWS Ireland) and EU2 (AWS Stockholm) regions available
• Source code. Closed source
• Replaces. LangSmith, Arize, Datadog LLM observability

Coralogix AI Center is one module of a general observability platform, and that is its distinguishing trait. Evaluators for hallucination, PII leaks, relevance and toxicity are assigned to each agent and score messages as they arrive, a session explorer follows a user journey end to end, and the vendor lists cost tracking per message, session and agent next to ordinary logs, metrics and traces.

Coralogix Ltd. is incorporated in Israel with its principal place of business in Ramat Gan, according to the master subscription terms, and Israel counts as European on this site.

The footer shows Coralogix Inc. and the terms allow an affiliate to be the contracting party, so confirm the entity on the order form. Documentation lists EU1 on AWS Ireland and EU2 on AWS Stockholm. AI data costs $1.50 per 1M tokens, and there is no self-hosted edition on the pages checked.

Standout feature. Coralogix AI Center is the only tool here that shows a failed model call next to the infrastructure logs and metrics beneath it.

#8 Qualifire

Israel (city not stated), Israel Free plan; Pro $550/month; Enterprise on request Demo on request

Best for: Teams that want real-time policy enforcement and evaluation, not only after-the-fact tracing

• Operating company. Qualifire Ltd.
• Jurisdiction. Israel (outside the EU/EEA)
• Where the data sits. SaaS, your cloud or on-premise; facilities in Israel per privacy policy
• Source code. Closed source
• Replaces. LangSmith, Arize, guardrail add-ons

Qualifire checks output while the request is still in flight. Small judge models, tailored to a business domain and its policies, test for prompt injection, PII, hallucination and grounding on each call, and the same results feed an observability view and a prompt studio with experiments. The vendor also links Rogue, an open evaluator and red-team tool, from its homepage.

Qualifire Ltd. issues the terms, governed by the laws of the State of Israel, and the privacy policy mentions facilities in Israel and relies on the European Commission adequacy decision for Israel. Founding year, city and customers are not stated. Plans include a free tier and Pro at $550 per month with 1 billion monthly tokens and 90-day logs, and Enterprise adds BYOC or on-premise deployment. The pricing page gives conflicting free-tier token limits, so verify them.

Standout feature. Qualifire runs small domain-tuned judge models in the request path, so a policy violation can be stopped rather than merely logged.

#9 Opper

Sweden, Sweden Pay as you go: 3% platform fee (Gateway), 5.5% (Control Plane); Enterprise custom Demo on request

Best for: Teams that want an EU gateway first and basic trace debugging with it

• Operating company. Opper Technology AB
• Jurisdiction. EU (Sweden)
• Where the data sits. EU-hosted API with EU routes; zero data retention by default
• Source code. Closed source
• Replaces. Helicone, LangSmith (tracing only)

Opper is a gateway first, with an Observe and Control section where every call is logged with model, provider, status, tokens, cost, latency and region. Tracing records a tree of spans for model calls, tool calls and the rules that fired, and a trace can be replayed in the playground. Per-call checks flag, redact, block or score a request, billed at a flat rate per check.

Opper Technology AB is registered in Sweden under organisation number 559446-3720, according to its terms; the founding year, city and customers are not stated. Calls are logged as metadata only by default, and inputs, outputs and spans are stored only when a retention rule says so. Gateway costs a 3 percent platform fee and Control Plane 5.5 percent, with full traces kept 1 to 30 days. The pages show no dataset or experiment tooling.

Standout feature. Opper stores only call metadata unless you explicitly set a retention rule, which makes zero data retention the starting point.

#10 ZenML

Munich, Germany Open source free; Scale SaaS $999/month; Enterprise on request Demo on request

Best for: ML engineers who want pipelines to move between orchestrators and clouds

• Operating company. ZenML GmbH
• Jurisdiction. EU (Germany)
• Where the data sits. Self-hosted on your infrastructure, or ZenML Pro SaaS
• Source code. Closed source
• Replaces. Weights & Biases, Comet, Databricks MLflow

ZenML is the odd entry in this list, because it orchestrates Python pipelines instead of tracing LLM calls, and it earns its place through the layer beneath an LLM project. You write steps once and run them on the orchestrator and cloud you already operate, and each run stores artifacts and metadata, which gives lineage from a model back to the data that produced it. A companion product, Kitaru, records and replays agent sessions on production data.

ZenML GmbH is registered in Munich (HRB 268487) and its terms are governed by German law. The site does not state a founding year and names no customers on the pages we reviewed.

According to the vendor, the home page lists SOC 2 Type II and ISO 27001, and the core is Apache 2.0. Expect no chart-heavy run-comparison dashboard on the listed plans, and note that the open source model registry is described as basic. Scale SaaS at $999 per month is a large step from the free plan.

Standout feature. ZenML lets a pipeline written on a laptop move to Kubernetes or another orchestrator without a rewrite, which none of the tracing tools attempt.