Giskard
Paris open-core testing and red-teaming for ML models and AI agents, with evaluation datasets and on-premise deployment
Quick Overview
| Company | GISKARD AI SAS |
|---|---|
| Category | MLOps Platforms |
| Headquarters | France |
| Founded | 2021 |
| EU Presence | Yes - France (EU) |
| Data Location | On-premise, private cloud or SaaS with data residency options |
| Open Source | Apache-2.0 (giskard-oss on GitHub) |
| GDPR Compliant | Yes |
| Self-Hosting | Yes - open source library, on-premise for the platform |
| Main Features | Agent red-teaming, 50+ adversarial probes, quality evaluation datasets, RAG metrics, human-in-the-loop review, CI/CD SDK, SSO and RBAC |
| Pricing | Open source free; Professional on request |
| Best For | Teams that need to test models and agents before release |
| Replaces | Weights & Biases Weave evaluations, Arize, Deepchecks |
Detailed Review
Giskard covers the testing step that most MLOps stacks leave thin. It generates adversarial and quality tests for ML models and agents, scoring problems such as prompt injection, data disclosure, hallucination and inappropriate refusals.
The site footer names GISKARD AI SAS and says it is built in France. The About page says the company was founded in 2021 by people with experience at Dataiku and Thales.
Security and quality tests
The security side covers sycophancy, prompt injection, data disclosure and inappropriate content. The quality side covers inappropriate denials, contradictions, hallucinations and omissions.
The Professional tier lists 50+ automated adversarial probes including multi-turn attacks and coverage of agent risk frameworks such as OWASP.
Open core
The Giskard library is published under Apache-2.0 on GitHub. Its README states that v3 is a fresh rewrite, which matters if you used an earlier version.
The open source plan is for local agent experiments, with community support and best-effort maintenance.
Deployment
The enterprise tier lists secure deployment on-premise, in a private cloud or as SaaS, data residency and isolation, a zero-training policy, SOC2, HIPAA and GDPR compliance and dedicated support.
Collaboration features include SSO, role-based access, task prioritisation and scheduled email alerts.
Limitations to Consider
Giskard tests and evaluates; it does not track training runs, version datasets or serve models, so it sits beside tools such as ZenML or Valohai. The open source tier is limited to local experiments with best-effort maintenance, and the Professional price is not published. The v3 rewrite means older scripts may need migration. Its current focus is AI agents more than classical tabular models.
Who Should Use Giskard
European teams shipping LLM agents or regulated ML that need repeatable pre-release testing from a French vendor.
Pros and Cons
Pros
- Apache-2.0 library
- Security and quality tests in one tool
- On-premise and private cloud deployment
- French company, European data residency options
- CI/CD integration through an SDK
Cons
- Testing only, no tracking or serving
- Professional price not published
- v3 is a rewrite of earlier versions
- Open source tier has best-effort support
Alternatives to Giskard
Other European mlops platforms worth comparing before you decide:
ZenML
Germany · Munich open source pipeline orchestration with artifact and model tracking
Valohai
Finland · Turku platform for versioned ML runs, pipelines and GPU scheduling
Hopsworks
Sweden · Stockholm feature store with model registry and serving, free tier
DagsHub
Israel · Experiment tracking, data versioning and annotation on one platform
See all European mlops platforms or compare them directly against Weights & Biases Weave evaluations on our alternatives page.
Frequently Asked Questions
The Giskard library on GitHub carries an Apache-2.0 licence, and the pricing page lists an open-source tier for local agent experiments.
No. It tests and evaluates models and agents, while experiment tracking is handled by tools such as DagsHub or ZenML.
The enterprise tier lists on-premise, private cloud and SaaS deployment with data residency and isolation.