Inspect is an open-source framework for evaluating large language models, developed by the UK AI Security Institute (AISI) together with Meridian Labs. It provides building blocks for datasets, agents, tools and scorers, a large set of prebuilt evaluations and a web-based log viewer. Inspect is installed as a Python package (pip install inspect-ai) and runs on your own infrastructure; to test a model you add that provider's package and an API key. Organisations use it to test the safety and capabilities of models in a reproducible way, as an alternative to commercial AI evaluation and red-teaming platforms.
The project comes from a UK government body and therefore sits under UK jurisdiction, outside the EU and the US. The software is published under the MIT licence, so you can inspect and modify the code and use it internally without a vendor relationship. No cloud service from the maintainers is needed; which data goes to an external model provider depends on the model you choose to evaluate.