Directory / Inspect
Inspect
Verified
UK AISI’s open framework for evaluating language model agents.
Inspect is an open-source eval framework from the UK AI Security Institute (formerly AISI). It is built for rigorous agent and model evaluations, including sandboxed tasks. It is closer to science than to a PR bot.
- Model independence
- Multi-model
- Deployment
- Self-hosted
- Open source
- Yes
- License
- MIT
- Pricing model
- Free
- Added / updated
- 19 Aug 2026 / 19 Aug 2026
inspect.aisi.org.uk · Docs · Source
Editorial take
Inspect is what you use when the factory must survive an audit, not a demo. Heavier than Promptfoo. Listed because “CI for agents” sometimes means national-lab-grade evals. Not a merge commenter.
Strengths
- Serious, public eval framework with sandbox-aware tasks.
- MIT, model-agnostic.
- Designed for agents, not only chat prompts.
Limitations
- Steeper than product eval tools.
- Not a GitHub app.
- You will write tasks. That is the work.