Cloze: Open-source platform for controlled AI therapy research
Supports 4 model families, runs fully on-premises for patient data privacy.
Cloze is an open-source web platform designed to standardize and secure the study of human-AI conversations in mental health research. Developed by Matthew Flathers, Francesco Cipriani, and John Torus, it directly addresses a critical gap: consumer LLM products like ChatGPT, Claude, and Gemini offer researchers very little experimental control, inconsistent data export, and no shared safety scaffolding that works across different providers. Cloze gives research teams a single environment to configure which models participants converse with, how the AI is instructed, how conversations are scheduled over time, and which safety constraints apply unconditionally. Every message is captured with full provenance, including model version, prompt configuration, and timing, enabling rigorous reproducibility.
The platform currently supports OpenAI, Anthropic, and Google’s models, as well as locally hosted open-weight models served through Ollama behind a unified interface. Critically, Cloze can run in the cloud or entirely on-premises, so participant data never has to leave the institution—a major boon for health-privacy compliance (HIPAA, GDPR). The platform is released under the AGPL-3.0 license, and the paper (7 pages, 2 figures) is available on arXiv. The authors emphasize that Cloze is strictly research infrastructure for building an evidence base on human-AI interaction in mental health contexts, not a therapeutic product.
- Supports OpenAI, Anthropic, Google, and locally hosted models via Ollama in a unified interface
- Full provenance capture: model version, prompt config, and timing for every message
- Can run fully on-premises to keep participant data within institutional control
Why It Matters
Cloze gives mental health researchers a reproducible, privacy-compliant sandbox to safely study AI conversations.