Syracuse University researchers have open-sourced DiSCoKit, a GPLv3-licensed toolkit for embedding live, controlled LLM interactions directly within web survey platforms.
Researchers from Syracuse University published a paper introducing DiSCoKit (Daemon for in-Survey Conversation Toolkit), an open-source toolkit for embedding and deploying live large language model (LLM) interaction experiences directly within online survey platforms. The toolkit was developed by Jaime Banks, Jonathan Stromer-Galley, Samiksha Singh, and Collin Capano, and published on 31 July 2026.
DiSCoKit was created to address the trade-off in human-AI interaction research between controlled laboratory studies with small sample sizes and scalable online surveys that previously relied on participant recall, hypothetical scenarios, or non-interactive static content.
Syracuse University’s DiSCoKit operates through a three-layer architecture consisting of front-end JavaScript code embedded in the survey platform, a Flask 3 middleware application, and back-end integrations connecting to an LLM service and a database. The toolkit’s source code is distributed under the GNU General Public License v3 (GPLv3) and is hosted in an open-access repository on GitHub.
DiSCoKit is designed to bridge proprietary web survey tools, templated for Qualtrics XM, with cloud-hosted LLM services, templated for Microsoft Azure portal serverless deployments, while maintaining persistent data logging to a back-end database. The middleware acts as a daemon-like service that transparently brokers message data, injects layered system prompts, and handles real-time interaction states without exposing underlying model credentials to study participants.
The framework enables experimental control by passing participant metadata and assigned condition variables from the survey platform to the middleware, allowing researchers to dynamically manipulate LLM behaviour across experimental groups. DiSCoKit supports multi-stage or shifting AI behaviour trajectories within a single session by layering system prompts and adjusting model parameters like temperature during an ongoing conversation.
The toolkit automatically logs full conversational transcripts alongside participant IDs and experimental condition metadata to ensure full data traceability for social science and human-computer interaction (HCI) research.
