int-brain-lab/ibl-ai-agent

IBL AI Agent is a repository that helps neuroscientists use AI coding assistants, such as OpenAI Codex or Claude Code, to analyze data from the International Brain Laboratory (IBL). You clone the repository, open an AI coding agent inside that directory, and ask scientific questions about brain activity data. The agent writes the analysis code, generates plots, produces a report, and can optionally publish the report to a GitHub Pages site. The underlying dataset is the IBL Brain Wide Map (BWM), a large collaborative project that recorded neural activity across the mouse brain while mice performed a decision-making task. The BWM data covers spike times for all high-quality neurons at 0.1 millisecond resolution, along with behavioral traces including wheel movements, stimulus and response events, and video keypoint detections. A compressed version of this data fits under 10 GB and is downloaded into your local repository when you run the install step. The workflow is deliberately interactive rather than one-shot. The intended sequence is: first, work with the agent to clarify and refine the scientific question, second, split the data into an exploration set (which you can analyze repeatedly) and a held-out confirmation set, third, use the exploration set to settle on metrics, statistical tests, and parameters, fourth, lock the analysis plan and run it once on the confirmation set, fifth, let the agent write a Quarto report, sixth, optionally publish. The split between exploration and confirmation sets is meant to reduce false positives by preventing repeated testing on the same data. The README is explicit about limitations. The agent can make scientific mistakes. Users are responsible for checking all assumptions, code, plots, and statistics before trusting any conclusion. The repository is not designed for unattended runs and is not suitable for processing raw Neuropixels recordings or video files, only the preprocessed spike and behavioral data. Users are also responsible for AI API costs, which can accumulate quickly in long sessions. Setup requires cloning the repository, installing an AI coding agent subscription, and typing "install" to let the agent handle the rest of the setup. Example reports from early use are available at a linked GitHub Pages site.