About the project
Many researchers now use AI every day: to summarise papers, to search the literature, to comment on drafts and to help with analysis. But these tools rarely keep a record of what they did. Six months later, most of us could not say which AI model wrote a summary, which version of which paper it read, or what we had asked it to do.
At the same time, more and more people are asking for that record. Journals ask authors to say how they used AI. Funders want research that others can repeat. Ethics committees want to know what data was shared, and who agreed to it. Memory is not a good enough answer to any of these questions.
The Research Workbench keeps that record for you, automatically. You collect your sources in one place, search them, and use AI to help you work with them. Each time AI is used, the workbench saves what was asked, which model answered, which parts of which sources it used, what it cost, and what you decided to do with the answer. You can look at this record at any time, and export it to share with others. The workbench is for research in any subject.
What we promise
Your data stays yours. The local workbench runs on your own computer, on Windows, macOS or Linux. You download it with a free account on this site; once installed it needs no account, no subscription and no internet connection, and that includes the AI. If your ethics approval says material must not be sent to anyone else, the workbench refuses to send it.
No hidden limits. When a feature works less well in some setting, the workbench tells you, and tells you why. The table on the overview lists exactly what works where.
Answers show where they came from. When the workbench answers a question, it shows the page and the passage the answer is based on. If it cannot point to a source, it says so, instead of presenting the answer as fact.
Every use of AI is recorded. The workbench records every request it sends to an AI model, including the ones that fail. If it cannot save the record, it does not send the request.
Where it comes from
The workbench started as the research tool for a doctorate. That doctorate studies how people supervise AI tools that write software: how they notice when the AI is wrong, when they insist on approving its work, and whether their trust matches how reliable the AI really is. The study needed controlled experiments, mistakes planted in the AI's answers on purpose, and records detailed enough to replay a session months later.
That is why the workbench keeps such a thorough record, more than a reading tool would normally need. It is also why the workbench is useful for other kinds of research. The doctorate is the first study to use it, but not the only one it is built for.
Designed in the open
The whole design was written down before the code, and it is kept alongside the code. It covers how documents are imported and searched, how the use of AI is recorded, how experiments and approvals work, how research ethics are respected, and the questions that are still open. Open questions are marked as open, not quietly decided.
The design will be published in the documentation, with progress notes as the work goes on. If you work in a subject the workbench should help with, and it does not, please tell me in the forum: now is the easiest time to change it.