A research workbench that keeps a record of how AI was used.
Collect your reading in one place, search it, and use AI to help you work with it. The workbench keeps a permanent record of every request made to AI: what was asked, which model answered, which sources it used, and what you decided. It works for research in any subject.
Free The workbench for your own computer is free on Windows, macOS and Linux. Download it with a free account. The workbench in the cloud is free too during the beta, for invited researchers: ask for it.
Free beta. Everything in the roadmap is built, and it is free while in beta: registered accounts download it for Windows, macOS and Linux, and the cloud workbench is open to invited researchers. It imports and searches a library, including a Zotero library, slides and pages saved from your browser, answers questions from your own papers, helps you write and check papers, records every AI call it makes before the call is made, and lets Claude read your vault with the same record.
- On your computer
- Cloud and records
- Limited, or needs approval
- Mistakes and risks
- Human decisions
An example of the record the workbench keeps for each AI request. One is made for every request, including requests that fail and test runs where a mistake was planted on purpose. If the record cannot be saved, the request is not sent.
Get started
Free. Register once on this site, and everything below opens to you.
Register
A free account on this site. It is what lets you download the workbench, post in the forum and ask for the cloud workbench.
Create an accountDownload the workbench
For Windows 10 and 11, macOS on Apple silicon, and Linux as an AppImage. One installer with everything in it. Once installed it needs no account and no internet.
DownloadAdd the browser extension
For Edge, Chrome and Firefox, from each browser's own store: save the page or the paper you are reading into a project, with a note on why you kept it.
How to add itAsk for the cloud workbench
The same workbench in your browser, synced with your computer. It is free during the beta, for invited researchers: tick the box when you register, or ask from your account page.
Ask for itConnect Claude, if you use it
Let Claude read your vault while you talk to it, through the Model Context Protocol (MCP): search your papers, read them by page and cite them, with every request recorded. From Claude Desktop or Claude Code on your computer, or as a connector to the cloud workbench.
How to connect it
Questions, ideas or something that did not work? Ask in the forum: every question is read, and the answers help the next person.
What already works
Free during the beta, on Windows, macOS and Linux, and in the cloud by invitation. The pictures are taken from it.
Collect your reading
Everything you read, in projects of your own, however it arrives.
- A start page that shows what is in every project, what needs your attention, and what came in last.
- Projects made from templates - a doctorate, a literature review, a taught module - with folders you can change.
- Importing files, whole folders and links to web pages: PDFs, Word documents, and PowerPoint slides read slide by slide with their speaker notes.
- Saving the page you are reading from Edge or Chrome, with a note on why you kept it. Behind a publisher's reader, the PDF is saved.
- Bringing in a Zotero library: the collections you choose, with their files.
- Writing or pasting a document, with where it came from kept as a reference.
- A public code repository as a source, pinned to the commit you read.
Read and search
Find what you need, and read it with your notes beside it.
- Searching every project by keyword, and by meaning: ask in your own words and get the passages that answer, each linked to its page.
- Reading a paper with your notes beside it, in an editable copy with its figures and tables. Edits are saved as you type.
- Tagging documents, and narrowing a folder by name or by tag.
- Long lists shown a page at a time, and table columns you can resize.
Work with AI
A model on your computer, or your own key for Anthropic, OpenAI or Gemini.
- Summarising a document.
- Answering a question from your own papers, with every passage the answer used recorded.
- Redrawing a figure from a paper as a diagram or a table, the original picture kept.
- Workflows of several steps - summarise, critique, revise - with limits on passes, time and cost.
Write and check papers
From the first draft to the version you send, with every source checked.
- Writing in sections, with citations in APA 7, Harvard, IEEE or Chicago, exported to Word, PDF, EPUB or Markdown, each with a contents list from its headings.
- A paper written in Word checked in as a version, its Zotero citations linked to your references.
- Sealed versions compared: what changed in the text and in the references.
- Sharing a version with a supervisor as one file they open in a browser, their comments brought back into the paper.
- Checking each claim against the source it cites, and each version's citations against its reference list.
The record, and experiments
What every AI call was asked, shown and answered - and studies built on it.
- An Evidence page recording every AI call: the prompt, the model, what it was shown and what it cost, written before the call is made.
- Replaying a recorded AI call, to see whether the same question still gets the same answer.
- Experiments across models and settings, with mistakes planted on purpose, approval asked where the risk calls for it, and a package others can rerun.
- Code written by a model, checked, built and tested in a sandbox with no network, every attempt kept.
Claude and your vault
An assistant reading your papers while you talk to it, with the same record.
- Claude searching, reading and citing your vault: in the cloud as a connector, on your computer from Claude Desktop or Claude Code.
- Every request it makes recorded in Evidence, documents your labels keep from hosted models never given to it, and nothing given in privacy mode.
- On your computer, Claude proposing a note on a document or a reference by its DOI: each waits under Approvals, showing exactly what would be added, until you allow it.
Plan your writing
Limits, deadlines and the pace to meet them, for every paper and project.
- A word or character limit for each paper, shown in its head and beside each section, with sections you leave out of the count.
- Deadlines for a project or a paper, with a time, a time zone and days of warning, and what is coming up on the start page.
- A calendar of every project's deadlines by month, week or agenda, and a file to add them to Outlook or another calendar.
- Statistics for each paper: words, sections, citations, how it grew version by version, and the words a day to reach its deadline.
- Each paper and each project with settings of its own in a drawer, and each project with a colour and an icon you choose.
- Light and dark, following your computer or as you choose.
People
The people your research involves, kept with it.
- An address book shared by your projects, with a list of people in each, imported and exported as CSV or vCard, and messages sent through your own mail program.
The cloud and sync
The same workbench in your browser, free during the beta for invited researchers.
- The workbench in the cloud, signed in with your account on this site, and yours to ask for.
- Searching by meaning in the cloud with your own OpenAI key, keeping back every document whose label does not allow it.
- Your computer linked with a code, and each project synced by hand or by itself every couple of minutes.
- Changes on both sides merged field by field: your writing is never overwritten, and you choose where both sides wrote differently.
- Erasing your whole cloud workspace yourself, from your account page.
See it working
The real workbench, on a made-up doctorate about supervising AI coding assistants. The papers, people and findings in the pictures are invented.
An answer from your own papers
Reading, with your notes beside it
The record of every AI call
What Claude read, recorded
Coming next
Not released yet: some built and being tested, some still planned. Each one says where it sits in the roadmap; the full plan is in the documentation.
Working together in the cloud
Next, after the cloud workbench
Projects shared with the people you invite - a supervisor, a co-author, a lab - each with a role that says what they may do, notes written together, and a log of who did what that cannot be altered unnoticed.
Assistants that can write, in the cloud too
Next
The proposals Claude can make on your computer - a note, a reference by its DOI - in the cloud workbench as well, and more of them, such as putting a citation into a draft. Each waits for you to allow it.
Paying for the cloud, after the beta
When the beta ends
A subscription for the cloud workbench, with storage on top, and the choice of AI on the workbench's keys billed monthly instead of bringing your own. Everyone in the beta is told at least 30 days before; the workbench on your computer stays free.
Surveys
Not scheduled yet
Designing a questionnaire in a project and piloting it on your computer, then sending it to a contact list and collecting answers through the cloud. Once a survey has an answer, changing it makes a new version, and every answer records the version it was given, so the data always means what it meant when it was collected.
How they work: the workbench in the cloud and syncing your computer with the cloud. Every guide is in one place.
Who it is for
Researchers in every subject face the same problems: too much to read, and too little record of how conclusions were reached.
Reviewing the literature
Import thousands of papers and ask questions about all of them at once. Each answer shows the page and passage it came from, so you are not left checking a summary that only sounds right.
Working with research participants
Interview transcripts and field notes record what each participant agreed to, including which AI services may process their words. Anything a participant did not agree to share stays on your computer: the workbench refuses to send it.
Writing and defending your work
Keep numbered versions of your drafts and see what changed between them: the text, the references and the ideas. The workbench can also check each citation, by reading the source again to see whether it really supports the sentence that cites it.
Studying how people use AI
Run controlled experiments across different AI models and settings. Plant mistakes on purpose, record what the people supervising the AI did about them, and export the whole study so that others can repeat it.
Two products, one core
Each works fully on its own. You can connect them, and disconnect them again, without losing anything.
The local workbench
One installer with everything included, for Windows, macOS and Linux, downloaded with your free account on this site. Once installed it needs no account, no subscription and no internet connection. Your documents stay on your computer, AI models run on your computer through Ollama, and the full record is kept there too.
- Storage
- A database and files on your computer
- Search
- By keyword and by meaning, offline
- AI
- Ollama on your computer, or your own AI service keys
- Assistants
- Claude Desktop or Claude Code, through MCP: reading your vault, and proposing changes you approve
The cloud workbench
The workbench in a web browser, with nothing to install, signed in with your account on this site. It does not need the local workbench, and syncs with it when you want it to. Free during the beta, for invited researchers.
- Built
- Your projects online, synced with your computers
- Built
- Search by meaning, sending only what each document's label allows
- Built
- Claude connected through MCP, reading and citing your vault
- Planned
- Sharing projects, roles, and a log of who did what
Keeping your data under your control is built into the design, not a setting you have to find. If your ethics approval does not allow data to be sent to other companies, the local workbench still does everything on its own, including the AI, using models on your own computer.
What works where
A promise, not an advertisement. If a feature is limited, the table says how. Nothing is shown as working when it is not.
There are three columns, because working without the internet and working without our cloud service are different things. Treating them as the same makes software that runs on your own computer look weaker than it is.
| Feature | Your computer offline | Your computer with your own AI keys | Cloud shared service |
|---|---|---|---|
| Importing documents and web pages | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Searching by keyword | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Searching by meaning | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Searching by keyword and meaning together | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Putting the best search results first | Your computer, offline LimitedBasic ordering only | Your computer, with your own AI keys LimitedNeeds an online ranking service set up | Cloud service Works |
| Summaries | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Finding the key ideas in a document | Your computer, offline LimitedLess accurate, and marked as low confidence | Your computer, with your own AI keys Works | Cloud service Works |
| Analysing many documents together | Your computer, offline Not availableToo much text for models on your computer | Your computer, with your own AI keys Works | Cloud service Works |
| Analysing very long documents | Your computer, offline Not availableNeeds a model that can read very long texts | Your computer, with your own AI keys Works | Cloud service Works |
| Maps of how ideas connect | Your computer, offline LimitedBuilt up a little at a time, not across the whole collection at once | Your computer, with your own AI keys Works | Cloud service Works |
| Writing code | Your computer, offline LimitedSmaller models, short pieces of code | Your computer, with your own AI keys Works | Cloud service Works |
| Checking, building and testing code | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Not availableThe cloud's hosting cannot run code in isolation, so it is checked only on your computer |
| Several AI models working in turn | Your computer, offline LimitedOne AI service at a time | Your computer, with your own AI keys Works | Cloud service Works |
| Experiments with mistakes planted on purpose | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Asking a person to approve risky steps | Your computer, offline LimitedRisk judged by fixed rules, not by a model | Your computer, with your own AI keys Works | Cloud service Works |
| Reading diagrams and images | Your computer, offline LimitedNeeds an image-reading model downloaded | Your computer, with your own AI keys Works | Cloud service Works |
| Recognising a document's title, authors and other details | Your computer, offline LimitedBasic recognition only | Your computer, with your own AI keys Works | Cloud service Works |
| Checking references against Crossref and OpenAlex | Your computer, offline Not availableNeeds an internet connection | Your computer, with your own AI keys Works | Cloud service Works |
| Working with Zotero | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Recording every AI request, and replaying it | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Exporting a study so that others can repeat it | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Sharing a copy with a reviewer | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Writing papers with citations, downloaded as Word, PDF, EPUB or Markdown | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service Works |
| Saving the page you are reading from the browser | Your computer, offline Works | Your computer, with your own AI keys Works | Cloud service LimitedSaved by the workbench on your computer, then synced to the cloud |
| An assistant such as Claude reading your vault (MCP) | Your computer, offline Not availableClaude runs on Anthropic's servers, so it needs the internet | Your computer, with your own AI keys Works | Cloud service Works |
| An assistant proposing changes, which you approve | Your computer, offline Not availableClaude runs on Anthropic's servers, so it needs the internet | Your computer, with your own AI keys Works | Cloud service LimitedReading only, for now |
| Adding notes together | Your computer, offline Not available | Your computer, with your own AI keys Not available | Cloud service Works |
| Live sharing, and projects with several people | Your computer, offline Not available | Your computer, with your own AI keys Not available | Cloud service Works |
| Separate organisations, roles and administration | Your computer, offline Not available | Your computer, with your own AI keys Not available | Cloud service Works |
| A log of changes that cannot be altered unnoticed | Your computer, offline Not available | Your computer, with your own AI keys Not available | Cloud service Works |
| One collection on several devices | Your computer, offline Not available | Your computer, with your own AI keys Not available | Cloud service Works |
| Reprocessing very large collections on the server | Your computer, offline Not available | Your computer, with your own AI keys Not available | Cloud service Works |
26 of 32 features work without the cloud service. The 6 that need it are all about working with other people or with very large collections: the cloud adds collaboration, not research features.
Why the record matters
More and more journals and funders ask researchers to say how they used AI. Memory is not a record.
Repeating a result
To repeat a result, you need to know exactly which instructions were used, which version of which model answered, with which settings, and exactly what text it read. Without that, nobody can repeat the work: not a reviewer, not a colleague, and not you, a year and a half later.
Explaining how you used AI
Export the full record of how AI was used in a project: which models, on which sources, at what cost, and which answers you accepted or rejected. That is enough for a methods section, an ethics committee or a data statement.
Knowing when the AI was wrong
AI models sometimes invent references, and write summaries that sound right but are subtly wrong. The workbench checks each claim against the source it cites, and marks any finding it cannot trace to a source.
Deciding what a person must approve
Before an important step, the workbench scores it: can it be undone, how sensitive is the material, and how well is it supported by sources? Risky steps wait for a person to approve them. You choose where the line is, and the record shows where it was.
The workbench started as the research tool for a doctorate on supervising AI coding assistants, which is why its record-keeping is so thorough. That study is the first to use it, but not the only one it is built for.