MediCrab guide

Before you start

• The draft you are submitting or revising (Word, Markdown, or LaTeX).
• The data table behind the numbers, if you have one.
• The references you actually cited. PDFs you already have can go in too.
• A Windows 10/11 PC or a Mac, with an internet connection.

Put them in one folder. That folder is the project. The files stay on your computer and are never uploaded to this website.

Where manuscripts and data go

MediCrab is a desktop app that runs on your computer, and project files are stored there. Content is sent to a large language model only when you use an AI feature.

Stored locally

Original and working manuscripts, data tables, PDFs, check records, and edit history are kept in the project folder. Statistics are computed locally, and the de-identification key is stored only on your computer.

Sent to a model

When you use AI to answer or revise, the passages involved and any files or images you attach are sent to a large language model; a full check or read-through sends the whole manuscript. Every request is logged under “What left this machine”.

Our server

With the AI models provided through MediCrab, requests are relayed through our server to the model provider. The server handles sign-in, usage billing, and relaying only: it does not write manuscript text to its database or logs, and model replies are held in memory for at most 15 minutes so an interrupted connection can resume, then cleared.

Literature and web search

Only search terms, titles, and DOIs are sent — never the full manuscript. Search terms may be written by AI based on the topic of your manuscript.

Your own API key

With your own API key, or Codex or Claude on your computer, requests go directly from your computer to that provider and never pass through our server.

Using a general chatbot instead? What ChatGPT, Claude and Gemini keep, and a checklist before you upload: Is it safe to upload an unpublished manuscript to ChatGPT?

Our responsibilities, and our advice

• The original stays read-only. AI edits reach the working manuscript only after you accept them.
• We do not train any model on your manuscripts or data, and do not sell them to anyone. The exception is content sent to model providers to complete an AI request you start.
• MediCrab models use only providers that commit not to train on your content; the list is in our Privacy Policy. Retention and safety review on the provider’s side are outside our control.
• De-identify before sending. MediCrab includes a de-identification tool; whether and how thoroughly to use it is your call. Identifiable patient records should not be given to AI. Where your institution has its own rules, they apply.

How model providers handle your content

Once content reaches a model provider, whether and how long it is kept is set by that provider’s terms.

MediCrab models in your account: every model provider we use commits in its terms not to train models on the content you send. Providers usually keep requests and replies for a period (generally no more than 30 days) to detect abuse, then delete them; content flagged by their safety systems may be kept longer and reviewed by authorised staff. Which providers we use, and their terms, are listed in section 6 of the Privacy Policy on our website.

Your own API key: the provider’s API terms apply. Under the published terms of OpenAI and Anthropic, API content is not used for training by default and is generally deleted within 30 days.

Codex or Claude subscriptions on your computer: the terms of your personal account apply, which differ from API terms. Personal ChatGPT and Claude accounts may use conversations to improve models; when a personal Claude account allows training, data can be kept for years. If you do not want this, turn off “improve the model” in that account’s data settings first.

Could unpublished research ideas leak to a model or to others?

According to providers’ published descriptions, a model does not remember your content while handling a request, and does not use it to answer other people. Content could only surface indirectly in a model’s answers if it were used for training, and the providers behind MediCrab models all commit not to do that. What remains is the provider’s retention period: usually no more than 30 days, during which content is kept under the provider’s security rules and may be reviewed by people if flagged.

If an idea is unpublished and highly sensitive, you can send only the methods and framework and leave out the key novel details, use an enterprise API contracted by your institution, or keep that part away from AI altogether.

Your first check

1. Choose the project folder

   Open MediCrab and pick the folder that holds the paper.

2. Import the manuscript

   Import the draft you are working on. The original file stays read-only; edits happen on the working manuscript MediCrab manages.

3. Run the full check

   Start it from Checks. Numbers, citations, and contradictions come back in one list.

Model review and online literature lookup use your MediCrab usage. MediCrab models need a subscription (Plus / Pro), extra usage, or gifted credit; your own API key or Codex / Claude is not limited by MediCrab.

Reading the results

Every finding comes with its evidence: the sentence in your manuscript, the source record that was found, or where the lookup stopped. Results are sorted into the groups below, kept separate so that “not found” is never mistaken for “wrong”.

Verified

Matches the source record. Nothing to do.

Needs attention

A number disagrees, a citation points to another paper, or the record differs from the manuscript. Start here.

Not found

The identifier matched no record. That does not make the paper fake; check the number.

Not covered

The databases used for this check do not include the item. Common for some Chinese journals.

Request failed

The lookup did not finish. Try again later; do not read it as “not found”.

What the citation check looks at

A DOI that opens only proves the identifier exists. MediCrab asks three separate questions about each citation:

1. Is it the right paper?

   Does the PMID or DOI point to the work you meant? A common error is a real identifier that belongs to another article.

2. Are the details right?

   Do the authors, title, year, volume, and pages match the record?

3. Does it support the sentence?

   This needs the text. When the full text is in the project, MediCrab pulls out the relevant passages for you to read; without it, this step is limited.

Lookups use PubMed, Crossref, and Chinese DOI services where they apply. When a work has a DOI, Crossref update records are also checked for retractions.

To check a reference list on its own, without the app, use the free reference checker.

Finding literature and checking sources

You do not need to decide whether something is a “search” or a “chat”. Research directions, clinical questions, and the evidence for a sentence in your manuscript all go in the right-hand panel.

1. Ask in the right panel

   To ask about a sentence in your manuscript, type @ and pick it.

2. Read the answer while sources are checked

   When the answer cites papers, a sources card appears and the Evidence tab opens. In the background MediCrab checks each paper: does it exist, has it been retracted, is there a newer version, and can the answer’s numbers and conclusions be found in its abstract.

3. Add the useful ones

   Only the bibliographic record is saved, with a link to the full text. Download the PDF and drag it onto that row; MediCrab checks that it is the same paper.

Once you have a manuscript, Evidence in the left bar lists your literature table, grouped into Has full text, Missing full text, Title not recognized, and Needs attention, so you can see which references still lack a PDF.

MediCrab checks the papers themselves, not every sentence of the answer. “Not found in the abstract” is a prompt to read the full text, not proof that the paper does not support the claim.

De-identifying data and building table 1

Put an Excel or CSV file in the project folder and it appears under Data in the left bar.

1. Review what identifies people

   The De-identify tab explains, column by column, what could identify someone: a header that says name, or values that look like phone, ID, or record numbers. You can correct any call.

2. Create a de-identified copy

   Names, phone numbers, ID and record numbers become codes such as ⟦P001⟧; dates shift together so follow-up intervals stay true. The same person keeps the same code in every table and chat. The original file is untouched and the key stays on your computer.

3. Build table 1 under Statistics

   Pick a grouping column and the variables to include. You get n (%), mean ± SD, or median (IQR), with optional two-group comparisons. A local program does the arithmetic; nothing is sent out.

Columns that identify people cannot be used as groups or variables. If the table still has unresolved content such as exact dates, results can be viewed and exported locally but not added to a chat.

Statistics currently cover two-group and paired comparisons. For three or more groups, regression, or survival analysis, keep using R, SPSS, or Python.

Identifiers buried in free text are not always recognized. Look through the copy yourself before sharing it with collaborators or publishing.

Drawing figures for your paper

Open Figures in the left bar, choose New figure, and describe what you want in the chat at the bottom right, for example “a diagram of how post-PCI depression may raise MACE risk”.

1. Say what you need

   Describe the content. To draw from your manuscript, say so (for example “based on the manuscript”); otherwise the manuscript is not read.

2. Use Purpose when it helps

   Make it a paper figure, graphical abstract, poster, tall or wide image; draw it as a diagram or a realistic picture; with a picture, edit it or draw from it. Leave everything unchecked and MediCrab decides from your description.

3. Refine and export

   Not right yet? Draw another or say what to change. Export SVG or PNG for your manuscript.

Check the text, numbers, and medical details in every figure yourself, especially in realistic pictures.

Making a graphical abstract? See graphical abstract examples for five study types, size rules and journal AI policies.

What you still decide

• Whether to change anything. Suggestions are written to the manuscript only after you confirm them; the original file is never touched.
• What to do with a disagreement. You can ask a second model for an independent opinion. Disagreement means it is worth a look, not that either side is right.
• What leaves your computer. Anything sent to a model or search service is shown and logged. Keep patient identifiers and data your institution does not allow out of it.

MediCrab does not write the paper, guarantee acceptance, or replace statistical review or reading the papers you cite.

This guide explains how to use MediCrab. For how data is handled, the Privacy Policy and Terms of Service on our website prevail.