Every finding shows its evidence
The sentence in your draft and the passage in the source, side by side. No guessing why it was flagged.
For medical researchers submitting or revising a paper
MediCrab reads your whole manuscript the way a careful reviewer does: numbers that disagree, statistics that can’t be right, citations that don’t say what you claim. Every finding shows its evidence. You decide what to change.
Windows · macOS · Free to use · 0.34.0 Beta
Full-manuscript check · CitationsDoesn’t match the source…pooled effectiveness after two doses was 92% [11,12].
“Across six studies, effectiveness of two doses was 82%.”
The sentence in your draft and the passage in the source, side by side. No guessing why it was flagged.
AI edits arrive as proposals. You decide what changes.
Projects live locally. Anything sent to a model is shown and logged.
Ask another model to read your manuscript independently. MediCrab combines the findings so you can review them in one place.
What it actually catches
From a single number to a whole conclusion, these slip past tired eyes. Caught after submission, they cost far more.
Abstract: 410 participants · Results: 412.
Both numbers look familiar, so your brain reads them as the same.
OR = 0.454 · 95% CI 1.124–4.311
Both can’t be right. One of them was copied wrong.
The draft says incidence is “shifting to younger patients”. The evidence is one cross-sectional comparison.
Age groups differ; that doesn’t show anything is changing over time.
Beyond the check
All in one project, using the manuscript and data on your own computer.
Search PubMed for real papers and check each one: title, authors, DOI, and whether the abstract supports your sentence. Open-access full text downloads straight into the project.
Names, phone numbers and record IDs become codes, column by column, before anything reaches a model. The original table never leaves your computer.
Baseline tables and two-group comparisons are computed by code, not improvised by a model, and go straight into the manuscript.
Say what the figure is for and how it should look. Data figures are drawn by code from your own table; the model only writes the spec.
Three steps
Select the folder that holds the draft, supporting data, and references.
Import the draft you are revising so the check has a clear manuscript to follow.
Review the collected number, citation, and consistency findings, then decide what to change.
MediCrab checks four kinds of problems: numbers that disagree, such as the abstract vs the results or a table vs the text; wrong statistics, with χ², P values and proportions recalculated by a program; citation problems, such as a PMID or DOI that points to the wrong paper, or authors, title or year that don't match; and contradictions or conclusions that go further than the evidence. Every finding comes with its evidence, and you decide whether to change anything. See Features.
Yes, MediCrab can check whether the references in your paper are real. It looks up the source records in PubMed, Crossref and Chinese DOI services to check that each PMID or DOI points to the paper you meant to cite and that the authors, title and year match; for papers with a DOI, it also checks for retractions. A program queries the databases; it does not rely on a model's memory. When a record isn't found, the result says whether it was "Not found", "Not covered" or "Request failed", so a missing record is never treated as a fake citation. Whether a cited paper actually supports your sentence can only be checked with that paper's full text. See What the citation check looks at. To check a reference list without installing anything, use the free reference checker.
ChatGPT can give feedback and polish the language, but you still need to check every number and reference it gives you. MediCrab adds program-based checks: it recalculates statistics such as χ² and P values, looks up reference records in PubMed and Crossref, and shows the evidence behind every finding. You can also ask another model for an independent read-through and review the combined findings in one place.
AI cannot replace peer review or a statistician. MediCrab helps you find problems with numbers, statistics, citations and the wording of conclusions before you submit, but it does not replace peer review, statistical review or your own reading of the cited papers in full. When two models disagree, that spot deserves a second look; it does not show which one is right. MediCrab does not write your paper or guarantee acceptance.
MediCrab does not use your manuscript to train AI, and your project files stay on your computer. Content goes to a model only when you use an AI feature (a full check sends the whole manuscript), and every request is logged. Models provided through MediCrab come only from providers that commit not to train on your content. With your own API key or a Codex / Claude subscription on your computer, that provider's terms apply; personal accounts may use content to improve models, and you can turn this off in your account's data settings. For details on how long providers keep your data, see Where manuscripts and data go and the Privacy Policy.
MediCrab is free to download. With your own API key or a Codex / Claude subscription on your computer, MediCrab adds no charge or usage limit; the provider’s own charges and limits still apply. Checking reference records and statistics doesn’t need a model at all. To use AI models provided through MediCrab, subscribe to Plus ($20/month) or Pro ($99/month), or buy extra usage that doesn’t expire. See Pricing.
You should usually disclose using an AI tool such as MediCrab to the journal; follow your target journal's instructions for authors. ICMJE says authors should describe whether and how they used AI in both the cover letter and the manuscript, and AI cannot be listed as an author. Elsevier asks for a separate AI declaration just before the references; basic checks of grammar, spelling and punctuation do not need one (checked October 2026). MediCrab can export a draft AI-use declaration (Word) from its record of the models actually used, for you to review before adding it to your manuscript and cover letter.
A Word, Markdown or LaTeX manuscript and an internet-connected Windows 10/11 PC or Mac are all you need. Put your data tables and reference PDFs in the same project folder so you can bring them into a check or a chat. Your original file stays read-only: edits appear as suggestions on a working manuscript and are written only after you confirm, and you can export the finished manuscript to Word for submission. See Before you start.
Start with a real paper
Built for clinical and biomedical researchers who already have a complete draft, supporting data, and references.
Review the findings, check the sources, and decide what to revise before submission.