A desktop app for Windows and macOS that lets researchers read scanned and handwritten documents, transcribe interviews and analyse texts with the AI provider of their choice, or offline, without Python, Colab or the command line. Every result is checked against its source side by side: where models disagree is highlighted, and corrections never overwrite the AI's text.
Kalamu is Swahili for pen, from the Arabic qalam. The test releases up to 0.2 were called Scriptorium; a newer version moves their jobs, settings and keys over on its first start.
Windows 11 (x64 and ARM64) · macOS 13 and later (Apple Silicon) · English interface (French and German planned) · MIT licence
Free, with every feature included. If it is useful to you, you can support its development.
Status: version 1.1.2. Kalamu is in the Microsoft Store, signed and updated through the Store. Unsigned installers for x64 and ARM64 (Windows warns on first launch) and, from 1.1.2, a disk image for Macs with Apple Silicon (see On a Mac) are on the releases page. Text extraction, transcription and analyses have run against the live OpenAI, Claude and Mistral APIs; Gemini, offline Whisper and the cheaper batch mode have had fewer live runs, so try a small job first. See the roadmap and the 1.1.2 release notes.
- Extract text from PDFs and images, printed or handwritten, with Gemini, OpenAI, Claude or Mistral. Lines kept as on the page, or joined into running text; pages read one at a time or several together, still reviewed page by page. Choose how much each model thinks; every answer may use the whole output the model allows.
- Transcribe audio and video, with speakers and timestamps, in the cloud or offline with Whisper. Speakers can be given their real names. Public YouTube videos are transcribed from their link with Gemini, without downloading them.
- Analyse texts with fields you define: a summary, the people, places and organisations a document names, its tone with a reason, themes from your own list. Ready-made analyses cover the common needs, including the question, argument, methods and findings of a scholarly publication for a literature review.
- Make a detailed table of contents of a long document: a newspaper or magazine issue, an archival file scanned in one piece, a book or report. One model reads each page, another, cheaper one (from any provider) combines the pages into one entry per article or item, with its pages, authors, date or summary.
- Correct existing text: the noisy OCR of a PDF's text layer or of a digital library's text file, corrected page by page with its spelling of the time kept, every change shown against the original.
- Organise your material in projects: documents grouped in collections (one document can be in several), each with every transcription, extraction and analysis made of it, whatever job made it. Add files before reading them, add jobs you ran before without paying again, start a job on a selection, and mark the result you rely on.
- Reconcile names in a project's dictionary: the people, places and terms its analyses listed, their spellings gathered into entries under the label you choose, with merges a model can propose and that you accept, change or turn down.
- Check every result against its source, side by side, with corrections kept apart from the AI output, which is never overwritten. Search a document or a recording across all its pages (Ctrl F), and report inappropriate output from the review.
- Compare models on the same documents, with their differences highlighted in the text and error rates against corrected pages.
- Never pay twice: work is saved piece by piece, interrupted jobs resume, and nothing is sent again with the same settings.
- Pay half for large collections: send a job to the provider's batch service, answered within a day, and close the app in the meantime.
- Save a job's settings as a pipeline and start the next batch in one click. Export it, with its models, settings and full prompts, to cite in a methods section or share with colleagues.
- Then analyse the text: at the end of New job, choose an analysis to run once the text is read and you have checked it. For longer chains, workflows run your pipelines one after another, from ready-made examples or your own, on a folder or a project's selection; each step is a job of its own, and running again pays only for the steps that changed.
- Export to text, Markdown, Word, JSON, CSV, subtitles and searchable PDF, each with a
provenance.jsonrecording the model, prompt and settings. - A glossary explains the words these tools use, from OCR and tokens to error rates and hallucination, for colleagues new to them.
Files go straight from your computer to the provider you choose, with your own API key, kept in Windows Credential Manager or the macOS Keychain. Offline, nothing leaves the computer. AI output can be fluent and wrong: check a sample against the source before relying on it.
Download Kalamu_<version>_aarch64.dmg from the
releases page (Macs with Apple Silicon, macOS 13
or later) and drag Kalamu to Applications. The app is not yet signed with an Apple Developer ID,
so the first time, macOS refuses to open it: open System Settings → Privacy & Security,
choose Open Anyway beside Kalamu, and confirm. macOS may also ask once to let Kalamu use the
key saved in your Keychain. Shortcuts use ⌘ where Windows uses Ctrl; closing the window keeps
Kalamu in the Dock, and ⌘Q asks first while jobs run.
- Specification: what the app must do, and why.
- Roadmap: what is built and what is next.
- Development: building, checks, the browser preview.
- Output formats, decisions, design, and everything else.
- Privacy policy and security policy.
npm installnpm run tauri devRequirements, checks and the sample-data preview are in docs/development.md.
Kalamu is free, and every feature stays free. If it is useful to you, you can help fund its development on Buy Me a Coffee. Supporting is entirely optional and does not unlock additional features; AI providers bill their own usage separately. The same link is in the app, at the foot of the sidebar and at the end of Settings.
By Frédérick Madore, University of Bayreuth. An independent project. MIT licence.