Turn PDFs, Word, Excel, PowerPoint, images, ZIPs, and web links into clean, token-efficient Markdown — so you can paste high-quality context into ChatGPT, Claude, or Gemini instead of uploading raw files and burning tokens.
Powered by Microsoft MarkItDown. Built by CAIO Group.
AI assistants bill by the token, and a token is roughly four characters. A raw PDF, spreadsheet, or slide deck carries layout, encoding, and repetition the model never needs — often 3–4× more tokens than the same content as clean Markdown.
Markdown Anywhere converts your documents locally and gives you tidy Markdown to paste instead. Less noise, fewer tokens, sharper answers — and nothing ever leaves your computer.
Typical savings: a 20-page scanned PDF ≈ −75% tokens · an Excel workbook ≈ −73% · a slide deck ≈ −75%.
- Drag-and-drop files (or a whole folder), or paste a web link (web pages and YouTube transcripts).
- Formats: PDF, Word (
.docx), Excel (.xlsx/.xls), PowerPoint (.pptx), CSV, HTML, JSON, XML, EPUB, Outlook.msg, images, audio, and ZIP (expands and converts everything inside). - Smart PDF structure — reconstructs headings (by font size) and reads multi-column pages in the correct order. Optional AI mode sends each page to a local vision model for best layout/tables fidelity.
- Local OCR for images and scanned PDFs via Ollama or LM Studio — your files never touch the cloud.
- Token-savings counter showing how much you save per session.
- Queue power tools: preview (rendered ⇄ raw), copy, download
.md, rename, drag-to-reorder, merge several into one document, download all as a zip, and save to a folder of your choice. - History of the session and a Settings panel that persists between launches.
- Light & dark themes.
- Single local app, no accounts, no telemetry, no uploads.
| Convert — light | Convert — dark |
|---|---|
|
 |
Pick whichever is easiest — you just need the project folder on your computer:
- Recommended — a release ZIP: go to the
Releases page
and under Assets download
Source code (zip)(or the attachedmarkdown-anywhere-vX.Y.Z.zip). - Or the green button: on the
repository home page, click the
green
Codebutton ▸ Download ZIP. - Git users:
git clone https://github.com/karagos/markdown-anywhere.git
Then double-click the downloaded .zip to unzip it. You'll get a
markdown-anywhere/ folder — that folder is the app.
One-time. Download from python.org/downloads. On Windows, tick “Add Python to PATH” during install.
Open the unzipped markdown-anywhere/ folder and:
- macOS: double-click
start.command(first time only: if macOS blocks it, right-click the file ▸ Open ▸ Open). - Windows: double-click
start.bat
The first launch finds your Python, creates a virtual environment, installs dependencies (a few minutes), and opens your browser at http://127.0.0.1:8400. Every launch after is instant. Close the terminal window it opens to stop the app.
Prefer the command line?
python3 -m venv venv ./venv/bin/pip install -r requirements.txt ./venv/bin/python -m uvicorn server.app:app --host 127.0.0.1 --port 8400
The app has four sections in the sidebar:
| Section | What it does |
|---|---|
| Convert | Drop files or paste a link. Each item lands in a queue you can preview, copy, download, rename, reorder, merge, zip, or save to a folder. |
| History | A dashboard of your conversions kept locally for a retention window you choose (Settings → History; default 7 days, or Off). Shows per-model performance (avg time, time/page), totals, and lets you re-download, merge, zip, or save past results. |
| Settings | Output folder + auto-save, OCR, PDF mode, theme. |
| About | Why Markdown saves tokens, with examples. |
- Fast — structured (default): a built-in extractor that detects headings by font size and fixes multi-column reading order. Instant, offline, never rewrites your text.
- AI — vision model: renders each page and sends it to your local vision model for the best layout, tables, and reading order. Slower (one pass per page) and requires a connected model (see below).
Images and scanned/image-only PDFs need a local vision model. Either works:
- Ollama — install from ollama.com, then:
ollama pull qwen2.5vl:7b # lighter machines: granite3.2-vision:2b OLLAMA_ORIGINS=* ollama serve # allow the local app to call it
- LM Studio — load a vision model (e.g. a Gemma or Qwen-VL model) and start its local server.
Then in Settings → Enable OCR, click Test connection to load the model list and pick your model. Now drop an image or scanned PDF (or switch PDF mode to AI).
Your files and OCR run 100% on your machine — nothing is uploaded, no account, no telemetry. The only outbound network request is when you paste a web link (that page or YouTube content is fetched from the internet to convert it).
Browser UI ──HTTP──▶ Local FastAPI server ──▶ Microsoft MarkItDown engine
(vanilla JS) (127.0.0.1:8400) + pdfplumber (structured PDF)
+ local vision model (OCR / AI PDF)
- Backend (
server/): Python + FastAPI. Wrapsmarkitdown(with thepdf,docx,pptx,xls,xlsx,outlook,audio-transcription,youtube-transcriptionextras); adds a structured PDF extractor (pdf_text.py), a vision-OCR fallback (pdf_ocr.py), and endpoints for saving, folder selection, model listing, and persisted settings. - Frontend (
web/): plain HTML/CSS/JS — no build step, no framework runtime. Markdown is rendered withmarkedand sanitized withDOMPurify. Fonts and libraries are vendored locally for full offline use.
| Method & path | Purpose |
|---|---|
POST /api/convert |
Convert an uploaded file (multipart) |
POST /api/convert-url |
Convert a web link / YouTube URL |
GET /api/ocr-status |
Detect a local vision model |
GET /api/models?endpoint= |
List models on a local LLM server |
POST /api/save |
Save Markdown files to a folder |
POST /api/open-folder |
Reveal a folder in the OS file manager |
POST /api/pick-folder |
Open the native folder chooser |
GET / POST /api/settings |
Read / persist settings |
markdown-anywhere/
├── server/ FastAPI app + conversion logic (converter, pdf_text, pdf_ocr, ocr, storage, settings)
├── web/ Vanilla UI (index.html, app.js, styles.css, icons.js, lib.js, vendored libs + fonts)
├── tests/ pytest (Python logic) + node --test (frontend helpers) + fixtures
├── start.command macOS launcher
├── start.bat Windows launcher
├── requirements.txt
├── LICENSE MIT
└── README.md
No build pipeline — edit files and refresh. Run the tests on your machine:
# Python (conversion logic, endpoints)
./venv/bin/python -m pytest -q
# Frontend pure logic (Node 18+)
node --test 'tests/*.test.js'Contributions welcome — open an issue or PR.
- Live per-page OCR progress (server-sent events).
- Smarter callout/pull-quote handling in the fast PDF extractor.
- A packaged double-clickable build with a bundled Python runtime.
- Conversion engine: Microsoft MarkItDown.
- PDF text/layout: pdfplumber · rendering: pypdfium2.
- UI libraries (vendored): marked, DOMPurify, JSZip, Geist fonts.
- Local vision models via Ollama / LM Studio.
Made by CAIO Group — AI is an intelligence to communicate with, not a tool to operate.
MIT © 2026 CAIO Group (Stefanos Karagos)