Content
<div align="center">
<sub>OMD.EXE // PUBLIC BETA 0.3.0b2</sub>
# OMD.EXE
**A local AI context inbox for Markdown, Obsidian, and agents.**
Turn documents, web pages, screenshots, audio, folders, and selected public
URLs into traceable Markdown that stays in folders you control.
[](https://github.com/omd-local/markdown-everything/actions/workflows/ci.yml)
[](https://www.python.org/downloads/)
[](LICENSE)
[](docs/privacy.md)
[Live demo](https://shionshine-omd-public-demo.hf.space) ·
[Install](#quick-start) ·
[Walkthrough](#walkthrough) ·
[Documentation](#documentation) ·
[Report an issue](https://github.com/omd-local/markdown-everything/issues)
</div>
## Walkthrough
<p align="center">
<a href="docs/assets/omd-walkthrough.gif">
<img src="docs/assets/omd-walkthrough.gif" alt="OMD.EXE walkthrough: add URLs or files, choose Markdown or an Obsidian vault, and run the local conversion" width="640">
</a>
</p>
<p align="center">
<strong>Try the hosted sample with a public webpage or non-sensitive document.</strong><br>
Private files, cookies, vault writes, media transcription, and local models belong in the local app.
</p>
## How conversion stays recoverable
<p align="center">
<a href="docs/assets/omd-context-pipeline.png">
<picture>
<source media="(max-width: 600px)" srcset="docs/assets/omd-context-pipeline-mobile.png">
<img src="docs/assets/omd-context-pipeline.png" alt="OMD source-to-context pipeline: inspect each source, select a document, web, or media adapter, normalise to Markdown, optionally run a model with raw-content fallback, and save a Markdown note plus a traceable OMD sidecar" width="640">
</picture>
</a>
</p>
Core conversion never depends on AI. Optional model work is isolated, so a
missing model or failed call leaves the raw Markdown intact.
## Quick start
```bash
brew install omd-local/omd/omd
omd doctor
omd-ui
```
Choose **Markdown file** for a normal export or **Capture to vault note** for an
Obsidian folder. An Obsidian vault is just a local folder; OMD does not require
an Obsidian plugin or account.
The custom Homebrew tap includes the project organisation in the command.
`omd-local` is the project owner, not a personal account. The beta installs the
CLI, MCP server, local browser UI, common document converters, and `yt-dlp`.
Large transcription and local-model tools remain optional.
<details>
<summary><strong>SETUP MENU // manual install and optional local tools</strong></summary>
### Manual install
```bash
git clone https://github.com/omd-local/markdown-everything.git omd
cd omd
pip install -e '.[all]'
omd doctor
```
Python 3.10 or newer is required. A minimal `pip install -e .` registers `omd`
and `omd-mcp`; the `all` extra adds the UI, MarkItDown, and Python `yt-dlp`.
### Optional local model
OMD never downloads Ollama models automatically. Install and start Ollama, then
run the model install command in Terminal. OMD recommends a conservative
instruct model from the machine's total memory; this is a 16 GB example:
```bash
ollama pull qwen3:4b-instruct
```
Keep the UI host at `http://localhost:11434` for fully local model calls. Core
conversion still works when Ollama is absent or stopped.
### Optional transcription and OCR
```bash
# Apple Silicon speech-to-text for audio, podcasts, and supported videos
brew install pipx
pipx install mlx-whisper
# Text recognition inside screenshots and article images
brew install tesseract tesseract-lang
```
**OCR** means optical character recognition: it reads visible text from an image.
English uses `eng`; mixed Chinese and English can use `chi_sim+eng`. `mlx-whisper`
is Apple-Silicon only and is intentionally excluded from the base install because
its model stack is large.
</details>
<details>
<summary><strong>SOURCE MENU // formats, public URLs, and cookie-gated routes</strong></summary>
- **Documents:** PDF, DOCX, PPTX, XLSX, HTML, CSV, JSON, XML, EPUB, and ZIP
use MarkItDown.
- **Images:** PNG, JPG, WEBP, TIFF, and BMP use Tesseract OCR.
- **Audio:** MP3, WAV, M4A, FLAC, and OGG use local Whisper when installed.
- **Web:** articles, WeChat, and public webpages use a source adapter or web
conversion.
- **Public posts:** Reddit, X, Bluesky, Mastodon, Threads, Hacker News, and
Telegram use bounded public adapters.
- **Media:** Apple Podcasts, YouTube, TikTok, and Bilibili preserve metadata and
use local transcription when available.
- **Local batches:** folders and saved one-item-per-line lists route each item
independently.
OMD does not bypass paywalls, login gates, captchas, access controls, or platform
restrictions. Public posts can be deleted, private, quarantined, region-blocked,
or rate-limited; a URL that opens in your signed-in browser may still reject an
anonymous converter.
### Douyin and Xiaohongshu / Rednote
These advanced local-only routes normally need separate Netscape `cookies.txt`
exports. Export cookies only from an account and content you are authorised to
use, then select the matching file in the UI:
- **Douyin cookies:** export while signed in to `douyin.com`.
- **Xiaohongshu cookies:** export while signed in to `xiaohongshu.com`.
- Do not reuse one platform's cookie file for the other platform.
- Cookies are disabled in the hosted demo and should never be committed.
Use **Inspect source / cookies** before starting. OMD warns when the current
source list contains one of these platforms but its matching cookie file is
missing.
</details>
<details>
<summary><strong>OUTPUT MENU // Markdown, Obsidian, polish, and memory cards</strong></summary>
### Direct Markdown
```bash
omd report.pdf -o report.md
omd screenshot.png -o screenshot.md --lang eng
omd bilingual.png -o bilingual.md --lang chi_sim+eng
```
### Obsidian-compatible vault capture
```bash
omd capture report.pdf --vault ~/Obsidian/AI-Memory --tags research,pdf
omd capture "https://example.com/article" --vault ~/Obsidian/AI-Memory
```
Capture writes a readable note under `Sources/<source type>/`, updates
`Index/OMD Captures.md`, and keeps hashes, route diagnostics, model errors, and
other debug metadata in the adjacent `.omd.json` sidecar.
### Optional local AI sections
```bash
omd report.pdf -o report.md --polish-md --polish-md-keep-raw
omd capture report.pdf --vault ~/Obsidian/AI-Memory --memory-cards
```
`--polish-md` cleans parser/OCR noise. `--memory-cards` adds a summary, useful
tags, `[[links]]`, and evidence-oriented cards above the preserved
`## Full Content`. Review all generated content before relying on it.
Read the [Obsidian guide](docs/obsidian.md) and
[memory cards guide](docs/memory-cards-user-guide.md).
</details>
<details>
<summary><strong>TOOLS MENU // CLI recipes, MCP, and agent-safe mode</strong></summary>
```bash
# Inspect routing, tools, cookies, and local readiness without converting
omd inspect "<url-or-file>" --with-readiness
# Convert a folder or reusable one-item-per-line list
omd batch sources.txt -o out/
omd capture ~/Downloads/sources/ --vault ~/Obsidian/AI-Memory --batch
# Quiet deterministic output for agent-facing runs
omd --agent-safe report.pdf -o report.md
```
`omd-mcp` exposes four tools:
- `convert_to_markdown(uri, output?, output_format?, lang?, reel_options?)`
- `inspect_source(uri, include_readiness?, cookies?, cookies_from_browser?)`
- `capture_to_vault(uri, vault, lang?, tags?)`
- `list_supported_formats()`
Minimal MCP configuration:
```json
{
"mcpServers": {
"omd": {
"command": "omd-mcp"
}
}
}
```
Treat converted Markdown as untrusted input. MCP restricts local paths and
private-network URLs by default; use narrow `OMD_MCP_ALLOWED_ROOTS` only for
folders you intend the client to access.
</details>
<details>
<summary><strong>HELP MENU // common failures and what survives</strong></summary>
| Symptom | What to check |
|---|---|
| `markitdown` missing | Reinstall the Homebrew package or `pip install -e '.[all]'`. |
| OCR language unavailable | Install `tesseract-lang`; use `eng` or `chi_sim+eng`. |
| Local model warning | Start Ollama and run the displayed `ollama pull <model>` command. Raw Markdown is retained. |
| Web or Reddit HTTP 403 | The source rejected automated access. Save an authorised copy as HTML or PDF and convert the local file. |
| Douyin / Xiaohongshu warning | Export a fresh platform-specific cookie file and select it in the matching UI field. |
| Partial failure in a multi-file run | Open the output folder; successful items and partial raw outputs are retained when safe. |
Run `omd doctor` for local capability checks. For a reproducible report, include
a non-sensitive sample, the warning text, and `--verbose` output. Verbose logs
are shown in the process log and are not added to the user-facing Markdown note.
</details>
<details>
<summary><strong>PROJECT MENU // development, acknowledgements, and licence</strong></summary>
```bash
git clone https://github.com/omd-local/markdown-everything.git omd
cd omd
pip install -e '.[all,test,audit]'
make smoke
make test
```
OMD builds on
[Microsoft MarkItDown](https://github.com/microsoft/markitdown),
[yt-dlp](https://github.com/yt-dlp/yt-dlp),
[f2](https://github.com/Johnserf-Seed/f2),
[MLX Whisper](https://github.com/ml-explore/mlx-examples/tree/main/whisper),
[Ollama](https://ollama.com/), and
[Tesseract](https://github.com/tesseract-ocr/tesseract).
Issues and focused pull requests are welcome. Read
[SECURITY.md](SECURITY.md) before reporting a vulnerability. OMD is released
under the [MIT License](LICENSE).
</details>
## Local-first, with explicit boundaries
> **Personal note use only.** OMD is designed for personal research,
> note-taking, and AI-assisted knowledge workflows. It is not a legal,
> compliance, evidentiary, or archival system. Output may omit, reorder, or
> reformat source content. You are responsible for ensuring you have the right
> to access, process, and store each source. OMD does not bypass paywalls,
> access controls, or platform restrictions. Review AI-generated summaries,
> tags, Evidence, and `[[links]]` before relying on them.
Local-first does not mean every URL workflow is offline: URLs contact their
source platform, and explicitly configured remote model endpoints receive the
content sent to them. Keep private material in the local app with loopback
Ollama. See the [privacy model](docs/privacy.md) and
[security policy](SECURITY.md).
## Documentation
| Read this | When you need it |
|---|---|
| [Examples](examples/README.md) | Copy-ready conversion, capture, inspect, batch, and polish commands |
| [Obsidian vault capture](docs/obsidian.md) | Folder layout, sidecars, indexes, and repeated captures |
| [Memory cards guide](docs/memory-cards-user-guide.md) | Local model setup and generated note sections |
| [Privacy model](docs/privacy.md) | What stays local and when network access is used |
| [Positioning](docs/positioning.md) | Product scope and what OMD is not |
| [Changelog](CHANGELOG.md) | Release history and current beta changes |
<div align="center">
**LOCAL SOURCES -> TRACEABLE MARKDOWN -> CONTEXT YOU CONTROL**
[Star this repository](https://github.com/omd-local/markdown-everything) ·
[Open the demo](https://shionshine-omd-public-demo.hf.space) ·
[Report a bug](https://github.com/omd-local/markdown-everything/issues) ·
[MIT License](LICENSE)
</div>
MCP Config
Below is the configuration for this MCP Server. You can copy it directly to Cursor or other MCP clients.
mcp.json
Connection Info
You Might Also Like
No recommendations available