# Markdown Editor (md.giit.ch)

> Free Markdown editor in the browser with a live preview, no sign-up and no upload. Write Markdown with syntax highlighting, see the document as you type, click in the preview to jump to the code or edit the text right there, and export to HTML, PDF (print dialog), plain text or the .md file with its pictures.

- Language: English (interface also in German with ?lang=de)
- Cost: free
- Data: documents stay in the browser (localStorage, pictures in IndexedDB) unless the user sets up a webhook to their own server.
- Rendering: markdown-it (CommonMark) in the browser plus KaTeX for math - nothing is loaded from other servers.

## Features

- Code and preview side by side; click in the preview to jump to its place in the code, the cursor's block is framed in the preview.
- Edit text in the preview: paragraphs, headings, list items, quotes and table cells that stand in the code as shown; Markdown characters typed there are escaped (\*, \_, \[ …); every edit is one undo step.
- CommonMark with GitHub's extensions: tables (with alignment), task lists (- [ ] / - [x], clickable in the preview), ~~strikethrough~~, autolinks, footnotes ([^1]); math in $…$ and $$…$$ (KaTeX); headings with GitHub anchors (#my-heading).
- Raw HTML in the Markdown is shown without scripts, styles, frames, forms, SVG and on… attributes; javascript: links are removed.
- Problems listed with their line: missing pictures, links to #anchors that are not there, code blocks that are never closed, table rows with another number of cells than the header, math errors.
- Pictures stored with the document, used by ![description](name.png) or <img src="name.png">; a ZIP with a .md and pictures opens as one document.
- Templates: README, article, meeting notes, checklist.
- My documents: any number of documents in the browser - open, rename, duplicate, download (.md, or ZIP with pictures) and delete.
- Light, dark or automatic color theme.

## Writing for a user (for AI assistants)

Any AI can put a document into the user's open editor with a link; no server, no API key:

- `https://md.giit.ch/#b64=<base64url of the UTF-8 Markdown>`
- `https://md.giit.ch/#md=<URL-encoded Markdown>`
- `https://md.giit.ch/#z=<base64url of the zlib-compressed Markdown>` (shorter)
- Add `&mode=add` to add text at the end of the current document.
- Add `&file=new&name=<name>` to make a new document next to the user's others, or `&file=<name>` to change that one. Without `&file=` the open document changes.
- Long documents: split the encoded text and send one link per part with `&part=N&of=M&id=<name>`.

Open the link in the user's browser if you can, otherwise give it to the user to click. The user can undo every change. In the editor, "AI Assistant" > "Copy prompt" gives the AI these instructions together with the current document.

## Saving on your own server (webhook)

Set under "Save to server" (cloud button), or for one document with `?webhook=<URL>`: only for that one document and only while the page is open - a link never stores a webhook. At the top level the user is asked first; inside a page (iframe or opener) it only works together with `?src=` and only for the document opened with it. On "Save" (Ctrl+S), or automatically a few seconds after each change if chosen, the browser sends:

    POST <webhook URL>
    Content-Type: application/json
    Authorization: Bearer <token>        (only if a token is set)

    {"editor":"md","event":"save","id":"<document id>","name":"<name>",
     "content":"<Markdown>","assets":{"<picture name>":"data:image/...;base64,..."},
     "updated":"<ISO date>"}

One request carries at most 40 MB (with the pictures; `hook.assets:false` leaves them out); the server's error text (e.g. for 413) is shown to the user. The server must answer CORS requests from https://md.giit.ch (Access-Control-Allow-Origin, and Access-Control-Allow-Headers: Content-Type, Authorization). `?src=<URL>` opens a .md file from a URL (CORS needed), `&name=<name>` names it.

## Embedding in a website

    <iframe src="https://md.giit.ch/?embed=1" style="width:100%;height:80vh;border:0"></iframe>

Messages (postMessage; check `event.origin === "https://md.giit.ch"` and `event.source === iframe.contentWindow`):

- Editor -> page: `{type:"mdeditor:ready", version:2, editor:"md", window:false}` when it can take a document.
- Page -> editor: `{type:"mdeditor:load", name, content, assets:{"logo.png":"data:..."}}` opens the document (a document of that name is replaced, otherwise made).
- Editor -> page: `{type:"mdeditor:loaded", name}`, then `{type:"mdeditor:change", id, name, content}` on each edit (not after load), and `{type:"mdeditor:save", id, name, content, assets, editor, event, updated}` when the user clicks Save or presses Ctrl+S.
- Page -> editor: `{type:"mdeditor:get"}` asks for a `mdeditor:save` message; `{type:"mdeditor:save"}` saves as if the user had clicked Save.

In embedded mode the page text below the editor is hidden and a Save button is shown. The LaTeX Editor (latex.giit.ch) and the SVG Editor (svg.giit.ch) speak the same protocol with the prefixes `latexeditor:` and `svgeditor:`.

Optional fields of `mdeditor:load`:
- `context: {host, title, rules}` - plain text about the page (rules up to 2000 characters). "Copy prompt" in the AI Assistant then tells the AI that the document belongs to that page and is saved back there, includes the rules, and leaves out how to build a webhook server.
- `hook: {url, token, assets}` - a webhook (https) for this document only, used on Save in addition to the message; kept in memory only (never stored, gone on reload). `assets: false` sends the body without `assets`. If the server answers with an error, its `{"error": "…"}` (or `message`, or a short text) is shown to the user.

AI Assistant in a frame: add `allow="clipboard-write"` to the iframe so "Copy prompt" can use the clipboard (without it the editor falls back to the old copy command or shows the prompt to copy by hand). Links from an AI cannot reach an editor inside another website (the browser keeps its storage and channels apart), so in a frame the prompt and the dialog ask to paste the AI's whole answer into the editor; use the own window (below) if links should work.

`mdeditor:get` is always answered with a `mdeditor:save` message, also when nothing changed.

### All messages at a glance

Editor -> page (sent to the page around the editor, or to window.opener in its own window):
- `mdeditor:ready` {version:2, editor, window} - can take a document
- `mdeditor:loaded` {id, name, target?} - after each load; remember the id (target:"new" for a new document from the AI, which is not yours)
- `mdeditor:change` {id, name, content, assetsChanged, assetsRemoved} - a real edit of your document (not after load); adding, replacing or removing a picture counts
- `mdeditor:save` {editor, event, id, name, content, assets, assetsChanged, assetsRemoved, updated} - Save / Ctrl+S (from anywhere in the editor), or the answer to get. assets holds ALL pictures and files of your document as data URIs; assetsChanged / assetsRemoved name those that differ from what your page has (after load, save or get)
- `mdeditor:closing` {id, name, content} - own window closed with unsaved changes
- `mdeditor:reload` {id, unsaved} - inside your page, F5 / Ctrl+R / Cmd+R do not reload the editor; ask the user yourself (Ctrl+W cannot be caught by a page)
- `mdeditor:key-request` {reason:"missing"} - the editor has no AI key in this session yet

`?embed=1&save=0` hides the editor's Save button (Ctrl+S still sends `mdeditor:save`) - for pages with their own Save button.

`?lang=de` shows the editor's interface in German (menus, buttons, dialogs, messages), `?lang=en` in English (the default); the choice is remembered in the browser. The AI prompt is always English.

Page -> editor (only from the page around the editor or the opener; after its first message only from that origin; pages without a real origin are ignored):
- `mdeditor:load` {name, content, assets?, context?, hook?, ai?, from?, target?} - open your document; with assets, they replace the document's pictures (send all of them; data URIs named as the Markdown uses them, e.g. "logo.png" or "img/logo.png" - PNG, JPG, GIF, WebP, SVG show in the preview, a PDF shows as a frame with its name, other files are kept untouched and come back with save); from:"ai" shows a note, target:"new" opens it as a separate document in the user's browser
- `mdeditor:get` - answered with `mdeditor:save` of your document (id and content null before the first load)
- `mdeditor:save` - press Save
- `mdeditor:ai` {label?, prompt?, endpoint?, key?, base?} or {clear:true} - AI through your server

Your document is the one your page loaded (or opened with ?src=). change, save, closing and get are only ever about it - never about other documents in the user's browser, also when the user opens them in the editor. A load replaces only a document that came from your page before (same name and origin), never the user's own; the AI settings and the hook from load apply only while your document is open.

### AI through your own server (prompt and session key)

The editor has no backend; your site keeps the keys and the results. Hand the editor your prompt, an endpoint and a key for this editor session - in `mdeditor:load` as `ai: {…}`, or at any time as its own message:

    {type:"mdeditor:ai", label:"Example Admin", prompt:"<your instructions, up to 8000 characters>",
     endpoint:"https://your-site.com/api/ai-result", key:"<session key>",
     base:true}

Fields left out keep their value; `{type:"mdeditor:ai", clear:true}` ends it. One key per editor session: once the editor has a key it keeps it until the session ends or `clear`, and does not take a different one. `base:false` leaves out the editor's own instructions. The AI dialog shows "Connected to <label> (<real host>)", what your page tells the AI, and a choice: change the current document (default) or make a new one. "Copy prompt" copies the editor's short instructions, your prompt (fenced, marked with your real host), the current document and how to send the result:

    POST <endpoint>
    Authorization: Bearer <key>
    Content-Type: application/json
    {"editor":"md","target":"current"|"new","name":"<short name>","content":"<the complete Markdown>"}

An AI that cannot send HTTP requests answers with a link instead, which the user pastes into the editor. When your server has a result, your page sends `{type:"mdeditor:load", name, content, from:"ai", target}`: target "current" (or none) puts it into your document (the user sees a note; Undo takes it back; it counts as not yet saved), target "new" opens a separate document that stays in the user's browser - its edits and Save never go to your page. The key stays valid in the editor for the whole session; only when it has none at "Copy prompt", the editor sends `{type:"mdeditor:key-request", reason:"missing"}`; answer with a `mdeditor:ai` message within 5 seconds.

Safety built in: key, endpoint and prompt are only taken from the page around the editor (after its first message, only from that origin), kept in memory, never stored and never sent by the editor itself - the key only goes into the prompt the user copies. The endpoint must be https on your page's own host (with or without www.) or a subdomain of it; others are refused. Keys must be 16-512 characters of A-Z a-z 0-9 . _ ~ + / = -.

Your backend should: make keys with at least 128 random bits, store only their hash, bind each key to the logged-in user, the document and this editor session, make exactly one key per session, give it no fixed lifetime but keep it valid only while your open page reports the session (e.g. every minute; void it 3 minutes after the last report), void it when the editor, its window or your page closes and when the user logs out, accept only JSON of a sensible size, keep results as drafts until the user saves in the editor, answer key requests only for the logged-in user, and not send CORS headers on the endpoint (AIs call it from a server or a command line, not from a browser).

### Own window

    const win = window.open("https://md.giit.ch/?embed=window", "mdeditor-<id>");   // no "noopener"

The full editor (not the reduced embedded view) with a Save button, in a top-level window, so links from an AI reach it. It talks to `window.opener` exactly as an iframe talks to its parent: `mdeditor:ready` goes to the opener, messages are only taken from it, and once the page has sent a message, answers go only to its origin. Save sends `mdeditor:save` to the opener (if it is still open) and to the webhook from `load`. Closing the window with changes that were not saved sends `{type:"mdeditor:closing", id, name, content}`; if the opener is gone, the browser asks before closing. Calling `window.open` again with the same name reloads the window - focus the existing one instead. The site sends no Cross-Origin-Opener-Policy, so `window.opener` is kept.

## Tools for browser agents (WebMCP)

In browsers with WebMCP (`document.modelContext`), the editor page registers tools that run in the page itself: `get_document`, `get_problems`, `list_sections` (headings with their #anchors), `set_document` {markdown}, `insert_text` {markdown, where: "end" | "cursor" | a heading's title}, `replace_text` {find, replace, all}, `show_line`, `undo`, and `list_files`, `open_file`, `new_file` {name, markdown?} for the user's other documents. No tool deletes a document. Every change is an undo step for the user.

## Pages

- [Markdown Editor](https://md.giit.ch/)
- [Imprint](https://md.giit.ch/imprint.html)
- [Privacy policy](https://md.giit.ch/privacy.html)

A project by [Grünwald Innovation](https://giit.ch/). Sister projects: [LaTeX Editor](https://latex.giit.ch/), [SVG Editor](https://svg.giit.ch/).