Write faster: your comments become your AI's instructions

You annotate a passage in the editor, the way you would for a proofreader. Your AI reads that comment through Sarmate's MCP server — with the exact excerpt and line it is anchored to — and edits the source. Nothing to copy, nothing to paste back, nothing to explain twice.

Connect your AI Try the editor

Annotate, your AI applies

You select, you comment

In the editor, highlight a passage and write what you want: “too easy”, “I need two questions using letters”, “the fractions touch the table rules”.

Your AI reads your comments

It calls read_comments on the MCP server. With no path, the tool says which files carry comments; with a path, it returns the full threads.

Your AI edits the source through the MCP server

It knows which passage to rework, because the comment quotes it. It edits the file, compiles, and you read the PDF.

The comment carries its own position

That is the whole difference with a chat. In an ordinary exchange you have to describe where to act — “the third exercise on page 12, the one about percentages” — and hope it lands. A comment is already attached to the right place: it carries the quoted excerpt and its line range.

And it stays attached when the text moves. If your AI rewrites the passage, changes the indentation or re-wraps the lines, the comment re-anchors on what it quotes rather than on a line number that has gone stale.

What your AI actually receives

Each comment arrives with the excerpt it is attached to, its line range, its author and any replies — enough to act without asking anything further.

// The AI asks for a file's comments
read_comments({ path: "Livres/EAM_1spe/p2_automatismes.tex" })

// What it gets back
{
  "path": "Livres/EAM_1spe/p2_automatismes.tex",
  "count": 3,
  "comments": [
    { "quote": "\item Soit $f(x) = 2x^{2} - 3x + 1$. ...",
      "line_start": 476, "line_end": 476,
      "author": "You",
      "body": "too easy",
      "replies": [] }
  ]
}

The tool is read-only and limited to your own drive. It never creates, edits or resolves a comment: it hands them over to be read. What changes the document are the editing tools, which you control separately.

A real session

On a maths textbook being written, six annotations left in four minutes, then a single call from the AI to read them all:

07:37  line 362  “the fractions touch the table rules”
07:38  line 380  “I need two questions using letters”
07:38  line 398  “slightly harder questions”
07:40  line 474  “Too easy. Give the area, ask for the radius”
07:41  read_comments  →  6 comments

None of these sentences says where to act. None needs to: the comment is already in the right place.

Getting started

Create an MCP token from your account, then connect your AI — Claude, Gemini, ChatGPT or Le Chat. The step-by-step tutorial covers each client.
Open your document in the editor, select a passage, click “Comment”.
Ask your AI to read the file's comments and apply what they ask for.

The MCP server is open on every plan, free included. Review comments are written in the editor, by you and by your collaborators.

What the tool will not do

It only reads comments attached to your files. Those you wrote on someone else's document belong to that person's drive.
By default it returns only unresolved comments. The ones you marked as handled do not come back to clutter the work.
It does not mark anything as resolved for you. You close a thread yourself, once you have seen the result.