Skip to main content

Mark one sentence, hand over one line, review once

· 7 min read
The MindMax team
Building MindMax

Mark the exact sentence, copy the short line MindMax gives you, paste it into your connected app, and review the change that comes back. That is the whole loop: one mark, one line, one decision. The app reads the mark in place through MindMax MCP; MindMax writes nothing until you accept.

The awkward part was never the wording of the request. It is that a chat instruction and the sentence it is about live in two windows, so every round starts by finding the place again. A mark removes that: the position, the kind and the instruction travel together. What is left to get right is the hand-off — and the fact that copying a line is not the same as sending it.

What do I mark in MindMax?​

Mark the smallest thing you can point at. A sentence, a number, one card — not the whole document.

  1. In an editable document, select the words and add a comment with one short instruction: "This figure is from 2022; use the 2024 one." Pick a kind — note, question, rewrite, doubtful, needs a source — when the words alone are too thin. Working in the margin with an AI agent explains what a comment carries and whose turn it starts.
  2. On a PDF or an ebook, comment the passage. The text stays read-only; the answer lands in the thread. When you want an editable copy, MindMax can put a Markdown file beside the original. Reading and reviewing with an agent covers page images, long books and review lenses.
  3. On the canvas, mark the card you want continued or corrected and use its request action. The card is recorded as a canvas request, with the prompt you will hand over. Canvas requests with an agent has the full life cycle.
  4. Leave the rest alone. Folders you do not want read can be hidden from AI in Settings; controlling agent access lists the two connection levels and the audit trail.
  5. Check the permission before you start. read only shows the app only the reading tools. For comments, proposals or new cards you need collaborate, set per app in Settings → External AI.

How do I hand the mark to my AI app?​

You send the line yourself. MindMax does not wake an app up, does not run a queue and does not assign work.

  1. Copy the prompt from the comments panel, the canvas card or the request bubble. One sentence is enough — the mark already says where.

  2. Paste it into your app and press send there. A workable shape:

    Read the comments and canvas requests I marked in the workspace I have open in MindMax. Answer each one in place, propose the wording changes, and leave every other paragraph alone.

  3. Say what should come back: notes in the margin, proposed edits, new cards, or a short answer in the thread. Handing work to your own external AI agent is the short guide to scope and hand-back.

  4. If the app shows no tools at all, check the connection page again. Claude, Claude Code, Codex and ChatGPT desktop are verified end to end; Cursor, Windsurf and WorkBuddy are written and detected, and show as connected after their first real tool call. ChatGPT desktop shares Codex's configuration and needs a restart after connecting.

Copying is not submitting. Paste and press send in the app; until then nothing has happened and no model has been called. This is the single most common reason a canvas request stays "waiting".

Where does the result land?​

Every change to an existing document arrives as a proposal, not a replacement.

  • Editable documents. The app proposes edits and comments. They appear on a review card as tracked changes, grouped with reasons, so you accept or reject a group at a time. Nothing is overwritten and nothing is deleted. Revising documents with an agent covers withdrawals, replies and what cannot be changed.
  • PDFs and ebooks. Answers go into the comment thread. The source file does not change; an adjacent editable Markdown copy is created when there is something to write into.
  • Canvas. New cards appear under the card you marked. The app may revise only the cards it added, and the canvas can be brought to the front so you can see the result in MindMax.
  • New documents. A file that did not exist before can be created directly, since there is no earlier text to protect.
  • The record. The status bar activity capsule, the external session, version history with the app's name and a local audit kept for 30 days all stay in MindMax.
  • The bill. The server binds to 127.0.0.1 only. Nothing is uploaded, no MindMax model runs, no MindMax credits are spent; the tokens are on the connected app's bill.
The mark, the reply and the proposed edit end up on the same passage, so the next round starts where the last one stopped.

Where it goes wrong​

  • The prompt was copied, never sent. A request waiting in MindMax is a sentence in your clipboard until you press send in the app.
  • "Make this better." A vague instruction gets a confident guess. Name the problem: a repeated sentence, a stale figure, a claim with no source.
  • read only and then waiting for a proposal. Proposals belong to the collaborate level.
  • The app wrote the file itself. If it edits the file directly instead of through MindMax's tools, there is no review card to accept. Ask it to use the tools.
  • Expecting a PDF's text to change. Read-only formats take comments and answers, not edits.
  • A long answer in chat, nothing in MindMax. Chat text is not a tool call; the result has to be written back through the workspace.
  • Accepting a pass you only skimmed. Forty comments is forty decisions. A skimmed pass is a pass you did not really make.
  • Assuming the app's model is free of charge here. MindMax charges nothing on this path, but the app's own token bill still applies.

What is not claimed​

MindMax does not fetch requests on its own, does not run a remote endpoint and does not offer this desktop local-MCP arrangement on mobile or in the browser. The app's model is not checked by MindMax for correctness; a proposal is a suggestion, and you are the one who decides. Seven hosts are supported, but only four are verified end to end. And the browser hand-off connectors are a separate, one-way direction: they import one finished document or mind map, and do not read your folder. The capability map draws all three directions apart.


Working with an external AI agent · MindMax MCP · Download MindMax

Cite this page