Works with your AI app
MindMax MCP: your library and tools inside the AI app you already use
MindMax works for you — and for your AI, drawing on the same material and the same method. Claude Code, Codex, ChatGPT and Cursor can reach the folder MindMax has open: the PDFs, the books, the notes, the highlights. Your app brings the model; MindMax brings the tools, the shelf, the reading map and the review gate in front of anything it writes. The model calls those tools, and the work and the collaboration are the better for it.
- Connects
- One click — no config file to edit
- Costs
- Nothing in MindMax — your app does the thinking
- Runs on
- macOS (Apple silicon) · Windows
- Price
- Free app · AI plans from $10/mo, or your own key
- AI
- Included, or your own API key
An agent that can read a file is not an agent that has read your library. Pointed at a folder it opens what it can find, reads what fits in its context, and cites what it can copy — which on a shelf of papers and books means a lot of guessing about structure, and no memory of what you marked last month. The usual fix is to upload the files somewhere. This goes the other way: the library stays here, and the agent comes to it.
It reads the shelf, not just the file you happened to name
The first thing an outside agent lacks is not intelligence. It is the map.
- Search runs over every format you have opened in MindMax — PDF, EPUB and Kindle, Word, PowerPoint, Excel, images, Markdown — and each hit comes back with the section it sits in and a handle you can cite.
- Ask for a long book and the agent gets a map, not a wall of text: two levels of structure, a one-line note per chapter, and a section id for each. It picks a chapter, reads that section whole, and can go back for another.
- A question can be passed along with the map as a lexical hint, so the plausible chapters are marked before the model chooses. The choosing stays with the model — MindMax never picks the chapter for it.
- Figures, tables and scanned pages can come back as page images, so a chart is looked at rather than guessed at from its caption.
It can think on your canvas, with the methods in your library
- The card you have selected, the canvas in front, the document open behind it: one call hands the agent what you are looking at, so “make this branch sharper” needs no pasting.
- MindMax’s named methods — first principles, a decision matrix, red-teaming, SWOT — are readable one at a time, with the card shape each one produces, and an outside agent can apply one onto a canvas instead of inventing a structure on the spot.
- Longer methods run as pipelines: MindMax plans the steps and validates the output, the agent’s own model does the writing, and the result lands where MindMax’s own run would have — the reading panel, the comment threads, the translated file, the recognised pages.
You point at the sentence. The work happens on it.
An instruction about one sentence should live on that sentence — not three hundred words away in a chat window.
- Comment on the line that is wrong, in one line: “this repeats the paragraph above”, “this needs the 2024 figure”. The agent reads the comment together with the passage it is attached to, so nothing is pasted and no paragraph is counted.
- A comment can also be a request with no words — mark a passage and set it to rewrite, verify or find a source. When the edit is proposed it answers in that thread, so a pass ends with every comment answered, changed, or still open with a reason.
- The channel runs both ways. Ask for a review rather than a rewrite and the notes arrive in the margin, addressed to you; reply to one and the agent picks the work up from your reply rather than starting a new conversation.
Its edits arrive as suggestions, not as a fait accompli
The second thing an outside agent lacks is restraint. This is where that is enforced.
- A new file can be created outright. A change to something you already wrote — one word or a complete rewrite — arrives as a proposal, shown on your own text as grouped tracked changes, each with its reason. Nothing is written until you accept it.
- A proposal that answers a comment says which one: the comment gets a reply, and accepting the change marks it done. Reject it and the thread stays where it was, so nothing closes without your say-so.
- There is no overwrite and no delete in the tool set. Renaming and moving go through MindMax so history, comments and open tabs follow the file. Every write is attributed in version history to the app that made it.
The rest of the connection
Read-only, or collaborate
Each connected app is set to one of two levels: read and search, or read plus suggestions, comments and new files. Collaborate is the default because a library you cannot write to is half a tool.
You can see it working
While an app is using the library, the status bar says so by name. Open it to see the last calls, pause the app, or disconnect it — without hunting for a settings file.
A session you can read afterwards
Every connected app gets its own entry in the sessions list, with each call it made, each proposal it raised and each pipeline it ran. Not an audit log you have to decode — the same list your own conversations are in.
Seven apps, one entry each
Claude, Claude Code, Codex, ChatGPT, Cursor, Windsurf and WorkBuddy are detected on this machine and written into their own configuration, with the old file backed up first.
The folder explains itself
A guide file in the workspace tells a coding agent the formats used here, and that new documents and changes should go through MindMax. On by default, and only while an app is connected.
Ready-made jobs in the “/” menu
Hosts that support prompts list MindMax’s own methods: proofread a document, verify its claims against your library, handle the comments waiting on it, write something new from your sources.
Diagrams checked before they land
A diagram is drawn with the same renderer the editor uses before it goes into a document, so a broken block comes back as an error to fix rather than as a grey box in your notes.
Delivered as Word or PDF
The agent can export through MindMax’s own exporter, keeping tables, editable equations and numbered sources — the same file the Export menu produces.
The older connectors still workThe web connectors that hand a document or a mind map over from claude.ai or ChatGPT into the online editor are unchanged, and still need no desktop app. This page is the other direction: a connected app working against your folder. Mind maps from Claude and ChatGPT covers those.
Connecting takes a click, not a configuration file
The app writes each host’s own configuration, keeps a backup, and offers to restart the app for you. If a host cannot be updated automatically, it shows the one line to paste — that is the only case where you touch a file.
Open the folder in MindMax
The library an outside app can reach is the workspace you have open — or the most recent one, if MindMax is in the background.
Go to Settings → External AI
MindMax has already looked for the apps installed on this computer and lists each one with its state.
Press Connect
Every app found is connected at once. No permission dialog, no workspace picker: collaborate access, all recent workspaces.
Restart the app if it asks
Claude, ChatGPT, Cursor and Windsurf need a restart, and MindMax offers the button. Claude Code and Codex pick it up in their next session.
Ask for something ordinary
“Using my MindMax library, what does the trial in these papers actually show?” The status bar shows which app is reading, and the answer arrives with quotes that open on the page.
Coming from pasting files into a chat window: what changes
Most people arrive here already handing documents to an assistant. Here is what moving the work to your own library changes — including the rows where the chat window still wins.
| Topic | A chat window with uploadsDrag a file in, ask, start again next week | In MindMax |
|---|---|---|
| What the assistant can see | The files you uploaded, one conversation at a time. | Every file in the workspace, in the formats MindMax reads, searched or read directly — a folder of papers, a shelf of books, last month’s highlights. |
| A book-length document | Whatever survives the upload and the context limit. Long books are usually shortened, sampled or refused. | A map of the whole book, then whole sections read on demand. The agent navigates instead of skimming. |
| Checking an answer | A quote you hope was copied correctly. | Quote handles that MindMax expands into the verbatim passage with a clickable anchor — checked against the text before it lands. |
| What you marked last month | Gone with the conversation. | Highlights, comments and bookmarks are readable by the agent, including the meaning you gave each highlight colour. |
| Saying where to change something | Describing the place in words — “the third paragraph”, “the bit about the trial” — and pasting the passage in to be sure. | A comment on those words. The agent edits that line, replies in the thread, and the next round starts from the reply. |
| Changes to your writing | A rewritten block to copy back yourself. | Tracked changes on your own text, grouped with reasons, accepted or rejected one by one. |
| What it costs | A subscription that covers the model and the storage. | The model is the one you already pay for, in the app you already use. MindMax charges nothing for the connection and runs none of its own models on this path. |
| A one-off question on a phone | Open a tab, upload, ask. Hard to beat. | The connection is to a desktop app on this machine. For a single file on a phone, MindMax Read in the browser is the shorter path. |
| Working from a borrowed computer | Nothing to install. | The library lives on your machine, so this is a desktop arrangement by design — there is no cloud copy of your folder to log into. |
Where it gets hard
- Support is not identical across apps
Claude Code, Codex and Claude are the ones verified end to end. Cursor, Windsurf and WorkBuddy are written and detected, and show as connected after their first real call. The ChatGPT desktop app shares Codex’s configuration, but whether it surfaces local tools depends on its build — MindMax tells you when an app has been set up but never called a tool, rather than claiming it works.
- It does not make the model smarter
The thinking is done by whichever model your app runs. A weak model with a good map is still a weak model; what changes is the material it works from, the tools it can call, and what it is allowed to do to your files.
- Reading a whole book is not free — to your app
MindMax spends nothing, but the pages the agent reads are tokens in its own context. A full reading-guide pass over a book costs roughly one pass of the book, and the tool says so before it starts.
- No overwrite, no delete, by design
An outside app cannot replace a document wholesale or remove one. That is deliberate: a tool set with a delete button is one prompt injection away from an empty folder. Deleting is yours to do in your file manager.
- Local only, for now
The connection listens on 127.0.0.1 and requires MindMax to be running — it will start it in the background if it is not. There is no remote endpoint, so an agent running on a server cannot reach this library.
- Scanned pages are still scans
Page images let the agent look at a page it cannot read, and recognition can be run with the agent’s own model, but quality depends on the scan. MindMax reports which pages came out badly instead of quoting them as if they were clean.
Common questions
What is MindMax MCP?
MindMax MCP is a local Model Context Protocol server inside the MindMax desktop app. It lets an AI app on the same computer — Claude Code, Codex, Claude, ChatGPT, Cursor, Windsurf or WorkBuddy — search your documents, read whole papers and books, work on canvases, and propose edits to your writing. The app uses its own model; MindMax supplies the tools, the library and the review step — and it is by calling those tools that the work and the collaboration come out better.
Does connecting an AI app upload my documents?
No. The connection is a server on 127.0.0.1 inside the app, and the calling app talks to it on the same machine. Your files are not copied to MindMax or to us. What leaves the computer is what the app sends to its own model provider, exactly as it would for anything else in that app.
Does this use MindMax credits?
No. Nothing on this path calls a MindMax model, so no credits are spent. The reading, writing and reasoning are done by the model your AI app is signed in to, and billed there.
Which AI apps can connect?
Claude, Claude Code, Codex, ChatGPT desktop, Cursor, Windsurf and WorkBuddy. MindMax detects the ones installed on your machine and writes their configuration in one click; Claude Code, Codex and Claude are the ones verified end to end so far.
Can a connected app change or delete my documents?
It can create new documents, append to existing ones, move or rename files, and propose changes to what you wrote. Those proposals arrive as tracked changes you accept or reject. It cannot overwrite a document outright or delete one, and every write it makes is attributed to it in version history.
Can I just comment on the sentence I want changed?
Yes, and that is the shortest path. Comment on the sentence, and the agent reads the comment with the passage it is attached to — no paragraph to count, no text to paste. It proposes the change on that line and replies in the thread; accept it and the comment is marked done. The same channel works the other way: ask for a review and its notes wait for you in the margin.
Can I use my own model, or a local one?
The model belongs to the app you connect. If that app runs a local model or your own API key, the whole path stays on your machine — MindMax never substitutes a model of its own.
Does it work when MindMax is closed?
Yes. The first tool call starts MindMax in the background, without taking focus, and waits for it. If the app cannot be started, the call comes back telling you to open MindMax rather than failing silently.
Can I connect from a server or a phone?
Not yet. The server binds to the loopback address of the machine MindMax runs on, so remote agents and mobile apps cannot reach it. For a single document in a browser there are the online readers; for a whole folder the desktop app is the only path.
How is this different from the MindMax connectors in claude.ai?
Those connectors hand one document or one mind map over to the online editor through a link, and store nothing. This is the reverse and the larger half: an app on your computer working against your actual folder, over many turns, with your existing highlights, comments and files.
What does it do to a long book?
It returns a map — two levels of structure, a note per chapter and a section id for each — then reads the sections the model chooses, one at a time, whole. A 550-page book is about 12k characters of map, which is enough for an agent to navigate it without reading every page.
See also
The one-click path, the two access levels, and what to do when an app does not show up.
Every tool by group and access level, from search_knowledge to export_document, plus the methods it can run itself.
A worked path: a question across a shelf, a book read by its map, a cited draft, a reviewed edit.
A comment on the sentence that is wrong, the change on that sentence, and a thread for the next round.
What the agent is writing on when it adds a card, and what those cards can carry afterwards.
Why a change to your writing arrives marked, and what accepting it actually writes.
The standard behind the connection, in plain terms, and where it goes wrong.