Connect Claude, Codex or Cursor to your library
MindMax runs a small server on your computer that AI apps can call. Once an app is connected, it can search your documents, read whole papers and books, work on a canvas, and suggest edits to your writing — using its own model, inside its own window.
Nothing is uploaded to MindMax or to us. The connection is on your machine, between the app and the running MindMax window.
- The MindMax desktop app for macOS or Windows, with the folder you want to share opened as a workspace.
- An AI app that speaks the Model Context Protocol: Claude, Claude Code, Codex, ChatGPT desktop, Cursor, Windsurf or WorkBuddy.
Connect in one click
- Open the folder you want the app to reach, so it is the workspace in front.
- Go to Settings → External AI. MindMax lists the apps it found on this computer.
- Select Connect on one app, or Connect all. MindMax writes the app's own configuration
and keeps a
.bakcopy of the old one. - If the app needs a restart, MindMax offers the button. Claude Code and Codex pick up the connection in their next session.
- Start a conversation and ask for something — the status bar shows which app is reading your library while it works.
| App | What MindMax writes | When it takes effect |
|---|---|---|
| Claude | claude_desktop_config.json | After a restart (MindMax offers it) |
| Claude Code | claude mcp add, user scope | Next session |
| Codex | ~/.codex/config.toml | Next session |
| ChatGPT desktop | The same config.toml entry as Codex | After a restart |
| Cursor | ~/.cursor/mcp.json | After a restart |
| Windsurf | ~/.codeium/windsurf/mcp_config.json | After a restart |
| WorkBuddy | ~/.workbuddy/mcp.json | Immediately |
Every app is written under one name, mindmax-desktop. Disconnecting removes only that entry; your
other MCP servers are left alone.
Two access levels
Each connected app is set to one of these, from the ⋯ menu next to it in the same settings page.
| Level | The app can |
|---|---|
| Collaborate (default) | Read, search, propose edits, comment, create documents and canvases, run MindMax's methods with its own model |
| Read only | Read, search, and read your annotations — no proposals, no new files |
Neither level allows an app to overwrite an existing document or delete one.
What to say first
The connection is invisible until you use it, so start with something you actually wanted:
- “Using my MindMax library, what does the trial in these three papers actually show?”
- “Read the third chapter of the book I have open in MindMax and tell me what it argues.”
- “Write a two-page brief on this topic from my sources, with quotes I can check.”
- “Make the document I have open in MindMax more concise.”
- “Go through the comments I left on this draft and answer them — or fix what they point at.”
- “Comment on the chapter I have open, the way a referee would. Don't change anything yet.”
In the app's / menu, hosts that support prompts also list MindMax's own jobs: proofread a document, verify its claims against your library, handle the comments waiting on it, write something new from your sources.
What happens when an app writes
A new file is created directly, with a line in the workspace history naming the app that wrote it.
A change to a file that already exists arrives as a suggestion: MindMax opens a review card, the proposal shows as grouped tracked changes on your own text, and each group says why. You accept or reject it — and until you do, the file on disk is untouched.
Comments and highlights the app adds behave like your own: they sit on the passage, they carry the app's name, and MindMax's AI can answer them later. Which makes a comment a two-way channel — one you leave is a request the app can act on, and one it leaves waits for you. See working in the margin with an AI agent.
Where to look afterwards
- The status bar. While an app is using the library it says so by name. Open it to see the last calls, pause the app for a while, or disconnect.
- The sessions list. Each connected app has its own entry: what it read, what it proposed, and any long-running method it ran.
- Version history. Every write is attributed, so a document an app touched can be compared and rolled back like any other save.
When an app does not appear
It is not installed, or MindMax cannot see it. MindMax looks in the usual places for each app,
and asks your login shell whether claude or codex is on the PATH. If an app is installed
somewhere unusual it may be listed as not installed.
The app is connected but never used a tool. An app that has been configured but has not made a real call is shown as set up rather than connected — that is deliberate, because a written configuration is not proof that the app works. Make one call and the state updates.
ChatGPT does not see the tools. The ChatGPT desktop app shares Codex's configuration. Whether a given build surfaces local tools in a conversation varies; if it does not, Codex on the same machine uses the same entry.
The manual entry. If a configuration file cannot be updated — permissions, or a file that no
longer parses as JSON — MindMax shows the exact entry to merge, with a copy button. Paste it into
the mcpServers object of that app's configuration and restart the app.
“Open MindMax first”. A tool call starts the app in the background if it is closed, and waits. If it cannot start, the call returns a message asking you to open MindMax rather than hanging.
An absolute path outside the workspace. Only read_document accepts a file outside the open
workspace, and only for reading. Writing outside the workspace is refused; open that folder as a
workspace instead.
Privacy and limits
- The server listens on
127.0.0.1only, with a separate token per app. There is no remote endpoint, so a server or phone cannot reach your library. - MindMax runs none of its own models on this path. What the app sends to its model provider is governed by that provider's terms, as with anything else you do in that app.
- Calls are recorded in a local audit log for 30 days, and in the app's session entry.
- A connected app can read the workspaces you have opened recently and nothing else.
For the full list of tools and what each one may do, see the agent tool reference. For the reasoning behind the design — why there is no delete, why quotes are expanded by MindMax rather than trusted from the model — see MindMax MCP.