v1.4.1 — a complete guide to working with your own AI agent

This entry is about the website, not a new desktop build. The app stays at v1.4.1, where the local MCP channel, canvas requests and comment threads already work. What is new is that every one of them now has a guide, a worked use case and a Chinese version — so a person who has connected Claude Code, Codex or ChatGPT desktop can find out what to ask for and where the answer lands.

What is in it

• One capability map. Working with an external AI agent is the hub: three directions into MindMax, the two method catalogues, the ten document prompts, the seven long-job methods, and where each kind of result lands for review. It is linked from the MCP page and from the homepage.
• Seven operation guides. Handing work over, revising documents, reading and reviewing, canvas requests, long jobs and controlling access sit behind the hub. Canvas requests and external-model long jobs had no guide at all before this update.
• Three worked use cases. Review a paper with your own AI agent, brainstorm a product with your own AI agent and run a long job with your own AI agent now appear on the use-case index, in order of how much they ask of you.
• A corrected tool reference. The tools an agent can call page now matches the code: show_canvas, page images and regions on PDFs, read_canvas with since, canvas request annotations, thread replies, fromCanvas and the full prompt list. The two connection levels are separated from the three tool kinds.
• Corrected host facts. Claude, Claude Code, Codex and ChatGPT desktop are verified end to end. The other three hosts 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.
• A troubleshooting section and AI-engine indexes. Common issues now lists the symptoms of this path — a copied prompt that was never sent, a suggestion that never reached a review card, a job that stopped halfway. llms.txt and llms-full.txt carry the same facts in a form search and answer engines can quote.
• Chinese mirrors for all of it. Every new guide, use case, the corrected connection and tool pages, and this changelog are available in Chinese with explicit heading anchors.

Known rough edges

• Four hosts, not seven, are verified end to end. Cursor, Windsurf and WorkBuddy are connectable, but their status is only confirmed on the first real tool call.
• Copying a prompt is not sending it. MindMax does not wake an external app, and a canvas request stays waiting until you paste the line into the app and press send there.
• This channel is local and desktop-only. The server binds to 127.0.0.1, there is no remote endpoint, and no mobile or web version of this arrangement. Nothing is uploaded and no MindMax credits are spent, but the connected app's own token bill still applies.
• Some formats stay read-only. PDFs take comment replies rather than edits to their text, and PowerPoint files are read as text rather than as page images on this path.
• Three real screenshots are still outstanding. The MCP page keeps its existing illustrations in those slots rather than substituting a mock-up for a screenshot.

Where to start

Open Working with an external AI agent for the map, then connect your AI app if it is not connected yet. MindMax MCP is the shorter product overview. If you want one task rather than a map, Mark one sentence, hand over one line, review once walks through the three moves.