MiniMax Code Mini Apps: Persistent Workspaces for AI Agents

MiniMax launched Mini Apps in the MiniMax Code desktop app on October 22. They combine a persistent backend, an interactive page and two-way communication with the agent. Instead of returning to a chat to reconstruct an ongoing task, users can keep its data and controls in a working application.

The distinction is persistence. A generated dashboard can be useful on the day it is created. A dashboard that keeps fetching data, saves decisions and lets the agent act on them is a different kind of tool.

What MiniMax launched

The announcement includes the Mini App runtime, the /miniapp-creator creation workflow, official marketplace apps and an open community repository.

A Mini App is packaged as a plugin, with a miniapp.json manifest, a frontend and a Node.js backend. The backend can expose its capabilities as MCP tools, so the agent can use the same application that the person sees. The app can also ask the agent to work on a task.

MiniMax Code handles processes, ports and credentials, leaving developers to build the interface and application logic.

The creation workflow starts by checking whether each data source can actually be read. It then builds and publishes the app and checks the result in the built-in browser. That is the vendor's described workflow, not evidence that every generated app will be correct.

A live workspace rather than another chat artifact

The runtime overview explains what survives: a Mini App's data directory persists across restarts and updates, while its Node service starts on demand. This lets the interface represent continuing work rather than a frozen response.

Consider a repository review. A chat can explain a branch once, but a persistent view can keep the selected repository, filters and changes visible while the user asks for another investigation. The important state stays attached to the work instead of being repeatedly described in prompts.

The same separation helps with human control. The page can expose precise selections and structured information that would be awkward to express in a message. The agent remains useful for interpretation and changes, while the interface handles inspection.

MiniMax's documentation says page content sent to a conversation first appears in the user's input box for review and editing. The page is not supposed to press Send on the user's behalf.

Availability and installation limits

Mini Apps are available in local sessions of MiniMax Code on macOS and Windows. Cloud, SSH and Remote Control sessions do not start these apps on the user's machine. A desktop feature should not be mistaken for a generally hosted web-app platform.

The community repository already includes a Git history viewer, token-usage dashboards, a CadQuery model viewer and MiniMax Bot. These are concrete examples of the interface-plus-agent pattern, rather than only a proposed extension format.

Compatibility is checked per app. Apps using newer capabilities, including background agent tasks or separate windows, require MiniMax Code 3.1.2 or later. The repository also warns that its packages need manual installation; the client's GitHub plugin-import feature does not support these Mini App packages.

The repository has an MIT license, but individual apps have their own included licenses. Neither that repository license nor the launch establishes that every model call or connected service is free. The announcement does not provide a separate Mini App pricing table.

Local execution is an important boundary

MiniMax says host-held account credentials are not exposed to the app's page, and connectors and agent capabilities must be declared before use. Those controls help separate an application's needs from the host's authority.

They do not turn arbitrary plugin code into harmless content. Mini App code runs locally as the user. Before installing a community package, review what it accesses and whether its source is trustworthy, just as you would with another desktop extension.

For teams building internal tools, the useful question is which part of the job deserves a stable interface. A recurring review queue, model viewer or operational dashboard may benefit. A one-off answer probably does not need an application around it.

Mini Apps make that choice easier to implement. The lasting value will come from keeping the right state and controls visible, rather than generating a new interface for every conversation.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain product architecture and launches in content that developers can understand, use and share.

Book a 30-minute call.