MiniMax Code Mini Apps: local workflows and permission limits
MiniMax Code Mini Apps turn an agent conversation into a persistent workspace with a page, a local backend and tools the agent can call. They are useful when a dashboard needs to keep its data and actions together, rather than ending as a static file. The important limitation is where that software runs: on your computer, with your user account’s permissions.
This is a practical guide, not a new launch report. MiniMax published its Mini App overview on October 22, 2026, but its desktop changelog already listed custom Mini Apps on September 8.
What runs where
According to the current Mini App documentation, the Node service can retain data across restarts and updates, expose business functions through MCP and hand work to background agent sessions. A task board connected to Notion or Jira is one documented example. That is a vendor example, not a workflow we tested.
| Access question | Documented answer |
|---|---|
| Where can a Mini App start? | Local sessions in the macOS or Windows desktop app. |
| What about cloud, SSH or Remote Control? | Those sessions do not start Mini Apps on your machine. |
| What does the desktop app require? | macOS 12 or later, or Windows 10 or later, plus MiniMax account sign-in. |
The operating-system requirements come from MiniMax’s download documentation. Local execution does not mean every connected service or model works offline.
Building and updating a workspace
The documented creation workflow starts with /miniapp-creator and a description of the desired app. Before building, the agent checks each data source with a read-only query. If that fails, the prescribed workflow stops rather than presenting sample data as a working integration.
It then creates the package, writes the page and Node backend, publishes, inspects and opens the result. Screenshot checks are part of MiniMax’s instructions, not independent evidence that every generated app works. Review the result before relying on it.
Later changes use the same package. Official marketplace Mini Apps cannot be edited and republished locally; customization requires a copy with a new plugin ID.
Host permissions are not a sandbox
MiniMax’s security model explicitly says the Node process has no operating-system sandbox. It can access files, start programs and make network requests as your user. Install only trusted packages; a restricted host API does not isolate malicious plugin code.
Within those host APIs, running an app’s agent sessions and reading your visible sessions are separate grants: session.run does not imply session.read. Declared permissions can be disabled individually. Host Connectors keep connected-account credentials out of plugin and page code.
The message.append operation requires confirmation and changes only an unsent input draft. You still send it yourself. These are documented host boundaries, not a guarantee about the safety of arbitrary Mini App code.
Check usage before assigning ongoing work
MiniMax’s M Plan documentation lists Go, Explore and Build with relative quotas of 1×, 3× and 7.5×. Consumption depends on the model, context and task complexity. Check the current account plan and usage display before assigning repeated agent work; a persistent interface is not a promise of unlimited model usage.
AI-generated editorial illustration by ByteForward. MiniMax logo: MiniMax. Not a product interface.
