MiniMax MCode Stack: Open Agent Skills, Setup and Marketplace Rollout

MiniMax has published MCode Stack, a repository of skills and workflows used by its own coding team. Developers can reuse them to manage agent context and review code changes. A plugin marketplace arrived on October 28, although support for adding marketplaces in MiniMax Code desktop is rolling out gradually.

The repository contains workflow instructions, not the source code of MiniMax Code or a new model. That makes it useful to teams trying to improve how an existing coding agent works without changing the model underneath it.

What MiniMax released on October 28

The initial repository commit is timestamped October 28 at 02:17:49 UTC. The plugin-marketplace merge followed at 11:17:48 UTC, and another contributor stack merged at 11:21:49 UTC. Those are separate changes within the same day's release history.

The README at that latest revision lists 15 skills in Ronny's stack. They cover decisions such as whether to split work across agents, how to handle a long-running command, and how to preserve useful state when a session gets large. The collection also includes engineering review and experiment checklists.

How to install MCode Stack

Individual skills can be copied into ~/.agents/skills/ and checked through /skills in a new session.

For a whole stack, MiniMax Code desktop supports adding MiniMax-AI/mcode-stack through Plugins and Add marketplace. That interface is still rolling out, so a missing option does not establish an installation error.

Claude Code users can add the same catalog with /plugin marketplace add MiniMax-AI/mcode-stack, followed by /plugin install ronny@mcode-stack. Check each workflow against your agent's available tools and permissions before using it.

A concrete example: independent code review

The independent-review skill starts a reviewer with the diff, the requirement and actual test results. It deliberately leaves out the implementing agent's conversation and self-assessment.

That boundary is useful because a reviewer can otherwise inherit the assumptions that produced the change. The skill asks for findings tied to a location and a reason, then requires the main agent to verify each finding before applying a fix. It does not make the review automatically correct.

Imagine a small API change whose tests all pass, but whose requirement includes rejecting a particular invalid input. A reviewer needs that requirement to notice the missing check. A longer account of how the implementation was written would not supply it. This is an illustrative use case, not a test result from this release.

Measure whether the workflow helps

The harness-experiment skill recommends changing one part of an agent setup at a time and repeating the same evaluation tasks. It tracks individual regressions alongside requests, tokens, cache hits and elapsed time. A higher average score can hide a task that now fails.

This is a sensible way to approach the collection. Start with a recurring problem in your own workflow, add the relevant skill, and compare results before installing everything. Extra instructions can consume context or conflict with existing rules.

The repository is MIT-licensed, with third-party notices for six engineering skills adapted from Google Engineering Practices under CC-BY 3.0. Preserve those notices when reusing the material. The release offers inspectable working habits; whether they improve your agent remains something to measure on your tasks.

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

I help technical founders explain developer tools and agent workflows in content that developers can understand, use and share.

Book a 30-minute call.