MiniMax Open-Sources MCode Stack: 15 Agent Skills Its Coding Team Uses Daily

MiniMax released the MCode Stack on October 28, 2026. This is an open-source collection of agent skills and workflows used by the team behind MiniMax Code, published under the MIT license.

What is the MCode Stack?

The release contains 17 distinct skills, five checklists, and five companion notes. It does not include source code for MiniMax Code itself. Instead, it provides plain Markdown files that install directly into the terminal agent. These files translate engineering habits into steps an agent can execute.

The repository pairs with a column of MCode Engineering Notes. Each note explains a specific concept, such as subagents, tool context, or memory compaction, and turns that idea into a runnable check. A note regarding context tax is marked as upcoming.

Skills and checklists

The main stack, attributed to Ronny, groups 15 skills by stage of the working day. Jiachen contributed two additional skills.

• Plan and split (3): subagent-split-check, engineering-small-prs, engineering-change-descriptions.
• Context and tools (2): tool-budget-audit, context-tax-audit.
• Long sessions and commands (3): long-session-check, handoff-doc, long-command-check.
• Review (4): independent-review, engineering-code-review, engineering-review-comments, engineering-review-feedback.
• Experiments and hotfixes (3): harness-experiment, trajectory-review, engineering-emergency-changes.

Several specific implementations stand out. The independent-review skill gives a reviewer exactly three inputs: the diff, the requirement, and test results. It deliberately withholds the author’s conversation history. The logic is that a reviewer inheriting the author’s context repeats the author’s blind spots.

The subagent-split-check instructs an agent to split work only when context size is the measured problem. Splitting by team or question type is not a valid reason.

The tool-budget-audit separates two costs: tool definitions and tool results. Tool search mainly cuts the first cost, while programmatic tool calling mainly cuts the second.

The harness-experiment skill enforces one change per experiment. It tracks flipped tasks, requests per task, tokens, cache hit rate, and wall-clock time.

The long-command checklist contains ten items. It asks whether moving a command to background re-runs the process. It also asks whether unread results survive compaction.

The six engineering-* skills are adapted from Google Engineering Practices. That material remains under CC-BY 3.0, per the stack’s NOTICE file.

Installation

There are three paths to install the skills, according to the README:

1. Clone the repo and copy a skill folder into ~/.agents/skills/.
2. Copy stacks/ronny into MiniMax Code’s local plugin directory to use the whole stack as a plugin.
3. Add the MiniMax-AI/mcode-stack marketplace in MiniMax Code desktop. This option is rolling out gradually.

The same catalog works in Claude Code with /plugin marketplace add MiniMax-AI/mcode-stack. MiniMax Code also reads ~/.claude/skills/ and ~/.codex/skills/. The skills use the common SKILL.md format, so they work in other agents that support skills.

Comparison with other packs

The MCode Stack is narrower than both Anthropic’s skills and obra/superpowers. It targets people who design agent harnesses, not only people who use agents. Superpowers covers far more harnesses today.

What it means

Developers can treat this release as a playbook rather than proof. There are no published evaluation results. The skills are plain Markdown files that install directly into the agent environment. For people building agents, this offers a concrete set of checks to enforce during planning, execution, and review.