Image: GitHub
Why it mattersBuyers who use coding agents keep asking what a team at a working lab actually does. One of those teams now shows the answer in installable form.
A model maker is releasing more than the agent. MiniMax has opened up the engineering habits its own team uses around the agent. MiniMax released mcode-stack on 28 October 2026. The GitHub API reports 291 stars for MiniMax-AI/mcode-stack seven days later. Trendshift ranked the repository 14th on its daily list on 29 October. The license is MIT.
The README calls it the workflows and skills the MiniMax Code team uses every day. Each item is a SKILL.md file. An agent can load it from its own description. The catalog installs into MiniMax Code with mcode plugin install. The same catalog also installs into Claude Code with /plugin marketplace add MiniMax-AI/mcode-stack.
What is in the stack
The stack holds two authors' setups. The main one, stacks/ronny, has 15 skills grouped by moment in the day. "Plan and split" skills decide whether work goes to a subagent. They also split a feature into reviewable pull requests. "Context and tools" skills audit what gets injected into every request and what each tool costs in context. "Long sessions and commands" skills cover compaction, hand-off documents, and a ten-point audit of long shell commands. "Review" skills cover three jobs. Reviewing a diff from a clean context. Writing severity-labeled review comments. Handling reviewer feedback as the author.
A second author, stacks/jiachen, adds two skills. One is called lifting-loops. It finds the loops where a user keeps acting as a manual step for their agents and lifts them out. The other is making-content, for turning notes into image posts and short vertical videos.
How it fits into an existing team
The skills use the common SKILL.md format. The README says most of them work in any agent that supports skills. MiniMax Code reads project skills from .minimax/skills/, .claude/skills/ and .agents/skills/. It also reads global skills from the matching user folders under ~/. A team already using Claude Code with its own skills folder does not have to move anything.
The repository pairs each skill with a published article in MiniMax's "MCode Engineering Notes" column. The article explains the idea. The repository makes the idea runnable. A buyer of a coding agent gets a free copy of how a working lab sets up its own agent. They do not have to write those habits from scratch.
What a reader should check first
Three skills in stacks/ronny stand out for a team that is already running a coding agent. tool-budget-audit measures what the installed tools cost in context, and recommends trimming, tool search, or programmatic tool calling. context-tax-audit reads what gets injected into every request and what that does to the prompt cache. handoff-doc writes a document that lets a new session pick up where a long one stopped.
A team should also read independent-review. The skill asks the agent to review a diff from a clean context: only the diff, the requirement and the test results, with no earlier conversation. The authors claim that is how they catch issues the main session has already talked itself out of.
What is missing
The README names no measurement. There is no "we cut tokens by X percent" or "we caught Y more bugs". The authors say harness engineering is a craft that moves forward by experiment. A buyer should treat each skill as a hypothesis to test on their own repo, not a proven win.
Source
- MiniMax-AI/mcode-stack, GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
