Skip to content
Rush Commerce
Software & Dev2 min read

MiniMax MCode Stack: open-source agent skills for Claude Code

MiniMax open-sourced MCode Stack, 17 MIT-licensed SKILL.md agent skills that install in Claude Code and Codex. What's in it and which to copy.

MiniMax open-sourced MCode Stack, the agent skills its own coding team uses daily. The MiniMax-AI/mcode-stack repo holds 17 skills in the common SKILL.md format, two personal workflow files and five engineering notes. It installs as a plugin in Claude Code and Codex, not only in MiniMax Code. It contains no MiniMax Code source. It is a look at how a model lab's engineers actually run coding agents.

What actually happened

Per the repo README:

  • Two stacks. ronny has 15 skills; jiachen has 2 (lifting-loops, making-content).
  • The ronny skills are mostly process checks: subagent-split-check, tool-budget-audit, context-tax-audit, long-session-check, handoff-doc, long-command-check, independent-review, trajectory-review, plus engineering-* skills for small PRs, change descriptions, code review and emergency changes.
  • Five notes cover subagents, tool context, compaction vs. memory, background bash tasks and prefix caching.
  • Install: claude plugin install ronny --marketplace MiniMax-AI/mcode-stack for Claude Code, or codex plugin marketplace add MiniMax-AI/mcode-stack for Codex. You can also copy one skill folder into ~/.agents/skills/.
  • License: MIT, except the engineering-* skills, which adapt Google's Engineering Practices and stay under CC-BY 3.0.

There are no benchmarks or eval results. These are one team's habits, published.

Why open-source agent skills matter for your business

If your team uses Claude Code or Codex, the hard part is no longer getting the agent to write code. It is keeping long sessions on track: context fills up, the agent forgets a constraint from an hour ago, a background command hangs, and nobody can tell what it did. Look at the skill names again. Context tax, tool budget, long session, handoff doc. Those are the problems a lab's own engineers hit every day.

That is the useful signal. You do not need to adopt someone else's workflow. Copy the checks.

  1. Install one skill, not the stack. handoff-doc or long-session-check is a good first pick. Run it for a week.
  2. Read every SKILL.md before you install. A skill is instructions your agent will follow. Treat it like a dependency.
  3. Keep the attribution. The engineering-* skills are CC-BY 3.0. If you fork them into a client repo, credit the source.
  4. Write your own versions. The real value is your team's rules (your branch naming, your deploy steps) in a file the agent reads every time.

Key takeaways

  • MCode Stack: 17 SKILL.md skills, 2 workflows, 5 engineering notes
  • Installs as a plugin in Claude Code and Codex, not only MiniMax Code
  • MIT licensed, except engineering-* skills under CC-BY 3.0
  • No benchmarks: these are one team's habits, not tested results
  • Read each skill before install; start with one

Your team uses coding agents but every session goes differently? We write the skills, config files and review checks that make agent work repeatable, in your repo. See what we set up or look at what we've shipped.

Sources: MiniMax-AI/mcode-stack (GitHub).

  • #minimax
  • #agent-skills
  • #claude-code
  • #codex
  • #coding-agents
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.

Keep reading

Software & Dev3 min read

TanStack AI 1.0: Stable Core, 23 Providers, Experimental Harness

TanStack AI 1.0 ships 23 provider adapters and a stable core, but the agent harness is still 0.x. Build on the core, keep the harness behind a wall.

Read it
Software & Dev3 min read

shopify app security check: put Shopify's app scanner in CI

Shopify's app security check CLI scans for unverified webhooks, cross-shop data access and leaked keys. It never fails by default. Run it in CI with --blocking.

Read it