Skip to main content

Concepts

The whole model fits in one picture:

Team  (a named workflow)
 └─ steps  (ordered)
     ├─ Member   → a role + what it does + [skills] + optional model
     └─ Loop     → until <condition> · max_rounds → steps → Members
Skill  (a reusable capability, referenced by a member; lives in a shared library)
Team lead  (implicit, on top of every run: briefs the team, synthesizes the result)

Team​

A team is an ordered list of steps, saved as <name>.team.yaml. You name it and reuse it on any task. Teams live in two scopes:

  • global (default): ~/.my-mini-team/teams/ — available anywhere.
  • local (mmt new --local): ./teams/ — belongs to a project. A local team shadows a global one of the same name.

Step: member or loop​

A step is either a single member or a loop of members.

Member​

A member is a role with a job (does), the skills it plugs in, and optionally a model (opus/sonnet/haiku). At run time each member executes as a fresh sub-agent that receives the task, the team lead's brief, and everything earlier members produced.

- member: reviewer
  does: review the diff for correctness bugs and real blockers
  skills: [github-comment]
  model: sonnet

Loop​

A loop repeats its inner members until a condition holds, capped by max_rounds. The first inner member is the gate: it ends its output with VERDICT: APPROVE (exit) or VERDICT: CHANGES (keep going). The classic use is a review→fix cycle.

- loop:
    until: reviewer approves
    max_rounds: 3
    steps:
      - member: reviewer
        skills: [github-comment]
      - member: coder
        skills: [github-pr]

Skill​

A skill is a real, reusable capability defined in a SKILL.md file — not just a label. A member plugs a skill by name; editing the skill once updates it everywhere. Skills are discovered from your mmt library (~/.my-mini-team/skills/), your Claude Code skills (~/.claude/skills/), and the current project. Skills are strictly opt-in per member — nothing is applied to everyone by default. See Skills.

The team lead​

Every run has an implicit team lead on top. It briefs the members up front (what matters, what each should focus on), then synthesizes the final deliverable at the end. You don't add it — it's always there.

Runs happen on your subscription​

mmt run opens an interactive Claude Code session and executes members as sub-agents within it, so it uses your Claude subscription — not claude -p / the Agent SDK credit. More in Running.