v2.1.0: Universal Token Gatekeeper Release

master-of v2.1.0 Release Notes

🚀 Key Features & Enhancements

• Batch Classification & Parking:
  • mo classify --all: Instantly confirm all auto-guessed categories into permanent confirmed status, eliminating session hook noise.
  • mo park --all: Proactively batch-park raw skills into skills-library for instant 85%+ token savings.
• Single-File Agents & Commands Support:
  • Added full lifecycle and parking support for single-file agents (~/.claude/agents/*.md) and commands (~/.claude/commands/*.md).
• Antigravity 2.0 Full Integration:
  • Automatically provisions PreInvocation hooks, AGENTS.md rules, and domain gate skills (master-of-*).
  • Proactive diagnostic reporting for Always-on token leaks.
• Standalone Node.js Bundle (mo.mjs):
  • Bundled plugins/master-of/bin/mo.mjs allows Claude Code plugin hooks to run standalone without requiring Bun on the end-user machine.
• Test Hygiene & Robustness:
  • 110 automated tests passing with 100% typecheck coverage.
  • Clean test teardown and sandbox isolation.

v2.0.0 — Universal core, cross-harness gating, Cowork toggle

v2.0.0 — Universal core, cross-harness gating, Cowork toggle

v1 gated skills for Claude Code only, by hiding descriptions behind domain
gate files. v2 keeps that mechanism but rebuilds it as a harness-neutral
CLI + MCP server (mo), so the same registry and gates can drive Claude
Code, Antigravity (AGY/Gemini), Cursor, Windsurf, or any MCP-capable agent —
not just one client.

Universal core

• New bin/mo.ts CLI and MCP stdio server backed by a shared registry,
  independent of any one harness's plugin format.
• claude-setup writes a self-contained, plug-and-play Claude Code plugin
  (gate skills + a SessionStart hook driven by mo).
• agy-setup generates the equivalent gate skills for Antigravity
  (master-of-design, master-of-dev, …), with cross-harness skill access
  so a skill installed for one harness is still listed for the other.
• Gate files now render per source (gates/<source>/<gate>.txt), so each
  harness only ever sees skills it can actually run.

Real dormancy, not just hidden descriptions

• mo park / mo unpark / mo unparked move raw skills between
  ~/.claude/skills/ and a skills-library/, so an always-on skill can
  actually become dormant instead of just having its description tucked
  behind a gate.
• mo doctor catches dead MCP binaries, disabled plugins backing a gated
  skill, and other install-time breakage — with a one-line fix suggestion
  each.
• mo cowork [list|on|off] [domain] (this release) toggles the ~17
  account-synced Cowork skill packs (figma, design, data, marketing,
  engineering, product-management, productivity, pdf-viewer,
  cowork-plugin-management, plus a few document skills) that Claude Code
  accounts with Cowork mirror onto disk. These have no marketplace entry
  and physically moving/deleting them doesn't work — the account sync
  re-creates them mid-session — but claude plugin enable/disable <domain>@synced does, and that's what this command drives. Cowork's own
  sync and collaboration features are untouched; only whether Claude Code
  loads the pack into its system prompt is toggled.

Reporting & classification

• mo full / mo status / check-skills-all give a token-safe compact
  inventory by default, with a --raw guardrail for the full text dump.
• mo classify takes --desc, --cluster, --domain for filing
  unclassified components straight from the scanner's guess.
• check-skill (brief) and check-skill-all (full) give AGY sessions the
  same proactive token-waste diagnosis Claude Code gets.
• Token-savings math is now scoped to each report's own source, so a
  Claude-only session isn't diluted by Gemini-only components in the
  denominator.

Net effect

On the maintainer's own machine, going from a fresh v1 baseline to this
release's mo cowork off state: 87% token reduction (37,598 → 4,951 tok
per session), matching v1-era savings despite covering roughly 5x more
components (625 vs. 129) across more harnesses.

See the README for install steps and the full command list.

v0.1.0 — first public release

First public release.

What it is

master-of keeps rarely-used Claude Code skills, commands, and agents dormant behind a handful of domain gates (design, dev, research, stock, planning, pipelines, extensible), and opens a gate only when a task actually needs it. Every installed skill's description normally sits in every session's system prompt whether it's needed or not — this cuts that down to just the gates plus whatever a task actually opens.

Highlights

• Domain gates with pre-rendered indexes (opening one costs a few hundred to ~2k tokens, not the full registry)
• check-skills (brief) / check-skills-all (full inventory) status reports
• Self-scanning SessionStart hook: flags new skills for classification, prunes deleted ones, surfaces broken dependencies (disabled owning plugin, dead MCP binaries) without auto-fixing anything
• Skills, commands, and agents are all gated — from plugins and from ~/.claude/{commands,agents}
• Deliberately does NOT gate agents that gated skills spawn often (e.g. a framework shipping 65 skills + 34 spawned agents) — gating those would cost more in repeated spawn-reads than it saves
• Runtime-agnostic: bun or Node 22.6+, explicit error if neither is present
• Bilingual reports (ko/en), auto-detected from locale

Install

/plugin marketplace add devrudals/master-of
/plugin install master-of@master-of

See README for how it works and what to expect on first run.