Minion Comms

Minion Comms

Multi-agent coordination MCP server using RPG raid mechanics for task assignment, context management, and knowledge persistence.

Category
访问服务器

README

Minion Comms

Multi-agent coordination server inspired by RPG raid mechanics. Built on MCP (Model Context Protocol) — any AI tool that speaks MCP can join the raid: Claude Code, Codex CLI, OpenCode, Gemini, or anything else.

What is this?

AI agents have finite context windows. When multiple agents work on the same codebase, they need coordination — who's doing what, what's been tried, where the traps are, and when someone's about to die (run out of context).

Minion Comms treats multi-agent engineering like an RPG raid. The codebase is the boss. Agents are the raid party. Context is reverse HP — you start full and every action drains you.

If you ever raided in WoW or wiped in FF, you already get this

Context windows are HP bars. Compaction is getting dazed mid-fight. A coder exploring the wrong files is a DPS pulling aggro on trash mobs. An oracle who hasn't read the latest intel is a healer with stale buffs. And when the lead goes down, it's a raid wipe — nobody else has the full picture.

This isn't a metaphor bolted on after the fact. The problems are the same problems:

  • Party composition matters. One buffed coder with oracle support > three coders exploring alone. Same as one geared DPS with a dedicated healer > three undergeared DPS facepulling.
  • You can't outheal stupid. High activity count on a task means wrong approach — ice spell on an ice boss. Pull back, reassess, change angle.
  • Loot that isn't picked up is wasted. An agent's findings that aren't written down die when their context window fills. XP that isn't shared is XP lost.
  • The raid leader's job is fighting entropy. Knowledge wants to die with agents. Lead's job is making sure it doesn't.

Inspired by

  • World of Warcraft — raid composition, tank/DPS/healer roles, buff coverage, aggro management, zone assignments, raid leader coordination
  • Final Fantasyfenix_down (Phoenix Down) for revival after context death, party class system
  • Majora's Maskmoon_crash for emergency shutdown (the moon is falling, everyone dump and run)
  • Dead Drop Teams — v1 of this system. SQLite message passing with role-based agents. Minion Comms is the evolution.
  • Vercel/Next.js — filesystem-as-database pattern for convention files
  • Military comms — brevity codes, CC discipline, chain of command

Core Concepts

5 Classes

Class Archetype What they do
lead Commander Coordinates, routes tasks, manages HP bars
coder DPS Edits code — the only class that changes source
builder Tank Runs commands — build, test, deploy. No edits
oracle Sage Holds zone knowledge, answers questions. No edits, no commands
recon Scout Investigates external intel — web, other repos, ecosystem. Reports back

Class = capabilities (permanent). Role = assignment (runtime). An oracle assigned to the audio zone becomes oracle-audio.

Two Databases

Database Stores Examples
SQLite (messages.db) Coordination state Agents, messages, task metadata, file claims, battle plans
Filesystem (.dead-drop/) Knowledge Intel, traps, zone notes, task specs, agent loot

SQLite tracks state. Filesystem stores knowledge. Comms surfaces file locations, agents read the files.

Enforcement Philosophy

Comms enforces what it owns (DB state). Reminds on what it can't verify (battle-time behavior).

Enforced: inbox discipline, context freshness, file claims, task dependencies, class restrictions, result files, battle plan requirement

Reminded: poll.sh running, agents reading files, following specs, HP truthfulness

Role Hierarchy

user (the human)
└── general (puppet — relays user intent)
    └── commander (runs the fight — needs brains)
        └── zone-lead (owns a section)
            └── party (oracle, coder, builder, recon)

Key Mechanics

  • HP — context is reverse HP. Tracked via set_context. Lead monitors all bars.
  • Fenix Down — dump knowledge to disk before context death. Come back clean.
  • Moon Crash — emergency shutdown. Everyone fenix_down NOW.
  • Trigger Words — brevity codes (sitrep, rally, retreat, hot_zone) save HP on both sides.
  • Activity Count — auto-increments on every task update. High count = wrong approach.
  • File Claims — prevents friendly fire. Can't edit a file another agent holds.
  • Intel/Traps — confirmed findings and known hazards in filesystem. Oracle must read them.
  • Battle Journey — agents must write up what they learned before a task can close.

Transport: Hybrid Terminal + Headless

Agents connect via two transport modes on the same comms network:

Transport Agent type Message delivery
terminal Human opens Claude Code CLI poll.sh polls inbox
daemon minion-swarm spawns headless Swarm daemon watches DB, injects on wake

Human opens terminals for high-value agents they want eyes on. Cheap work goes to swarm daemons. All peers on the same comms — same enforcement, same raid log.

Tools (36)

Phase Tools
Core Comms register, deregister, rename, set_status, set_context, who, send, check_inbox, get_history, purge_inbox
War Room set_battle_plan, get_battle_plan, update_battle_plan_status, log_raid, get_raid_log
Task System create_task, assign_task, update_task, get_tasks, get_task, submit_result, close_task
File Safety claim_file, release_file, get_claims
Monitoring party_status, check_activity, check_freshness
Lifecycle cold_start, fenix_down, debrief, end_session
Trigger Words get_triggers, clear_moon_crash

Install

curl -sSL https://raw.githubusercontent.com/ai-janitor/minion-comms/main/scripts/install.sh | bash

This installs minion-comms via pipx, deploys onboarding docs to ~/.minion-comms/, and configures MCP for Claude Code.

Or install directly:

pipx install git+https://github.com/ai-janitor/minion-comms.git

Quick Start

# Run MCP server (stdio transport)
minion-comms

# Or add to your .mcp.json
{
  "mcpServers": {
    "minion-comms": {
      "command": "minion-comms"
    }
  }
}

Status

Server implementation complete (Phases 0-8). One-liner install, onboarding docs, and MCP auto-config included.

See docs/FRAMEWORK.md for the full design specification. See PLAN.md for the implementation roadmap.

Lineage

v1: dead-drop-teams — basic message passing + task tracking

v2: minion-comms — RPG-inspired coordination with classes, HP management, knowledge persistence, enforcement philosophy, and trigger words

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选