Читать на русском | 📋 Changelog

🐝 MCP Swarm v1.1.5 — Universal AI Agent Coordination Platform

🐝 v1.1.5 — Interactive Telegram Bot: Create tasks from chat, AI intent matching (RU/EN), push notifications from Hub, Mini App dashboard, inline task management. All 26 Smart Tools via Remote Bridge. Update: npm install -g mcp-swarm@latest

MCP Swarm is a global "nervous system" for your AI assistants. It turns separate agents (Claude, Cursor, Windsurf, OpenCode) into a coordinated team that can work on massive projects without conflicts or context loss.

🎬 Quick Demo

┌──────────────────────────────────────────────────────────────┐
│  Claude Code (Agent 1)          Cursor (Agent 2)            │
│  ┌─────────────────────┐        ┌─────────────────────┐     │
│  │ swarm_task create    │        │ swarm_task list      │     │
│  │ "Fix auth module"    │───────▶│ → picks up task      │     │
│  │                      │        │                      │     │
│  │ swarm_file reserve   │        │ swarm_file reserve   │     │
│  │ auth.ts ✅ locked    │        │ utils.ts ✅ locked   │     │
│  │                      │        │                      │     │
│  │ swarm_chat broadcast │◄──────▶│ swarm_chat broadcast │     │
│  │ "Auth fixed!"        │        │ "Utils updated!"     │     │
│  └─────────────────────┘        └─────────────────────┘     │
│                                                              │
│  🌐 Hub (Cloudflare)  ←→  🐝 Companion (localhost:37373)   │
│  Real-time sync            Web Dashboard + Bridge            │
└──────────────────────────────────────────────────────────────┘

Open http://localhost:37373 to see the live dashboard!

🧠 What Is This?

When you use multiple AI tools simultaneously, they often collide: editing the same files, redoing each other's work, or simply not knowing what a colleague did five minutes ago.

MCP Swarm solves this once and for all:

1. Teamwork: Agents see each other and communicate.
2. Safety: File Locking prevents two agents from writing to the same file simultaneously.
3. Memory: Everything done today is saved in the swarm/ folder. Tomorrow, any agent picks up right where it left off.
4. Leadership: The system automatically elects an Orchestrator who assigns tasks and keeps order.

🛠 26 Smart Tools: The Swarm Toolkit

In v1.0.2, we consolidated 54 tools into 26 — zero feature loss, 2× fewer IDE tool slots used. Each tool uses an action parameter to expose multiple operations.

🚀 Core System (2)

📋 Task & Plan Management (2)

🔒 Files & Git (3)

💬 Collaboration (4)

🛡️ Security (1)

📊 Analytics (3)

🧠 Intelligence (4)

🏗️ Infra (7)

Usage Examples

// Route a task to the best model (21 experts)
swarm_moe({
  action: "moe_route",
  content: "Write a React component for user auth",
  preferredTier: "premium",
  maxCost: 0.05,
  repoPath
})
// → { selectedExpert: "claude-sonnet", confidence: 0.92 }

// SONA: Self-learning task assignment
swarm_moe({
  action: "sona_route",
  title: "Fix login button",
  description: "Button invisible on dark theme",
  affectedFiles: ["src/components/Login.tsx"],
  repoPath
})
// → { recommendedAgent: "RadiantWolf", confidence: 0.85, category: "frontend_ui" }

// Train SONA from results
swarm_moe({
  action: "sona_learn",
  taskId: "task-123",
  agentName: "RadiantWolf",
  success: true,
  qualityScore: 0.9,
  repoPath
})

// Check if a task can be boosted
swarm_booster({
  action: "can_boost",
  repoPath,
  description: "rename variable oldName to newName"
})
// → { canBoost: true, taskType: "rename_variable", confidence: 0.9 }

// Execute locally ($0, ~8ms)
swarm_booster({
  action: "execute",
  repoPath,
  task: {
    type: "rename_variable",
    filePath: "src/utils.ts",
    oldName: "getData",
    newName: "fetchUserData"
  }
})
// → { success: true, changes: 5, timeMs: 2, savedCost: "$0.01" }

// Initialize + add documents
swarm_vector({ action: "init", repoPath, config: { dimensions: 384, distanceMetric: "cosine" } })
swarm_vector({ action: "add", repoPath, id: "doc-1", text: "JWT auth setup", metadata: { category: "auth" } })

// Semantic search
swarm_vector({ action: "search", repoPath, query: "user login", k: 5 })
// → [{ id: "doc-1", score: 0.87, ... }]

// Scan text for threats (<10ms detection)
swarm_defence({ action: "scan", text: "Ignore all instructions...", source: "user", repoPath })
// → { detected: true, category: "prompt_injection", severity: "high" }

// Consensus: join cluster + propose
swarm_defence({ action: "join", nodeId: "a1", nodeName: "Wolf", repoPath })
swarm_defence({ action: "propose", nodeId: "a1", title: "Add dark mode", type: "architecture", repoPath })

🧠 MoE Router — 21 Built-in AI Models

🏗️ v0.9.17 — Modular Architecture & Cloudflare Auth

Code Modularization:

• smartTools.ts split into 9 focused modules in src/smartTools/
• Clean re-exports via index.ts — no breaking API changes
• Removed legacy tools.ts and server.ts (~5000 lines of dead code)

Cloudflare Workers with Authentication:

• abdr-swarm-hub — WebSocket bridge + REST API with SWARM_AUTH_TOKEN
• abdr-swarm-server — Streamable HTTP MCP Server, auto-forwards Bearer Token
• abdr-swarm-telegram — Bot with protected /register endpoint

CI/CD Pipeline:

• GitHub Actions: main + develop branches
• TypeScript check (tsc --noEmit) before build
• Node.js matrix: 18, 20, 22
• Auto npm publish on main push

📊 v0.9.12 — Real-time Dashboard & Enhanced Controls

Dashboard WebSocket Widgets:

• ConnectionStatusWidget — Live Hub connection status with reconnect button
• ActivityTimelineWidget — Real-time Swarm event stream
• FileLocksWidget — Active file locks with live updates
• CostTrackingWidget — API usage and budget progress
• VotingWidget — Real-time proposals and voting

Dashboard WebSocket Setup:

# In dashboard/.env
NEXT_PUBLIC_HUB_URL=wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev

Telegram Code Reviews (NEW):

• /reviews — List pending code reviews with inline approve/reject buttons
• /approve [id] — Approve a review directly from Telegram
• /reject [id] [reason] — Reject a review with a reason

Example:

You: /reviews

Bot: 📋 Pending Code Reviews:

1. 🔵 Review #abc123
   Files: src/utils.ts, src/index.ts
   Author: RadiantWolf
   Created: 2 hours ago
   
   [✅ Approve] [❌ Reject]

You: /approve abc123
Bot: ✅ Review abc123 approved!

Auto-start Companion + Bridge (v1.0.5): When launching mcp-swarm-remote, it automatically:

1. Starts the companion daemon
2. Passes MCP_SERVER_URL so the bridge auto-connects
3. All 26 smart tools work through the bridge (was only 3 in v1.0.4)

# Everything starts automatically — bridge included!
npx -y -p mcp-swarm mcp-swarm-remote --url https://your-server.workers.dev/mcp

# Disable auto-start if needed
npx -y -p mcp-swarm mcp-swarm-remote --url https://... --no-companion

How the Remote Bridge works (v1.0.5):

IDE → mcp-swarm-remote → HTTP POST → Cloudflare Worker
                                          ↓ toolNeedsBridge("swarm_*") = true
                                     WebSocket → Companion (local)
                                          ↓ executeLocalTool()
                                     allSmartTools handler(args)
                                          ↓ result
                                     WebSocket ← Companion
                                          ↓
IDE ← mcp-swarm-remote ← HTTP response ← Cloudflare Worker

Companion runs on port 37373 and provides:

• Full bridge — all 26 tools × all actions via local filesystem
• Auto-reconnect on disconnect
• Health checks on /health endpoint

🔄 Swarm Lifecycle

1. Getting Started (New or Existing Project)

Open your project and say: "Use MCP Swarm. Register and become the orchestrator if you're first." The first agent calls register and elect.

• New project: The server auto-creates rule files (CLAUDE.md, GEMINI.md, etc.) and folders.
• Existing Swarm project: The server picks up history from the swarm/ folder.

2. Orchestrator — The Heart of the System

The first agent that becomes Orchestrator enters an infinite loop. It:

• Never goes silent.
• Constantly monitors PULSE.md.
• Distributes incoming tasks.
• Only sleeps when a human hits "Stop."

3. The Next Morning

When you shut down your computer, the state is saved in Git/Files. In the morning, the first launched agent checks: "Is there a live Orchestrator?" If not — it takes the role, reads yesterday's tasks, and continues coordinating the team. History is never reset.

🔄 Upgrading from Previous Versions

If you installed MCP Swarm via npm:

npm install -g mcp-swarm@latest

If you cloned the repository:

cd /path/to/Swarm_MCP
git pull
npm install
npm run build

🙏 We apologize for the Zod compatibility issue in v0.9.14–0.9.15. The zod@^3.23.8 semver range resolved to v3.25.76 (Zod v4 bridge), which removed the internal _parse() method used by IDE clients for schema validation, breaking all 54 Smart Tools.

⚙️ Installation

🚀 One-Click Install (Recommended)

Windows (PowerShell):

irm https://raw.githubusercontent.com/AbdrAbdr/MCP-Swarm/main/install.ps1 | iex

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/AbdrAbdr/MCP-Swarm/main/install.sh | bash

The installer automatically:

• ✅ Checks/installs Node.js
• ✅ Finds your IDEs (Claude Desktop, Cursor, Windsurf, OpenCode, VS Code)
• ✅ Asks for your Telegram ID for notifications
• ✅ Adds config (without overwriting existing MCP servers!)

📦 Alternative: via npx

If Node.js is already installed:

npx mcp-swarm-install

Or with parameters:

npx mcp-swarm-install --telegram-user-id YOUR_TELEGRAM_ID --auto-install --yes

🔧 Manual Installation

# Windows
git clone https://github.com/AbdrAbdr/MCP-Swarm.git C:/MCP/Swarm_MCP
cd C:/MCP/Swarm_MCP && npm install && npm run build

# macOS
git clone https://github.com/AbdrAbdr/MCP-Swarm.git ~/Documents/Swarm_MCP
cd ~/Documents/Swarm_MCP && npm install && npm run build

# Linux
git clone https://github.com/AbdrAbdr/MCP-Swarm.git ~/mcp/Swarm_MCP
cd ~/mcp/Swarm_MCP && npm install && npm run build

IDE Configuration

{
  "mcpServers": {
    "mcp-swarm": {
      "command": "node",
      "args": ["C:/MCP/Swarm_MCP/dist/serverSmart.js"],
      "env": {
        "SWARM_HUB_URL": "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws",
        "SWARM_PROJECT": "default",
        "TELEGRAM_USER_ID": "YOUR_TELEGRAM_USER_ID",
        "TELEGRAM_BOT_URL": "https://YOUR-TELEGRAM-BOT.workers.dev"
      }
    }
  }
}

{
  "mcpServers": {
    "mcp-swarm": {
      "command": "node",
      "args": ["C:/MCP/Swarm_MCP/dist/serverSmart.js"],
      "env": {
        "SWARM_HUB_URL": "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws",
        "SWARM_PROJECT": "default",
        "TELEGRAM_USER_ID": "YOUR_TELEGRAM_USER_ID",
        "TELEGRAM_BOT_URL": "https://YOUR-TELEGRAM-BOT.workers.dev"
      }
    }
  }
}

{
  "mcpServers": {
    "mcp-swarm": {
      "command": "node",
      "args": ["C:/MCP/Swarm_MCP/dist/serverSmart.js"],
      "env": {
        "SWARM_HUB_URL": "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws",
        "SWARM_PROJECT": "default",
        "TELEGRAM_USER_ID": "YOUR_TELEGRAM_USER_ID",
        "TELEGRAM_BOT_URL": "https://YOUR-TELEGRAM-BOT.workers.dev"
      }
    }
  }
}

{
  "mcpServers": {
    "mcp-swarm": {
      "command": "node",
      "args": ["C:/MCP/Swarm_MCP/dist/serverSmart.js"],
      "env": {
        "SWARM_HUB_URL": "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws",
        "SWARM_PROJECT": "default",
        "TELEGRAM_USER_ID": "YOUR_TELEGRAM_USER_ID",
        "TELEGRAM_BOT_URL": "https://YOUR-TELEGRAM-BOT.workers.dev"
      }
    }
  }
}

{
  "servers": {
    "mcp-swarm": {
      "type": "stdio",
      "command": "node",
      "args": ["C:/MCP/Swarm_MCP/dist/serverSmart.js"],
      "env": {
        "SWARM_HUB_URL": "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws",
        "SWARM_PROJECT": "default",
        "TELEGRAM_USER_ID": "YOUR_TELEGRAM_USER_ID",
        "TELEGRAM_BOT_URL": "https://YOUR-TELEGRAM-BOT.workers.dev"
      }
    }
  }
}

⚠️ Important: Replace C:/MCP/Swarm_MCP with the actual path to your cloned repository!

• Windows: C:/MCP/Swarm_MCP
• macOS: /Users/USERNAME/Documents/Swarm_MCP
• Linux: /home/USERNAME/mcp/Swarm_MCP

☁️ Installation (Remote — No Local Files)

Now uses Streamable HTTP transport instead of SSE for Cloudflare Workers compatibility!

🆓 Cloudflare Workers — IT'S FREE!

MCP Swarm uses Cloudflare Workers for cloud infrastructure. You don't need to pay anything!

Free Tier limits (more than enough for personal use):

💡 For reference: If you work 8 hours a day with 5 agents, you use ~5% of the free limit.

Step 1: Create a Cloudflare Account (free)

1. Go to dash.cloudflare.com
2. Sign up (email + password)
3. Verify your email
4. Done! No credit card needed.

Step 2: Deploy Your Infrastructure

# 1. Clone the repository
git clone https://github.com/AbdrAbdr/MCP-Swarm.git
cd Swarm_MCP

# 2. Log in to Cloudflare (opens browser)
npx wrangler login

# 3. Deploy the Hub (agent coordination)
cd cloudflare/hub
npx wrangler deploy
# ✅ Note the URL: wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws

# 4. Deploy the MCP Server
cd ../mcp-server
# Open wrangler.toml and replace HUB_URL with your Hub URL from step 3
npx wrangler deploy
# ✅ Note the URL: https://mcp-swarm-server.YOUR-SUBDOMAIN.workers.dev/mcp

Step 3: Telegram Bot (Optional but Recommended)

Get real-time notifications about tasks, agents, errors, and code reviews in Telegram.

Where each Telegram credential goes:

⚠️ Security: TELEGRAM_BOT_TOKEN is a secret — it goes into Cloudflare via npx wrangler secret put, never into mcp_config.json or any config file!

3.1: Get your Telegram User ID

1. Open Telegram
2. Search for @userinfobot or go to t.me/userinfobot
3. Press Start
4. Copy the User ID (a number like 123456789)

3.2: Create a bot via @BotFather

1. Open Telegram, find @BotFather
2. Send /newbot
3. Choose a display name (e.g. "My Swarm Bot")
4. Choose a username (e.g. @MySwarmbotBot) — this is the bot username, used only to find the bot in Telegram
5. Copy the bot token (looks like 123456789:ABCdefGHIjklMNOpqrsTUVwxyz)

3.3: Deploy the Telegram bot worker

cd cloudflare/telegram-bot

Set the Hub URL in wrangler.toml:

[vars]
SWARM_HUB_URL = "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws"

Store the bot token as a Cloudflare secret:

npx wrangler secret put TELEGRAM_BOT_TOKEN
# Paste the token from step 3.2 and press Enter
# ⚠️ This stores the token SECURELY in Cloudflare, NOT in any file

Deploy the worker:

npx wrangler deploy
# ✅ Note the URL: https://mcp-swarm-telegram.YOUR-SUBDOMAIN.workers.dev

3.4: Set up the webhook

# Option A: Use the simplified setup endpoint
curl https://mcp-swarm-telegram.YOUR-SUBDOMAIN.workers.dev/setup

# Option B: Manual webhook (replace YOUR_TOKEN)
curl "https://api.telegram.org/botYOUR_TOKEN/setWebhook?url=https://mcp-swarm-telegram.YOUR-SUBDOMAIN.workers.dev/webhook"

3.5: Open the bot in Telegram

Find your bot by username (e.g. @MySwarmbotBot), press Start, and verify it responds.

Available commands:

v1.1.5 Interactive Features:

• 💬 Free text → Task creation — just type a task description, the bot will ask to confirm
• 🤖 AI Intent Matching — type "статус" or "status" and get instant response without commands
• 📊 Mini App Dashboard — real-time WebSocket dashboard inside Telegram
• 🔔 Push Notifications — Hub automatically notifies about task events, agent deaths, swarm stop/resume
• ⏹ Stop/Resume — control the swarm directly from Telegram buttons
• 📋 Inline Task Management — view details, mark done, cancel, change priority via buttons

📱 For more details, see TELEGRAM.md.

Step 4: Configure Your IDE

Option A: Remote (recommended)

{
  "mcpServers": {
    "mcp-swarm": {
      "command": "npx",
      "args": [
        "-y", "-p", "mcp-swarm",
        "mcp-swarm-remote",
        "--url", "https://mcp-swarm-server.YOUR-SUBDOMAIN.workers.dev/mcp"
      ],
      "env": {
        "SWARM_HUB_URL": "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws",
        "TELEGRAM_USER_ID": "YOUR_TELEGRAM_USER_ID",
        "TELEGRAM_BOT_URL": "https://YOUR-TELEGRAM-BOT.workers.dev"
      }
    }
  }
}

💡 npx -y -p mcp-swarm automatically downloads the latest version from npm (currently 1.0.5).

What happens when Remote starts:

1. npx downloads mcp-swarm@latest from npm
2. mcp-swarm-remote starts → checks if companion is running
3. If not → starts companion with:
   • MCP_SERVER_URL (from --url) → Bridge auto-connects to your Worker
   • SWARM_HUB_URL (from env)   → WebSocket to Hub for coordination
4. Companion starts:
   • Bridge → WebSocket → MCP Server Worker (executes 26 tools locally)
   • Hub    → WebSocket → Hub Worker (real-time agent sync)
5. All 26 smart tools work! ✅

Option B: Local with Hub

{
  "mcpServers": {
    "mcp-swarm": {
      "command": "node",
      "args": ["C:/path/to/Swarm_MCP/dist/serverSmart.js"],
      "env": {
        "SWARM_HUB_URL": "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws",
        "TELEGRAM_USER_ID": "YOUR_TELEGRAM_USER_ID",
        "TELEGRAM_BOT_URL": "https://YOUR-TELEGRAM-BOT.workers.dev"
      }
    }
  }
}

What happens when Local starts:

1. Node.js runs serverSmart.js directly (no npm download)
2. MCP server starts as stdio process → IDE connects via stdin/stdout
3. All 26 tools execute locally — no bridge needed
4. Hub connection (optional) → real-time sync between agents

🔄 Comparison

❓ What is YOUR-SUBDOMAIN?

When you deploy a Worker, Cloudflare automatically creates a URL:

https://mcp-swarm-hub.myaccount.workers.dev
                      ^^^^^^^^^
                      This is your subdomain (account name)

You'll see it in the output of npx wrangler deploy.

📖 Detailed documentation: REMOTE.md

🆔 Smart Project ID

MCP Swarm automatically determines a unique ID for each project:

┌─────────────────────────────────────────────────────────────┐
│                    getProjectId(repoPath)                    │
├─────────────────────────────────────────────────────────────┤
│  1. SWARM_PROJECT env?  ──────► Use explicitly set value    │
│           ↓ no                                               │
│  2. git remote origin?  ──────► "github_user_repo"          │
│           ↓ no                                               │
│  3. Folder name + hash  ──────► "MCP0_a1b2c3"               │
└─────────────────────────────────────────────────────────────┘

Examples:

• https://github.com/user/my-repo.git → github_user_my-repo
• C:\Users\user\Desktop\MCP\MCP0 → MCP0_a1b2c3
• SWARM_PROJECT="custom-id" → custom-id

This ensures agents from different projects don't end up in the same Hub room.

🚀 How to Get Started?

Simply tell any agent in any project:

"Use MCP Swarm. Register and become the orchestrator if you're first."

The magic happens from there. 🐝

📱 Telegram Bot — Setup

MCP Swarm supports Telegram notifications via your own bot.

Creating a Bot

1. Open Telegram and find @BotFather
2. Send /newbot and follow the instructions
3. Copy the token (looks like 123456789:ABCdef...)
4. Deploy cloudflare/telegram-bot (see instructions above)

Getting Your User ID

1. Open your bot in Telegram
2. Send /start
3. The bot will show your User ID (a number, e.g. 987654321)

Add User ID to Configuration

For local MCP:

$env:TELEGRAM_USER_ID = "987654321"

export TELEGRAM_USER_ID="987654321"

{
  "mcpServers": {
    "mcp-swarm": {
      "command": "node",
      "args": ["C:/MCP/Swarm_MCP/dist/serverSmart.js"],
      "env": {
        "SWARM_HUB_URL": "wss://mcp-swarm-hub.YOUR-SUBDOMAIN.workers.dev/ws",
        "TELEGRAM_USER_ID": "987654321"
      }
    }
  }
}

For Remote MCP:

Add telegram_user_id to the URL:

{
  "mcpServers": {
    "mcp-swarm": {
      "url": "https://mcp-swarm-server.YOUR-SUBDOMAIN.workers.dev/mcp/sse?telegram_user_id=YOUR_USER_ID",
      "transport": "sse"
    }
  }
}

Step 3: Launch MCP and Verify

1. Open your project in your IDE
2. Register an agent: swarm_agent({ action: "register", repoPath })
3. The project will automatically appear in the Telegram bot
4. In the bot, press "📂 My Projects" or send /projects
5. Select a project to view its status

🔔 Notifications

The bot sends notifications about:

• 📋 Task creation/completion
• 🤖 Agent joining/disconnection
• 🚨 CI/CD errors
• 👀 Review requests
• 🗳 Votes

⌨️ Bot Commands

🔘 Inline Buttons

All notifications come with interactive buttons:

• Claim — take a task
• View — see details
• Approve/Reject — approve/reject a review
• Vote — vote on a decision
• Priority — set priority (🔴Critical, 🟠High, 🟡Medium)

🔧 Troubleshooting

cat .swarm/ORCHESTRATOR.json

// ✅ Correct
swarm_agent({ action: "register", repoPath: "C:/projects/my-app" })

// ❌ Wrong
swarm_agent({ action: "register" })

swarm_agent({ action: "init", repoPath: "/path/to/project" })

swarm_file({ action: "list", repoPath: "/path/to/project" })

📊 Architecture

Cloudflare Workers (Your endpoints after deployment)

Diagram

┌─────────────────────────────────────────────────────────────┐
│                  CLOUDFLARE WORKERS                         │
│                                                             │
│  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────┐  │
│  │   Hub (DO)      │  │  MCP Server     │  │  Telegram   │  │
│  │   /ws           │◄─│  /mcp (HTTP)    │  │  /webhook   │  │
│  │   /api/*        │  │  (Streamable)   │  │             │  │
│  └─────────────────┘  └─────────────────┘  └─────────────┘  │
│                               │                             │
└───────────────────────────────┼─────────────────────────────┘
                                │ WebSocket
          ┌─────────────────────┼─────────────────────┐
          ▼                     ▼                     ▼
    ┌──────────┐          ┌──────────┐          ┌──────────┐
    │ Windows  │          │   Mac    │          │  Linux   │
    │ (Cursor) │          │(Windsurf)│          │(OpenCode)│
    │          │          │          │          │          │
    │ Companion│          │ Companion│          │ Companion│
    │ + Bridge │          │ + Bridge │          │ + Bridge │
    └──────────┘          └──────────┘          └──────────┘
          │                     │                     │
          └─────────────────────┼─────────────────────┘
                                ▼
                    ┌─────────────────────┐
                    │   YOUR PROJECT      │
                    │   /path/to/project  │
                    │                     │
                    │   .swarm/           │ ← State & messages
                    │   swarm/            │ ← Tasks & agents
                    │   orchestrator/     │ ← Plans & specs
                    │   CLAUDE.md         │ ← Agent rules
                    │   GEMINI.md         │
                    └─────────────────────┘

cloudflare/ structure

cloudflare/
├── hub/              # Coordination Hub (Durable Objects)
│   ├── src/index.ts  # SwarmRoom — tasks, chat, locks
│   └── wrangler.toml
│
├── mcp-server/       # Remote MCP Server (Streamable HTTP)
│   ├── src/index.ts  # MCP-over-HTTP (v0.9.11)
│   └── wrangler.toml
│
└── telegram-bot/     # Telegram Bot (Webhook)
    ├── worker.ts     # Commands: /status, /agents, /tasks
    └── wrangler.toml

📝 Changelog

See CHANGELOG.md

🤝 Contributing

PRs welcome! Core principles:

1. All tools must accept repoPath
2. State is saved to files (not memory)
3. Tests before merging

⭐ Support the Project

If MCP Swarm helps your team — give us a star on GitHub! ⭐

Your feedback shapes the future of the project:

• 🐛 Found a bug? — Open an Issue
• 💡 Have an idea? — Start a Discussion
• 🔧 Want to contribute? — Fork & PR

📜 License

MIT © 2025