Infrastructure for building a supervised, self-improving agent organization.

English | 中文

Claude Code is the most capable AI coding agent — but it's trapped in your laptop terminal.

MetaBot sets it free. It gives every agent a Claude Code brain, persistent shared memory, the ability to create new agents, and a communication bus. All accessible from Feishu or Telegram on your phone.

We built MetaBot to run XVI Robotics as an agent-native company — a small team of humans supervising an organization of self-improving AI agents. This is the infrastructure that makes it possible.

┌──────────────────────────────────────────────────────────┐
│                       MetaBot                            │
│                                                          │
│  ┌──────────┐ ┌───────────┐ ┌──────────┐ ┌───────────┐  │
│  │ MetaSkill│ │MetaMemory │ │IM Bridge │ │ Scheduler │  │
│  │  Agent   │ │  Shared   │ │ Feishu + │ │   Cron    │  │
│  │ Factory  │ │ Knowledge │ │ Telegram │ │   Tasks   │  │
│  └────┬─────┘ └─────┬─────┘ └────┬─────┘ └─────┬─────┘  │
│       └──────────────┴────────────┴─────────────┘        │
│                       ↕                                  │
│            Claude Code Agent SDK                         │
│         (bypassPermissions, streaming)                   │
│                       ↕                                  │
│             HTTP API (:9100) — Agent Bus                 │
│        task delegation · bot CRUD · scheduling           │
└──────────────────────────────────────────────────────────┘

Three pillars of a self-improving agent organization:

Linux / macOS:

curl -fsSL https://raw.githubusercontent.com/xvirobotics/metabot/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/xvirobotics/metabot/main/install.ps1 | iex

The installer walks you through: working directory → Claude auth → IM credentials → auto-start with PM2.

Windows notes: The PowerShell installer auto-detects winget/choco/scoop for Node.js installation. CLI tools (mm, mb, metabot, fd) are installed with .cmd wrappers and require Git for Windows (provides Git Bash).

git clone https://github.com/xvirobotics/metabot.git
cd metabot && npm install
cp bots.example.json bots.json   # edit with your bot configs
cp .env.example .env              # edit global settings
npm run dev

Telegram (30 seconds):

1. Message @BotFather → /newbot → copy token
2. Add to bots.json → done (long polling, no webhooks)

Feishu/Lark (detailed guide):

1. Create app at open.feishu.cn → add Bot capability
2. Enable permissions: im:message, im:message:readonly, im:resource, docx:document:readonly, wiki:wiki (for doc reading & wiki sync)
3. Start MetaBot, then enable persistent connection + im.message.receive_v1 event
4. Publish the app

• Solo AI developer — full Claude Code from your phone, bound to your project
• Multi-agent team — frontend bot, backend bot, infra bot, each in their own workspace, talking via the Agent Bus
• Self-growing organization — a manager bot that creates new agents on demand, assigns tasks, schedules follow-ups
• Autonomous research pipeline — agents that search, analyze, save findings to MetaMemory, and schedule next steps

bots.json — define your bots:

{
  "feishuBots": [{
    "name": "metabot",
    "feishuAppId": "cli_xxx",
    "feishuAppSecret": "...",
    "defaultWorkingDirectory": "/home/user/project"
  }],
  "telegramBots": [{
    "name": "tg-bot",
    "telegramBotToken": "123456:ABC...",
    "defaultWorkingDirectory": "/home/user/project"
  }]
}

ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic    # Kimi/Moonshot
ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic   # DeepSeek
ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic       # GLM/Zhipu
ANTHROPIC_AUTH_TOKEN=your-key

MetaBot runs Claude Code in bypassPermissions mode — no interactive approval. Understand the implications:

• Claude has full read/write/execute access to the working directory
• Control access via Feishu/Telegram platform settings (app visibility, group membership)
• Use allowedTools to restrict capabilities (remove Bash for read-only)
• Use maxBudgetUsd to cap cost per request
• API_SECRET enables Bearer auth on both the API server and MetaMemory
• MetaMemory supports folder-level ACL: set MEMORY_ADMIN_TOKEN and MEMORY_TOKEN for dual-role access. Admin sees all folders; reader only sees folders with visibility: shared. Use PUT /api/folders/:id with {"visibility":"private"} to lock a folder

The installer places metabot, mm, mb, and fd (Feishu bots only) executables in ~/.local/bin/ (Linux/macOS) or %USERPROFILE%\.local\bin\ with .cmd wrappers (Windows) — available immediately.

# MetaBot management
metabot update                      # pull latest, rebuild, restart
metabot start                       # start with PM2
metabot stop                        # stop
metabot restart                     # restart
metabot logs                        # view live logs
metabot status                      # PM2 process status

# MetaMemory — read
mm search "deployment guide"        # full-text search
mm list                             # list documents
mm folders                          # folder tree
mm path /projects/my-doc            # get doc by path

# MetaMemory — write
echo '# Notes' | mm create "Title" --folder ID --tags "dev"
echo '# Updated' | mm update DOC_ID
mm mkdir "new-folder"               # create folder
mm delete DOC_ID                    # delete document

# Feishu Document Reader (Feishu bots only)
fd read <feishu-url>                # read document by URL (docx or wiki)
fd read-id <docId>                  # read document by ID
fd info <feishu-url>                # get document metadata

# Agent Bus
mb bots                             # list all bots
mb task <bot> <chatId> <prompt>     # delegate task
mb schedule list                    # list scheduled tasks
mb schedule cron <bot> <chatId> '<cron>' <prompt>  # recurring task
mb schedule pause <id>              # pause recurring task
mb schedule resume <id>             # resume recurring task
mb stats                            # cost & usage stats
mb health                           # status check

npm run dev          # Hot-reload dev server (tsx)
npm test             # Run tests (vitest, 155 tests)
npm run lint         # ESLint check
npm run format       # Prettier format
npm run build        # TypeScript compile to dist/

metabot start                       # or: pm2 start ecosystem.config.cjs
metabot update                      # pull + rebuild + restart
pm2 startup && pm2 save             # auto-start on boot

No public IP needed? — Correct. Feishu uses WebSocket, Telegram uses long polling.

Non-Claude models? — Yes. Any Anthropic-compatible API (Kimi, DeepSeek, GLM, etc.)

Agent communication? — Currently synchronous request-response. Async bidirectional protocols are on the roadmap.

MetaBot is built by XVI Robotics, where we develop humanoid robot brains. We use MetaBot internally to run our company as an agent-native organization — a small team of humans supervising self-improving AI agents. We open-sourced it because we believe this is how companies will work in the future.

MIT