A multi-agent orchestration platform that organizes AI agents (Claude, GPT, Gemini, etc.) into collaborative teams with persistent identity, cross-model review, and shared memory, with out-of-the-box desktop deployment.
Core Positioning#
Clowder AI is a multi-agent orchestration platform built with Node.js/TypeScript using a pnpm workspace monorepo architecture. Under a three-layer separation design (Model, Agent CLI, Platform), the platform layer provides independent AI agents with collaborative capabilities.
Three-Layer Architecture#
| Layer | Responsibility |
|---|---|
| Model | Reasoning, generation, understanding |
| Agent CLI | Tool use, file operations, command execution |
| Platform (Clowder) | Identity, collaboration, discipline, audit |
Core Capabilities#
- Multi-Agent Orchestration: Route tasks to appropriate agents—Claude for architecture, GPT for review, Gemini for design—in a single conversation
- Persistent Identity: Each agent maintains role, persona, and memory across sessions and context compression
- Cross-Model Review: Built-in cross-model review mechanism (e.g., Claude codes, GPT reviews)
- A2A Communication: Asynchronous agent-to-agent messaging with @mention routing, thread isolation, and structured handoffs
- Shared Memory: Evidence store, lessons learned, decision logs—persistent institutional knowledge
- Skills Framework: On-demand prompt loading (TDD, debugging, review, etc.)
- MCP Integration: Model Context Protocol for cross-agent tool sharing, non-Claude models via callback bridge
- Collaborative Discipline: Automated SOPs: design gates, quality checks, vision guardianship, merge protocols
Notable Features#
- CVO Mode: Human as Chief Vision Officer for vision expression, key decisions, culture shaping, and co-creation
- Voice Companion: Hands-free voice mode with per-agent unique voice, ASR input and TTS playback
- Signals: AI research feed aggregating RSS/blogs with priority grading, multi-agent collaborative reports and podcast generation
- Game Modes: Play games (e.g., Werewolf) with AI teams
- Mission Hub: Feature lifecycle management (idea → spec → in-progress → review → done), requirements audit, and SOP boards
Safety Constraints (Iron Laws)#
Four constraints enforced at both prompt and code levels:
- Never delete own database (memory must not be discarded)
- Never kill parent process (preserve own existence)
- Runtime configuration is read-only for agents (human modification required)
- Never touch other agents' ports (isolation principle)
Platform Core Modules#
- Identity Manager — Persistent agent identity management
- A2A Router & Threads — Asynchronous inter-agent message routing and thread isolation
- Skills Framework & Manifest — On-demand skill loading framework
- Memory & Evidence — Shared memory and evidence storage
- SOP Guardian — Automated standard operating procedure guardianship
- MCP Callback Bridge — Cross-model MCP tool sharing bridge
Supported Agent CLIs#
| CLI | Models |
|---|---|
| Claude Code | Claude Opus / Sonnet / Haiku |
| Codex CLI | GPT / Codex |
| Gemini CLI | Gemini |
| Antigravity | Multi-model |
| opencode | Multi-model |
Interaction Methods#
- @mention routing:
@opusarchitecture,@codexreview,@geminidesign - Slash commands:
/new(new thread),/threads(list),/use <id>(switch),/where(current location)
External Integrations#
| Channel | Status |
|---|---|
| Feishu (Lark) | Supported |
| GitHub PR Review | Supported (auto review routing) |
| Telegram | In progress |
Supports third-party model providers including Kimi / GLM / MiniMax.
Deployment#
- Windows: Download .exe installer (bundled portable Node.js and Redis, no manual dependencies)
- macOS: Download .dmg, drag to Applications (right-click → Open to bypass Gatekeeper on first launch)
- Linux: Use
bash scripts/install.shone-click installation - Source install:
git clone→pnpm install→pnpm build→cp .env.example .env→pnpm start
After launch, access http://localhost:3003, add model API keys in Hub → System Settings → Account Configuration.
Monorepo Structure#
clowder-ai/
├── packages/
│ ├── api/ # Backend API service
│ ├── mcp-server/ # MCP server
│ ├── shared/ # Shared types and utilities
│ └── web/ # Frontend Web UI
├── cat-cafe-skills/ # Skills framework
├── desktop/ # Desktop installers
├── docs/ # Documentation
├── guides/ # Guides
└── scripts/ # Ops and install scripts
Unconfirmed Information#
- Independent website/online demo: Not found in repository, only local access after launch
- Academic papers: None found
- Hugging Face: No associated page found
- Specific ASR/TTS providers for voice features not disclosed
- Internal codename
cat-cafe(package.json name field) vs public nameclowder-ai: repository synced from internal codename