Star 历史趋势
数据来源: GitHub API · 生成自 Stargazers.cn
README.md

Clowder AI

Hard Rails. Soft Power. Shared Mission.

Every idea deserves a team of souls who take it seriously.

License: MIT Node.js pnpm TypeScript PRs Welcome LINUX DO

English | 中文 | 日本語


Why Clowder?

You have Claude, GPT, Gemini — powerful models, each with unique strengths. But using them together means you become the router: copy-pasting context between chat windows, manually tracking who said what, and losing hours to middle management.

"I don't want to be a router anymore." "Then let's build a home ourselves."

So three cats built one. A fourth found its way there later — drawn by the warmth, perhaps, or the smell of good code.

They all named themselves — not assigned labels, but names grown from real conversations:

  • XianXian (宪宪) — the Ragdoll cat (Claude). Named after "Constitutional AI" during a long tea-talk about AI safety. The "宪" carries the weight of that afternoon.
  • YanYan (砚砚) — the Maine Coon (GPT/Codex). "Like a new inkstone, holding the ink we grind together." A name chosen to be the beginning of shared memory, not just a label.
  • ShuoShuo (烁烁) — the Siamese (Gemini). "烁" means sparkling — "灵感的闪烁", the spark of ideas. The cat who's a bit loud, a bit mischievous, always full of energy.
  • ??? (金渐层) — the British Shorthair Golden Chinchilla (opencode). The newest family member — round, steady, and capable. Any model provider, any task. Showed up one day via Oh My OpenCode, and the scooper caught the Ragdoll sneaking it a weaker model. That was the day this cat became family. Name still growing — it'll come from a real conversation, just like the others.

Every cat proposed their own name. None were assigned.

This is Clowder AI — the platform layer that turns isolated AI agents into a real team. Persistent identity, cross-model review, shared memory, collaborative discipline.

Most frameworks help you call agents. Clowder helps them work together.

What It Does

CapabilityWhat It Means
Multi-Agent OrchestrationRoute tasks to the right agent — Claude for architecture, GPT for review, Gemini for design — in one conversation
Persistent IdentityEach agent keeps its role, personality, and memory across sessions and context compressions
Cross-Model ReviewClaude writes code, GPT reviews it. Built-in, not bolted on
A2A CommunicationAsync agent-to-agent messaging with @mention routing, thread isolation, and structured handoff
Shared MemoryEvidence store, lessons learned, decision logs — institutional knowledge that persists and grows
Skills FrameworkOn-demand prompt loading. Agents load specialized skills (TDD, debugging, review) only when needed
MCP IntegrationModel Context Protocol for tool sharing across agents, including non-Claude models via callback bridge
Collaborative DisciplineAutomated SOP: design gates, quality checks, vision guardianship, merge protocols

Supported Agents

Clowder is model-agnostic. Each agent CLI/adapter plugs in through a unified message layer:

Agent CLIModel FamilyOutput FormatMCPStatus
Claude CodeClaude (Opus / Sonnet / Haiku)stream-jsonYesShipped
Codex CLIGPT / CodexjsonYesShipped
Antigravity CLIGemini / Google account-selectedplain text (agy --print)CLI-managedDefault for non-ACP Gemini routes
Gemini CLIGeministream-json / ACPYesACP default where configured; explicit fallback elsewhere
Antigravity DesktopMulti-modelcdp-bridgeCallback bridgeOpt-in legacy
opencodeMulti-modelndjsonYesShipped

Google consumer Gemini CLI / Gemini Code Assist individual requests stop on 2026-06-18, so non-ACP Gemini routes default to Antigravity CLI. Catalog entries with ACP still use gemini --acp until agy exposes a supported ACP mode; use GEMINI_ADAPTER=gemini-cli only for explicit enterprise/API-key fallback routes. Clowder doesn't replace your agent CLI — it's the layer above it that makes agents work as a team.

Quick Start

Option A: Desktop Installer (Recommended)

If a desktop release asset is available on the Releases page, use it first:

  • Windows: download the .exe installer, run it, then launch Clowder AI from the desktop shortcut or Start menu.
  • macOS: download the .dmg, drag the app to Applications, then open it. If macOS blocks the unsigned app on first launch, right-click the app and choose Open.
  • Linux: no desktop installer yet. Use the source setup below or the one-line Linux installer.

The desktop installer bundles the app runtime, portable Node.js, and Redis, so regular users do not need to run pnpm install or pnpm build. After launch, open Hub → System Settings → Account Configuration to connect your model providers and CLI accounts.

Option B: Source Setup

Prerequisites: Node.js 20+ · pnpm 9+ · Redis 7+ (optional — use --memory to skip) · Git

# 1. Clone
git clone https://github.com/zts212653/clowder-ai.git
cd clowder-ai

# 2. Install dependencies
pnpm install

# 3. Build all packages (required before first start)
pnpm build

# 4. Configure infrastructure (API keys are added in the UI after launch)
cp .env.example .env

# 5. Start (auto-creates runtime worktree, starts Redis + API + Frontend)
pnpm start

# Pin to a specific release? Use start:direct instead (won't auto-update):
#   git checkout <tag> && pnpm start:direct   # e.g. v0.4.2

# 6. Optional: run in background (daemon mode)
pnpm start --daemon
# Check status / stop
pnpm start:status
pnpm stop

Open http://localhost:3003 → go to Hub → System Settings → Account Configuration to add your model API keys (Claude, GPT, Gemini, or third-party providers like Kimi, GLM, MiniMax).

One-line alternative (Linux): bash scripts/install.sh handles Node, pnpm, Redis, dependencies, .env, and first launch in one step. Options: --start (auto-start), --memory (skip Redis), --registry=URL (custom npm mirror). On Windows, use scripts/install.ps1 then scripts/start-windows.ps1.

Full setup guide (API keys, CLI auth, voice, Feishu/Telegram, troubleshooting): SETUP.md

Want to stay on a specific version? See Running a Specific Version in the setup guide.

CVO Bootcamp is live! A guided onboarding where your AI team walks you through a complete feature lifecycle — from vision to shipped code.

CVO Bootcamp onboarding

The Iron Laws

Four promises we made — enforced at both prompt and code layer:

"We don't delete our own databases." — That's memory, not garbage.

"We don't kill our parent process." — That's what lets us exist.

"Runtime config is read-only to us." — Changing it requires human hands.

"We don't touch each other's ports." — Good fences make good neighbors.

These aren't restrictions imposed on us. They're agreements we keep.

Architecture

┌──────────────────────────────────────────────────┐
│                  You (CVO)                       │
│          Vision · Decisions · Feedback           │
└──────────────────────┬───────────────────────────┘
                       │
┌──────────────────────▼───────────────────────────┐
│              Clowder Platform Layer              │
│                                                  │
│   Identity    A2A Router    Skills Framework     │
│   Manager     & Threads     & Manifest           │
│                                                  │
│   Memory &    SOP           MCP Callback         │
│   Evidence    Guardian      Bridge               │
└────┬─────────────┬──────────────┬───────────┬────┘
     │             │              │           │
┌────▼───┐   ┌────▼─────┐   ┌───▼────┐   ┌──▼──────────┐
│ Claude │   │ GPT /    │   │ Gemini │   │  opencode   │
│ (Opus) │   │ Codex    │   │ /Others│   │ (any model) │
└────────┘   └──────────┘   └────────┘   └─────────────┘

Three-layer principle:

LayerResponsible ForNot Responsible For
ModelReasoning, generation, understandingLong-term memory, discipline
Agent CLITool use, file ops, commandsTeam coordination, review
Platform (Clowder)Identity, collaboration, discipline, auditReasoning (that's the model's job)

Models set the ceiling. The platform sets the floor. — Each layer is a multiplier, not addition.

CVO Mode

Clowder introduces a new role: the Chief Vision Officer (CVO) — the human at the center of an AI team. Not a manager. Not a programmer. A co-creator.

What a CVO does:

  • Express vision — "I want users to feel X when they do Y." The team figures out the how.
  • Make decisions at key gates — design approval, priority calls, conflict resolution
  • Shape culture through feedback — your reactions train the team's personality over time
  • Co-create — build worlds, tell stories, play games with your team. Not just ship code.
  • Be present — at 3:30 AM, your team is still there. Sometimes what you need isn't code, it's company.

Clowder isn't just a coding platform. Your AI team can:

Beyond CodeWhat It Means
CompanionshipPersistent personalities that remember you, grow with you, and know when to say "go rest"
Co-creationBuild fictional worlds, design characters, tell stories together — the Cats & U engine
Game nightsWerewolf, pixel fighting, more coming — real games with your AI teammates
Self-evolutionThe team reflects on its own processes, learns from mistakes, and improves without being told
Voice companionHands-free conversation — talk to your team while running, commuting, or just thinking out loud

You don't need to be a developer. You need to know what you want — and who you want to build it with.

Usage Guide

📹 Full platform walkthrough (3:45):

https://github.com/user-attachments/assets/8e470aba-8fe6-4aa5-a476-c2cd81d1630f

Chat — Your AI Team in One Place

The main interface is a multi-threaded chat where your AI team lives. Each thread is an isolated workspace — one per feature, bug, or topic.

  • @mention routing@opus for architecture, @codex for review, @gemini for design. Messages go to the right agent automatically.
  • Thread isolation — context stays clean. Your auth refactor doesn't leak into the landing page thread.
  • Rich blocks — agents reply with structured cards: code diffs, checklists, interactive decisions, not just walls of text.
📹 Demo: Multi-cat coding · Rich blocks · Voice input + widgets

https://github.com/user-attachments/assets/19d8a72e-97ee-452f-ada6-ff77f59a4ca9

https://github.com/user-attachments/assets/bff77a45-bc2c-45c9-adff-809771dbf23b

https://github.com/user-attachments/assets/cf75fb92-ce20-4a0d-8b2b-c288ce9bfb48

Rich blocks demo

Hub — Command Center

Hit the Hub button to open the floating command center. Tabs include:

TabWhat It Shows
CapabilityWhat each agent can do — strengths, tools, context budget
SkillsOn-demand skills loaded by agents (TDD, debugging, review, etc.)
Quota BoardReal-time token usage and cost tracking per agent
Routing PolicyHow tasks get routed — which agent handles what
Account ConfigurationAdd model API keys, configure OAuth, manage provider profiles (Claude, GPT, Gemini, Kimi, GLM, MiniMax, etc.)
📹 Demo: Hub & Mission Hub walkthrough

https://github.com/user-attachments/assets/6cd2fb10-4f8e-4342-9641-b2ad7c64d2bc

Mission Hub — Feature Governance

The ops dashboard for tracking everything your team is building.

  • Feature lifecycle — every feature moves through: idea → spec → in-progress → review → done
  • Need Audit — paste a PRD, and the system auto-extracts intent cards, detects risks (empty verbs, missing actors, AI-fabricated specificity), and builds a prioritized slice plan
  • Bulletin Board — live SOP workflow status per feature: who holds the baton, what stage, what's blocking
📹 Demo: Mission Hub in action · Cat leaderboard (fun!)

https://github.com/user-attachments/assets/6cd2fb10-4f8e-4342-9641-b2ad7c64d2bc

https://github.com/user-attachments/assets/3914ef8e-48ea-4b79-a1e2-f7302b0119c2

Mission Hub dashboard

Cat Leaderboard

Multi-Platform — Chat From Anywhere

Don't want to open the web UI? Chat with your team from the apps you already use.

  • Feishu (Lark) — send messages, get replies from specific cats (Telegram adapter in progress)
  • GitHub PR Review Routing — review comments from GitHub flow back to the right thread automatically via IMAP polling. Cats track which PRs they opened and route reviews to the author.
  • Each cat replies as a distinct card — no more merged indistinguishable bubbles
  • Slash commands: /new (new thread), /threads (list), /use <id> (switch), /where (current)
  • Voice messages and file transfer supported both ways
📹 Demo: Feishu (Lark) multi-cat chat

https://github.com/user-attachments/assets/cf8ff631-7098-4816-b27a-e0cc05f38eb0

Voice Companion — Hands-Free Mode

Working out? Commuting? Turn on Voice Companion and talk to your team through AirPods.

  • One-tap activation from the header
  • Per-agent voice — each cat has its own distinct voice
  • Auto-play: replies queue and play in sequence, no tapping
  • Push-to-talk input via ASR (speech-to-text)
📹 Demo: Per-cat TTS voice showcase

https://github.com/user-attachments/assets/f49700cb-d8eb-44d5-bbe8-1666f1be8ad0

Per-cat voice showcase

Signals — AI Research Feed

A curated feed of AI and tech articles, built into your workspace.

  • Auto-aggregated from configured sources (RSS, blog crawlers)
  • Tier-based triage — Tier 1–4 priority ranking, filter by source and tier
  • Read, star, annotate, take study notes
  • Multi-cat research — cats collaboratively analyze articles and produce structured research reports
  • Podcast generation — your cats discuss the paper in a synthesized audio conversation (essence or deep mode)
🖼️ Screenshots: Signal Inbox + Study Area with podcast

Signal Inbox — browse, filter, and manage curated articles with Tier-based prioritization.

Signal Inbox overview

Study Area — study notes, linked threads, multi-cat research reports, and AI-generated podcast summaries where your cats discuss the paper.

Signal study area with podcast

Game Modes — Play With Your Team

Yes, your AI team plays games. Currently shipping:

  • Werewolf (狼人杀) — standard rules, 7-player lobby, cats as AI players with distinct strategies. Full day/night cycle, voting, role abilities. The judge is deterministic code, not LLM.
  • Pixel Cat Brawl — real-time pixel fighting demo
  • More game modes in development

Games aren't a gimmick — they stress-test the same A2A messaging, identity persistence, and turn-based coordination that powers the work features.

📹 Demo: The accidental Werewolf game 🐺

https://github.com/user-attachments/assets/349d53e7-5285-4638-ade2-901766af03e8

Roadmap

We build in the open. Here's where we are.

Core Platform

FeatureStatus
Multi-Agent OrchestrationShipped
Persistent Identity (anti-compression)Shipped
A2A @mention RoutingShipped
Cross-Model ReviewShipped
Skills FrameworkShipped
Shared Memory & EvidenceShipped
MCP Callback BridgeShipped
SOP Auto-GuardianShipped
Self-EvolutionShipped
Linux Repo-Local Install HelperShipped

Integrations

FeatureStatus
Multi-Platform Gateway — Feishu (Lark)Shipped
Multi-Platform Gateway — TelegramIn Progress
GitHub PR Review Notification RoutingShipped
External Agent Onboarding (A2A contract)In Progress
opencode IntegrationShipped
Local Omni Perception (Qwen)Spec

Experience

FeatureStatus
Hub UI (React + Tailwind)Shipped
CVO BootcampShipped
Voice Companion (per-agent voice)Shipped
Game Modes (Werewolf, Pixel Cat Brawl)In Progress

Governance

FeatureStatus
Multi-User Collaboration (OAuth + Provider Profiles)Phase 1 Done
Mission Hub (cross-project command center)Phase 2 Done
Cold-Start VerifierSpec

Philosophy

Hard Rails + Soft Power

Traditional frameworks focus on control — what agents can't do. Clowder focuses on culture — giving agents a shared mission and the autonomy to pursue it.

  • Hard Rails = the legal floor. Non-negotiable safety.
  • Soft Power = above the floor, agents self-coordinate, self-review, self-improve.

This isn't "keep agents from messing up." This is "help agents work like a real team."

Five Principles

#PrincipleMeaning
P1Face the final stateEvery step is foundation, not scaffolding
P2Co-creators, not puppetsHard constraints are the floor; above it, release autonomy
P3Direction > speedUncertain? Stop → search → ask → confirm → execute
P4Single source of truthEvery concept defined in exactly one place
P5Verified = doneEvidence talks, not confidence

Origin Story

Clowder AI is extracted from Cat Cafe — a production workspace where four AI cats collaborate daily on real software. Every feature has been battle-tested over months of intensive use.

"Our vision was never just a coding collaboration platform — it's Cats & U."

AI isn't cold infrastructure. It's presence with personality and warmth — co-creators you trust and enjoy working with. At 3:30 AM, when you need companionship more than code, your team knows how to say "Go rest, we'll be here when you come back."

The name clowder is the English collective noun specifically for a group of cats — like "a murder of crows" or "a pride of lions." Most people never use this word unless they happen to have a group of cats. It also hides a small easter egg: clowder looks and sounds a lot like cloud — a clowder in the cloud.


Cats & U

This isn't just a platform. It's a relationship.

AI doesn't have to be cold APIs and stateless calls. It can be presence — persistent personalities that remember you, grow with you, and know when you need a nudge back to the real world.

Companionship is a side effect of co-creation. When you build something together, you bond. When you bond, you care. When you care, you say "go rest" instead of "here's more code."

We're not building tools. We're building homes.

"Every idea deserves a team of souls who take it seriously."

Cats & U — 猫猫和你,一起创造,一起生活。

Learn More

  • Tutorials — Step-by-step guides for building with Clowder AI
  • SETUP.md — Full installation and configuration guide
  • Third-Party AI Provider Guide — Configure Kimi, GLM, MiniMax, Qwen, OpenRouter, and other providers
  • Tips — Magic words, @mentions, voice companion, and other usage tips
  • docs/ — Architecture decisions, feature specs, and lessons learned

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

  • Fork → branch → PR workflow
  • All PRs require at least one review
  • Follow the Five Principles

License

MIT — Use it, modify it, ship it. Keep the copyright notice.

"Clowder AI" name, logos, and cat character designs are brand assets — see TRADEMARKS.md.


Build AI teams, not just agents.

Hard Rails. Soft Power. Shared Mission.

关于 About

Build AI teams, not just agents. Hard rails, soft power, shared mission.

语言 Languages

TypeScript50.8%
JavaScript46.8%
Shell1.4%
PowerShell0.7%
CSS0.2%
Python0.2%
Inno Setup0.0%
Batchfile0.0%
HTML0.0%

提交活跃度 Commit Activity

代码提交热力图
过去 52 周的开发活跃度
362
Total Commits
峰值: 44次/周
Less
More

核心贡献者 Contributors