Clowder AI: a platform layer that makes Claude, Codex and Gemini work as one team

The problem Clowder AI solves, and who it is actually for

The README opens with a specific complaint: you have Claude, GPT and Gemini, each good at something different, and using them together makes you the router. You copy context between windows, remember which model said what, and spend the afternoon on coordination instead of the task. Clowder AI is the answer to that, and the README states it plainly: most frameworks help you call agents, Clowder helps them work together.

The target user is narrower than the tagline suggests. You need to already be running at least two agent CLIs from the supported list: Claude Code, Codex CLI, Gemini CLI or opencode. If you use one model, the platform layer has nothing to route. The README also says Clowder does not replace your agent CLI, it sits above it, so the tool assumes you are comfortable installing and authenticating those CLIs yourself. This is infrastructure for someone who has already felt the pain of being middle management between models.

How the platform layer works: identity, A2A routing and shared memory

The architecture diagram in the README shows one operator at the top, a Clowder Platform Layer in the middle, and the four agent CLIs at the bottom. The layer holds six named pieces: Identity, A2A Router, Skills, Memory, SOP Guardian, MCP Bridge and Plugins. The README's closing line for that diagram is the design intent in one sentence: models set the ceiling, the platform sets the floor.

Concretely, the capability table lists what each piece does. Multi-agent orchestration routes a task to the right agent, with Claude suggested for architecture, GPT for review and Gemini for design. Persistent identity means each agent keeps its role, personality and memory across sessions, so you are not re-priming it every morning. Cross-model review is described as built in rather than bolted on: Claude writes code, GPT reviews it. A2A communication is asynchronous agent-to-agent messaging with @mention routing and structured handoff. Shared memory is an evidence store, lessons learned and decision logs. The plugin framework covers MCP tools, IM adapters for Feishu and Telegram, and on-demand skills.

The repository layout backs this up. There are directories for sop-definitions, cat-cafe-skills, feature-specs and packages, plus a pnpm workspace file. The package.json is named cat-cafe, which matches the README's origin note: the project was extracted from Cat Cafe, described as a production workspace where four AI cats collaborate daily on real software. The name clowder is the collective noun for a group of cats.

Installing Clowder AI and getting a first run on localhost:3003

There are two paths. The README points to the Releases page for a desktop installer available for Windows (.exe) and macOS (.dmg), and states the installer bundles everything so no pnpm install is needed. That is the shorter route and the one to take if you do not intend to modify the code.

From source, the README gives this sequence. It requires Node.js 20+ and pnpm 9+, and Redis 7+ is optional.

git clone https://github.com/zts212653/clowder-ai.git
cd clowder-ai
pnpm install
pnpm start        # opens http://localhost:3003

After pnpm start, the README says a browser window opens at http://localhost:3003. The README also lists two flags worth knowing before you wait through a rebuild: pnpm start --quick skips the rebuild, and pnpm start --daemon runs the process in the background.

Redis is where the first real decision sits. The README marks Redis 7+ as optional and mentions a --memory flag to skip it. The .env.example file shows the default ports: frontend 3003, API 3004 and Redis 6399, with the convention that the API port is the frontend port plus one. Copy the example file before your first run if you need to change anything.

cp .env.example .env

The same file documents the binding behaviour. API_SERVER_HOST defaults to 127.0.0.1, so the API is localhost only. Setting it to 0.0.0.0 exposes the API to your LAN, Tailscale or Docker, and the comment warns that this trusts the entire private network. If you do that and want phones or tablets on RFC 1918 networks to call the API, CORS_ALLOW_PRIVATE_NETWORK=true is the second switch, and the file notes a restart is required after changing it.

The owner identity rule and the LAN mode trade-off

The .env.example file carries a constraint that deserves more attention than the README gives it. Owner identity is optional for choosing a default cat, which falls back to default-user. But sensitive environment variable writes require it to be explicitly set, and return 403 otherwise.

Capability install and delete is stricter. The file states it is supported only through the localhost UI, and that setting API_SERVER_HOST=0.0.0.0 disables capability writes entirely. So LAN mode and capability management are mutually exclusive in the current design. If you expose Clowder AI to your network, you lose the ability to install or remove capabilities from the same instance. That is a deliberate safety choice, and it is also the kind of thing that surprises people who enable LAN access first and read the environment file later.

For multi-user deployments the file says to set owner identity to your actual user ID. There is no guidance in the README on how roles are separated between multiple owners, so treat a shared instance as an area to test before you rely on it.

Where Clowder AI is the wrong tool

The clearest limitation is in the supported agents table. Four CLIs are listed as shipped: Claude Code, Codex CLI, Gemini CLI and opencode. A model with no supported CLI has no entry point into the platform. If your work runs through a provider outside that list, the routing layer cannot help you.

The second limitation is operational. The README's quick start ends at a local URL, and the environment file is written around localhost, private networks and a local Redis. This is a self-hosted tool. There is no hosted offering described, and the desktop installer still runs the stack on your machine. If you want something that works without a local process on port 3003, this is not that.

The third is that the project is young. The published releases run from v0.11.1 in June 2026 through v0.12.0 in July to v0.13.0 in September. The package.json still carries version 0.1.0, which is a private workspace version rather than the release line. The README also contains a visible TODO comment where the logo should be replaced once synced from assets, and the main README does not document rollback or what happens to shared memory when you switch profiles. For a system whose selling point is persistent memory, the absence of documented memory lifecycle management is the gap to probe first.

How Clowder AI differs from calling agents directly

The obvious alternative is the thing you are already doing: open two terminal windows, run Claude Code in one and Codex CLI in the other, and move context between them by hand. That approach has no install step, no ports and no Redis. It also has no persistent identity, no @mention routing and no shared evidence store, and the README's framing of the problem, that you become the router, describes exactly this setup. The difference is not capability, it is who holds the coordination state. In the manual approach it lives in your head and your clipboard. In Clowder AI it lives in the platform layer.

The other alternative is a single agent CLI with its own subagent or MCP tooling. That keeps everything in one vendor's model family and avoids cross-model review entirely. Clowder AI's specific claim is the opposite: Claude writes, GPT reviews. If your reason for wanting multiple models is redundancy of judgement rather than raw capability, that cross-model review is the feature to evaluate, and the README lists it as built in rather than bolted on.

Maintenance cost, licence and what the MIT grant does not cover

The repository is not archived and the last push was on 2026-09-20, with v0.13.0 released on 2026-09-02. That is a recent release cadence, and the version numbers moved three times between June and September 2026. Upgrading means pulling the repository and rerunning pnpm install, or downloading a new desktop build. The README does not describe a migration path between releases, so plan on reading the release notes for each jump.

Running from source also means Node.js 20+ and pnpm 9+ on the machine, plus Redis 7+ unless you use --memory. Those are the ongoing costs: a Node runtime, a package manager, and an optional Redis instance on port 6399 by default.

The licence is MIT, which permits use, modification and shipping. The README adds a boundary that MIT does not erase: the Clowder AI name, logos and cat character designs are brand assets covered separately in TRADEMARKS.md. Forking the code is permitted. Rebranding with the same name and cat designs is a separate question, and TRADEMARKS.md is where the project states its position. There is also a CLA.md in the repository root, which is relevant if you plan to contribute rather than just consume. None of this is legal advice, and if you intend to redistribute a modified build under a name, read both files.

Editorial conclusion

Adopt Clowder AI if you already run two or more agent CLIs and the copy-paste routing between them has become your job. Skip it if you only use one model, or if you want a hosted service with no local ports to manage. Before committing, verify that Node 20+ and pnpm 9+ are present, decide whether Redis 7+ is worth running or whether you will start with --memory, and read SETUP.md rather than the README alone, because the README does not document rollback or what happens to shared memory when you switch profiles.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. zts212653/clowder-ai on GitHub