Open-source project
zts212653/clowder-ai avatar
zts212653/clowder-ai

clowder-ai: the root package is still cat-cafe at version 0.1.0

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

3,878 stars926 forksTypeScriptMIT

At a glance

What is it?
clowder-ai is a platform layer that sits above four agent command line tools and gives them identity, shared memory, review passes and message passing. It was extracted from an internal project and never renamed underneath, ships installers for two operating systems, binds its API to loopback by default, and documents a flag that trusts the whole private network.
Who is it for?
clowder-ai fits someone who already runs more than one agent command line tool and wants them to share state and hand work to each other rather than copy-pasting context, and its security defaults are better than most self-hosted tools manage. Verify four things before you run it.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository received new commits within the last day.
What is it written in?
Mainly TypeScript, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • zts212653/clowder-ai3,878 ★ · +677
Star history
Repository2026-09-282026-11-15
zts212653/clowder-ai3.2k3.9k

677 Stars in the last 48 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The root package is still named cat-cafe at version 0.1.0

The repository, the product and the build manifest disagree about what this is. The repository is clowder-ai, the newest release is version 0.13.0, and the licence is permissive with installers attached. The root package manifest says the name is cat-cafe, the version is 0.1.0, the package is private, and the description is a fourth wording of the same pitch. Every workspace package in the scripts is scoped to the old name as well. The origin section explains why: the project was extracted from an internal workspace of that name, where four AI cats collaborated on real software. The extraction moved the product identity and left the build identity behind. Nothing is broken by it, and anyone reading the scripts will recognise the old name immediately, but the two identifiers have to be reconciled by hand in a fork.

Installers for two operating systems, and a setup guide missing one language

The quick start offers a desktop installer from the releases page for two platforms, one Windows executable and one macOS disk image, and states that it bundles everything so no package installation is needed. There is no Linux installer, which is worth noting for a Node and package-manager project that runs on Linux perfectly well and whose own badge row links to a Linux community forum. The source route is the other path, and it is short:

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

Requirements are stated in the same breath, including a cache server as an optional component that can be skipped in favour of an in-memory mode. Translation coverage is uneven in a similar way. The main file offers English, Chinese and Japanese, and the setup guide exists in English and Chinese with no Japanese version. The badge row has its own looseness: six links sit in badge position, and three of them point at the licence file, the contributor guide and a forum topic rather than at anything that renders as a status badge.

The API binds to loopback, and the file explains how to stop doing that

The environment example is bilingual down to every comment, which is unusually thorough. The port section names the frontend and API ports and records a convention that the API port is one above the frontend port, noting that an internal pair uses the same relationship at lower numbers. So four ports are in play rather than two. The host section defaults the API to localhost and says in plain words to set it to the all-interfaces address for network, overlay-network or container access. Right below that sits a commented-out cross-origin flag for letting phones and tablets on private ranges call the API, and its own comment warns that it trusts every device on the private network and that a restart is required. The default is right. The documentation of how to widen it is equally clear, which means the widening is a one-line decision.

The owner identity falls back to a shared literal, and plugin installs are local-only

There is a section on owner identity for privileged operations, and it is the most carefully written part of the file. Three rules are set out. The default identity selection needs no configuration and falls back to a literal shared user name, so a fresh install has no identity separation at all. Writes to sensitive environment variables require an owner identity to be set explicitly and return a forbidden status otherwise. And installing or deleting capabilities is restricted to a direct local connection with no owner needed, with that write path disabled entirely when the service is bound to the network instead. That last rule is the strongest thing in the file: the most dangerous operation, adding and removing plugins, is only reachable from the machine it runs on. The gap is the shared fallback, since the gate is whether a value was changed rather than whether a caller proved anything.

Orchestration is described as a fixed role table

The pitch is that combining several models normally turns you into the router, copying context and tracking who said what. The capability table then describes routing with a fixed assignment: one model for architecture, another for review, a third for design, and a review capability whose description is one model writing code while another reviews it, presented as built in rather than added on. That is a sensible default and a template, not the dynamic selection the framing implies. Four agent command line tools are listed, all with protocol support and all marked shipped with no beta distinction, and the file is explicit that the platform does not replace them but sits above them. The architecture diagram names seven platform components, and one of them appears nowhere else in the file: nothing explains what the SOP guardian does, though a directory of SOP definitions exists in the tree.

Seven process directories, three provenance dotfiles and three merge gates

The repository root has forty-five entries, and a large share of them are process rather than code. There are seven directories for specifications, discussions, review notes, guides, documentation, SOP definitions and skills, plus three dotfiles that record a runtime revision, a sync provenance record and a directory exceptions list. Three separate mechanisms gate a commit: a pre-merge check script, a merge execution script, and a set of git guards you install from a script. There are two worktree-based workflows with their own init, start and stop commands and a third script that tests the worktree script itself, and a template file for creating a new instance of the workspace. Even the test and task names carry ticket identifiers. Set against that, the full platform walkthrough offered at the top of the file runs for three minutes and forty-five seconds.

The licence section is three imperatives, and the brand is carved out

The licence section does not reproduce a licence notice. It names the permissive licence and then offers three verbs: use it, modify it, ship it. That is a summary, and a permissive licence requires its notice to travel with copies and substantial portions, so the notice lives in the file that the badge row links to rather than on the page describing it. Separately, a trademark file reserves the product name, the logos and the cat character designs as brand assets. That is a common and sensible split: the code is permissive, the name and artwork are not. It does mean a fork cannot simply drop the branding, and it means the artwork in the assets directory is not covered by the code licence even though it sits in the same repository.

Editorial conclusion

clowder-ai fits someone who already runs more than one agent command line tool and wants them to share state and hand work to each other rather than copy-pasting context, and its security defaults are better than most self-hosted tools manage. Verify four things before you run it. Set an owner identity, because the fallback is a shared literal and the only thing standing between a default install and a privileged environment write is whether you changed one value. Leave the API on loopback unless you genuinely need a phone on the network to reach it, and if you do, understand that the accompanying cross-origin flag trusts every device on that network. Read the licence section's three imperatives as a summary, not as the licence, because the notice the licence requires is in the file it links. And if you fork it, expect the manifests to still say cat-cafe, which is a rename you will have to do yourself.

Frequently asked questions

What is zts212653/clowder-ai?

A platform layer that sits above agent command line tools rather than replacing them. It gives each agent a persistent identity, memory and role across sessions, routes work between them, passes messages agent to agent with mention routing, keeps a shared store of evidence, lessons and decision logs, and can ask one model to review another's output. It supports four agent tools and is extracted from an internal project of a different name.

How do I install clowder-ai?

Either download a desktop installer from the releases page, which is published for Windows and macOS only and bundles everything, or build from source with the package manager, which needs Node 20 or newer, version 9 or newer of the package manager, and optionally a cache server that can be skipped in favour of an in-memory mode.

Does clowder-ai bind its API to my whole network?

Not by default. The API host defaults to the loopback address and the file says to change it to the all-interfaces address only for local network, overlay network or container access. Immediately below it is a commented-out cross-origin flag for private-range devices, and that flag's own comment warns that it trusts every device on the private network.

What does clowder-ai use an owner identity for?

Three things. Default identity selection works without one and falls back to a shared literal user name. Writes to sensitive environment variables require it to be set explicitly and are refused otherwise. Installing or deleting capabilities is restricted to a direct local connection with no owner needed, and that write path is disabled entirely when the service is bound to the network.

What license is clowder-ai under?

The permissive MIT licence, summarised on the page as three imperatives rather than reproduced as a notice, with the licence file itself linked from the badge row. Separately, a trademark file reserves the product name, the logos and the cat character designs as brand assets, so those are not covered by the code licence.

Official sources

  1. Issues
  2. License: MIT
  3. README
  4. Releases
  5. zts212653/clowder-ai on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/zts212653-clowder-ai.svg)](https://hysenlabs.com/projects/zts212653-clowder-ai)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/zts212653-clowder-ai.svg)](https://hysenlabs.com/projects/zts212653-clowder-ai)

Was this useful?