What is modu-ai/moai-cowork?

MoAI-Cowork ships 17 Korean-language coworker plugins behind a single /project command, with manifests for both desktop apps. It is aimed at non-developers, and the install path is a marketplace address rather than a terminal.

Can I use modu-ai/moai-cowork commercially?

Yes. Apache-2.0 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 modu-ai/moai-cowork still maintained?

Yes. The repository last received commits 13 days ago.

What programming language is modu-ai/moai-cowork written in?

Mainly Python, according to GitHub's language statistics.

MoAI-Cowork: Korean AI coworker plugins for Claude and ChatGPT

At a glance

What is it?: MoAI-Cowork ships 17 Korean-language coworker plugins behind a single /project command, with manifests for both desktop apps. It is aimed at non-developers, and the install path is a marketplace address rather than a terminal.
Who is it for?: Adopt MoAI-Cowork if your team works in Korean and your users live inside Claude Cowork or ChatGPT Work rather than a shell; the marketplace install and the /project router remove the blank-prompt problem that stalls most non-developer users. Do not adopt it if you need a Python library to import, if your deliverables are not Korean-market documents, or if you expect the repository to be a general agent framework rather than a plugin catalogue.
Can I use it commercially?: Yes. Apache-2.0 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 last received commits 13 days ago.
What is it written in?: Mainly Python, according to GitHub's language statistics.

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

DEEP OPEN-SOURCE ANALYSIS

Hysen Labs Editorial · September 15, 2026 · modu-ai/moai-cowork

The blank-prompt problem MoAI-Cowork is built around

Most AI tooling assumes you already know what to ask. MoAI-Cowork starts from the opposite premise. The README describes its target user as someone who is not comfortable with English-language menus and developer-oriented interfaces, who has to think about what to type each time, and who needs Korean working context: Hangul documents, SmartStore, public data, Korean speech levels. That is a narrower audience than "knowledge workers" and a more specific one than "Korean users".

The design answer is a family of job-specific agents instead of one general assistant. The README compares it to a company with departments, where each department has its own owner. The roster covers marketing, writing, media, commerce, legal, finance, HR, customer service, design, data analysis, recruiting and more, all installed from the single modu-ai/moai-cowork marketplace. The README states that the authoritative roster lives in .claude-plugin/marketplace.json and that its own table is a role summary, which is worth remembering: the table is illustrative, the manifest is the contract.

The entry point is a PM coworker, moai-pm, which the README describes as a router rather than a worker. It does not produce deliverables itself. It interviews you, works out which coworkers match, and assembles them into a chain.

How the /project router and skill chains actually work

The mechanism has two phases. On first run, /project starts an interview about your goals and deliverables. From the answers the PM detects the work involved and assembles a chain: the relevant coworkers plus their skills, in order. After you confirm, it writes project rules. The README lists three outputs from that final stage: a project instructions file, project-specific custom agents and a skill chain design, and, where needed, guidance for external service API keys.

The instructions file has a detail worth noticing. AGENTS.md is the canonical document, and CLAUDE.md contains only a single import line pointing at it. So the two desktop apps are not maintained as parallel copies of the same rules; one file is the source and the other is a pointer. That is a sensible choice for drift, and it also means anyone editing CLAUDE.md directly is editing a file that is not the source of truth.

After setup, the second phase is natural language. The README's example is that saying "write me a business plan" causes the matching skill to be selected automatically, with no slash command and no skill name to memorise. The README illustrates the chain as content generation, then document conversion, then quality review, where conversion handles PPTX generation and review handles removing the tone typical of AI writing. The PM's role is explicitly limited to finding who fits the job and forming the team.

What the README does not document is how conflicts are resolved when two skills match the same request, or what happens when a chain fails partway through. That is a real gap for anyone planning to rely on this in a production workflow.

Installing MoAI-Cowork: marketplace address, not pip

There is no terminal step. The README states that installation happens entirely inside the desktop app with mouse clicks, and that both Claude Cowork and ChatGPT Work use the same marketplace address, modu-ai/moai-cowork. The repository ships both manifest types, .claude-plugin and .codex-plugin, so the same catalogue serves either app.

Registration happens once per computer. In the app, go to Settings or Plugins, then Marketplace, and add the repository with the + control.

bash

modu-ai/moai-cowork

That string is the address you enter in the marketplace field, not a command to run in a shell. When synchronisation finishes, the coworker list appears.

Next, add the plugins you need from the registered marketplace, each with its own + button. The README recommends starting with two: moai-pm, the entry hub that runs /project, and moai-coworker, which the README describes as holding general-purpose work including review of text deliverables. Job-specific plugins come after, for example moai-marketer for campaign and content work, or moai-lawyer and moai-officer for legal and document work.

The README notes that PM plus coworker is enough to begin, and that other coworkers are detected and offered during setup when needed. Once installed, create a project with a connected working folder and type the command in the chat window.

bash

/project

The PM greets you and asks what you want to do. Answer in plain language, for example that you are preparing an online class launch. The PM then designs the project-specific custom agents and skill chain and generates CLAUDE.md for Claude Cowork or AGENTS.md for ChatGPT Work. From then on, one natural-language line drives the work.

Two caveats from the README itself. It points to the plugin install and management page on the documentation site for exact per-app click paths and troubleshooting, which suggests the in-app menus are not identical across the two apps. And the README contains placeholder image comments for the marketplace and plugin screens, so the visual walkthrough is not complete in the repository.

Where the design gets thin

The repository is Python, but it is not a library you import. The Makefile targets are maintenance plumbing: ci-local runs a local CI mirror of lint, vet, test and cross-compile; mcp-core-sync copies a canonical MCP core into every adopting server; mcp-core-check fails if a vendored copy has drifted; mcp-core-test runs the shared core test suite with uv run pytest in plugins/_shared/moai-mcp-core. None of that is the user-facing product. If you arrived expecting a Python package to call from your own code, this is the wrong repository.

The dependency on two specific desktop apps is the sharpest limitation. The README's own framing is that you install into whichever of Claude Cowork or ChatGPT Work you normally use. There is no documented path for a headless environment, a CI job, or a web-only session. Teams that want to script coworker behaviour into an existing pipeline have nothing in the README to build on.

The /project router is also a single point of interpretation. The README does not document rollback, so if the generated AGENTS.md or CLAUDE.md encodes the wrong assumptions about your project, the documented recovery path is not stated. Since AGENTS.md is canonical and CLAUDE.md is one import line, a mistake propagates to both apps at once.

Finally, the Korean-market fit is the selling point and the boundary. Hangul office documents, SmartStore, public data and Korean speech levels are exactly what the roster is built for. If your deliverables are in English, or your commerce runs on Shopify rather than SmartStore, the specialised coworkers lose most of their advantage and you are left with a plugin catalogue.

How this differs from a general agent framework

The obvious comparison is a general-purpose agent framework where you define your own agents, tools and orchestration. MoAI-Cowork inverts that. The agents are pre-built and opinionated by job function, the orchestration is a router that interviews you, and the extension surface is a plugin manifest rather than a Python API. You gain a working team on day one without designing anything. You give up the ability to shape agent behaviour at the code level, because the README describes the customisation path as project instructions files and generated skill chains, not as a framework you program against.

A second comparison is a plain chat session with a single general assistant. The difference is memory of role and sequence. A general assistant answers each prompt fresh; MoAI-Cowork writes project rules once and then routes each request through skills that were selected for that project. The README's PPTX conversion and AI-tone review steps are the concrete illustration: those are pipeline stages, not answers.

The repository's own structure supports the plugin-catalogue reading rather than the framework reading. The top level holds .claude-plugin, .codex, plugins, www, scripts and reports, plus AGENTS.md, LICENSE, LICENSE-OUTPUT.md, NOTICE and TRADEMARKS.md. That is the layout of a distributed content and plugin product with a documentation site, not of a library with a public API surface.

Licence, output rights and the cost of keeping up

Version 1.1.0 is described in the release notes as the switch to Apache-2.0 with commercial use of outputs fully permitted. The repository carries both LICENSE and a separate LICENSE-OUTPUT.md, and the README states plainly that results are yours with no commercial use restriction. Two separate documents for code and output is a deliberate split, and it means you should read both rather than assuming the Apache-2.0 grant covers generated deliverables. There is also a TRADEMARKS.md at the top level, so the name and marks are governed separately from the code. None of this is legal advice; the point is that the licence surface here is three files, not one.

The maintenance picture from the release history is concentrated. v1.0.0, v1.1.0 and v1.2.0 all landed on 2026-08-10, and v1.2.0 is described as adding an open source credits page, an MCP guidance section, and a rename of the connecting tools to moai-mcp-*. The last push to the repository was on 2026-09-03. Three releases on one day suggests a version renumbering event, which the v1.0.0 note confirms: the plugin versions were unified at 1.0.0 and a design system v2 was introduced.

The upgrade cost to watch is the moai-mcp-* rename in v1.2.0. If you had configured MCP connections under the old names, that release is the one that touches your setup. The Makefile's mcp-core-check target exists precisely to catch drift between vendored copies and the canonical source, which tells you the project treats MCP core consistency as a real failure mode rather than a theoretical one.

Editorial conclusion

Adopt MoAI-Cowork if your team works in Korean and your users live inside Claude Cowork or ChatGPT Work rather than a shell; the marketplace install and the /project router remove the blank-prompt problem that stalls most non-developer users. Do not adopt it if you need a Python library to import, if your deliverables are not Korean-market documents, or if you expect the repository to be a general agent framework rather than a plugin catalogue. Verify first: the current roster in .claude-plugin/marketplace.json, which the README names as the source of truth over its own table, and whether the two desktop apps you actually use both appear in the .claude-plugin and .codex-plugin manifests. Then read LICENSE and LICENSE-OUTPUT.md together, because the Apache-2.0 code grant and the output grant are separate documents with separate scopes.

Frequently asked questions

Do I need to use a terminal to install MoAI-Cowork?

No. The README states that installation happens entirely inside the desktop app with mouse clicks, and that no terminal is required. You register the marketplace address modu-ai/moai-cowork once per computer and then add plugins with the + button.

Does MoAI-Cowork work in both Claude Cowork and ChatGPT Work?

Yes. The repository ships both .claude-plugin and .codex-plugin manifests, and the README says both desktop apps use the same marketplace address. You install into whichever one you normally use.

What does the /project command in MoAI-Cowork actually do?

It starts the PM coworker, which interviews you about your goals and deliverables, detects the work involved, and assembles the matching coworkers and skills into a chain. After you confirm, it generates project instructions and custom agents for that project.

Which MoAI-Cowork plugins should I install first?

The README recommends moai-pm, the entry hub that runs /project, together with moai-coworker, which covers general-purpose work including review of text deliverables. Job-specific coworkers can be added later, and the README says setup detects and offers them when needed.

Can I use MoAI-Cowork outputs commercially?

The README states that results are yours with no commercial use restriction, and the v1.1.0 release notes describe the move to Apache-2.0 with commercial use of outputs fully permitted. The repository keeps LICENSE and LICENSE-OUTPUT.md as separate documents, so read both.

Is MoAI-Cowork a Python library I can import?

No. The repository is written in Python and its Makefile targets handle CI and shared MCP core maintenance, but the product is a plugin family installed into a desktop app. There is no documented library API for calling coworkers from your own code.

Official sources

Community notes