AI Game DevTools lists sixteen categories and fills three of the columns

Sixteen categories in the table of contents, and one table on the page

The table of contents names sixteen anchors: LLM (LLM & Tool), VLM, Game (World Model & Agent), Code, Image, Texture, Shader, 3D Model, Avatar, Animation, Video, Audio, Music, Singing Voice, Speech and Analytics. Each entry links to an anchor id, from `#tool` through `#analytics`.

The visible page then opens exactly one of them, `### <span id="tool">LLM (LLM & Tool)</span>`, and the first table ends partway down a ChatRWKV row. So the structure is announced in full and filled in one part. The other fifteen categories are named, spaced and anchored, and their rows are not on this page.

The framing above the table is worth reading as a specification of what this is. The heading is a name with an abbreviation, AI Game DevTools and AI-GDT, and the taglines ask the reader to discover a long list of media types in one sentence. Two links labelled Website and the Chinese equivalent of official site point at the identical address, `yuan-manx.github.io/ai-game-devtools/`, which is the GitHub Pages site rather than the repository.

So the practical shape is a README that mirrors a site, with the site as the fuller copy and this file as the one you can read in a terminal.

The first category mixes model weights, agent loops, Unity bridges and a game master

The eighteen rows visible under LLM (LLM & Tool) do not share a category so much as share a name. Four are language models: baichuan-7B, Baichuan-13B, Baichuan 2 and ChatRWKV. Four are agent frameworks or agent platforms: AgentGPT, AIOS, Auto-GPT and BabyAGI, the last described as an example of an AI-powered task management system. Three bridge a chat API into the Unity editor, AICommand, ChatGPT-API-unity and ChatGPTForUnity, and those three are the only rows in the sample with an engine filled in.

Then there are research systems with papers attached, AI Scientist, Character-LLM and ChatDev, and developer platforms such as Bisheng and Assistant CLI. The odd one out is Auferet, described as an AI game master for solo text adventures and tabletop-style RPGs with persistent memory of your story and your own uploaded lore.

Auferet is also the only row whose Type cell is not Tool. Everything else in the sample, including three language models and three Unity bridges, is typed Tool. So the single most game-shaped project in the category is the one the schema files under a different role, and the language models that carry no game integration at all share the same Type value as the Unity plugins.

Paper filled on three rows, Game Engine on three, Type filled on every row

The table header has five columns: Source, Description, Paper, Game Engine and Type. In the eighteen rows visible, the Type column is filled on all of them, with Tool on seventeen and Writer on the Auferet row.

The Paper column is filled on three rows, and each is an arXiv link: 2408.06292 for AI Scientist, 2310.10158 for Character-LLM and 2307.07924 for ChatDev. The Game Engine column is filled on three rows, all of them Unity. Every other cell in both columns is empty, including for baichuan-7B and for AgentGPT.

For a hub whose subject is game development, the engine column is the one that would carry the game-specific information, and it is empty for every entry that is not a Unity plugin. The Paper column, meanwhile, marks the research projects rather than the game-facing ones.

The row formatting is uneven as well. Most rows are padded into the five-column layout with wide spacing, while the Auferet row is written compactly with fewer empty cells, so it reads as a row that was added by a different route than the rest.

The Description column is the upstream project's own pitch

Read a few descriptions and the source becomes clear. They are the projects' own taglines: the AI Scientist as a project titled Towards Fully Automated Open-Ended Scientific Discovery, BabyAGI as a Python script that is an example of an AI-powered task management system, Assistant CLI as a comfortable CLI tool to use the ChatGPT service, Bisheng as an open LLM devops platform for next generation AI applications, Auto-GPT as an experimental open-source attempt to make GPT-4 fully autonomous.

That is a defensible choice for a link hub. It keeps the writer out of the entries and lets each project speak for itself. It also means the Description column carries no information the project page did not already carry, and it carries none of the things a reader would use to choose: licence, price, platform requirements, whether the tool is a model or a service, when it was last updated, or whether it still exists.

The presentation also carries its own layer. Several link labels carry emoji, one row labelled with a baby, a robot and a monitor for the BabyAGI UI, and one description ends with an emoji flame. That tone belongs to the source projects' marketing, not to the hub, and it is the only editorial voice in an otherwise neutral grid.

src/ and assets/ at the root, and no manifest that builds the site

The repository root holds five entries: `.github/`, `LICENSE`, `README.md`, `assets/` and `src/`. The project language is JavaScript and the recorded homepage is the GitHub Pages site, so there is site code in the repository.

What the root does not hold is a package manifest, a lockfile or a build configuration file. There is no `package.json` at the top level, which is the usual place where the scripts that turn `src/` into a published site would be declared. This page does not explain how the Pages site is produced, and it does not link to a deployment workflow, so the relationship between `src/` and the live site has to be read out of the repository rather than off the documentation.

The licensing question is the clean one. The licence is MIT and the LICENSE file sits at the root, so the hub itself is permissively licensed. Note the asymmetry that creates: MIT covers this list of links, while the licences of the eighteen projects it points at are not recorded anywhere in the table, so nothing on this page tells you what you may do with the tools it recommends.

Seventeen of eighteen links point at GitHub, and one at a vendor domain

Seventeen of the eighteen visible rows link to a `github.com` repository, from reworkd/AgentGPT through BlinkDL/ChatRWKV. The exception is Auferet, which links to `auferet.com`, a vendor domain rather than a source repository.

That single row is also the only one with a Type of Writer, the only one whose description is written in the second person about your own story and your own uploaded lore, and the only one whose row is formatted differently from its neighbours. It reads as a product listing rather than an open-source entry, and the table has no column that would let a reader see the difference deliberately.

The repository itself has 1,353 stars, 132 forks and 14 open issues. The fork count relative to stars is what you would expect from a resource list that people clone or copy rather than depend on, and the open issue count is the only place the page gives a signal about how the list is maintained, which is not a lot of signal.

No releases, no per-row dates, and a link list ages in place

There are no GitHub releases for this repository, and no version tags are shown. The default branch is `main`, and the most recent push is dated 2026-07-21. There is nothing to pin and nothing to diff between, which suits a document rather than a package.

The real risk with a link hub is not that the code breaks, it is that the links rot. Nothing in the table records when an entry was added, when it was last checked, or what version it pointed at. There is no archived marker for a project that has been renamed or discontinued, and no column that would distinguish a maintained project from a demo. A row can therefore sit unchanged for a year and look identical to a row added last week.

The three columns that would push back against that, last verified, licence and platform, are the three the table does not have. A `.github/` directory is present at the root, but this page says nothing about how an entry is submitted, reviewed or retired, so there is no documented path for a reader who spots a dead row to report it, other than opening an issue.

Editorial conclusion

Use this repository as a bookmark shelf, not as a shortlist. It is at its best when you want to know that a category exists at all, for instance that Singing Voice and Speech are treated separately from Music, or that the first category already mixes language models with Unity editor bridges. It cannot help you choose between two entries, because the table has no licence, price, platform, version or last-checked column, and the description text is copied from each project rather than written about it. Two checks are worth doing before you rely on it. Confirm that the tool you picked still supports your engine and your licence terms yourself, and if you intend to keep a copy of the list, remember that the repository has no releases and the default branch carries everything. Submitting a missing tool is also undocumented, so an omission here is not evidence that a tool does not exist.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. Yuan-ManX/ai-game-devtools on GitHub