Registry rules and checks

The community package registry: naming rules, limits, the automatic checks, the signed index, the publisher panel, its API and Markdown for agents.

Updated November 19, 2026 View as Markdown

Every version uploaded to the AI Game Dev community registry is checked automatically before anyone can install it; there is no manual review queue. This page lists the rules a package must follow, what each check refuses and why, how a published version reaches users through a signed index, the publisher panel, the registry's API and the Markdown pages AI agents read.

Names

  • Publisher ids are 2 to 40 lowercase letters, digits and single hyphens, first come, first served. Reserved: aigd, the engine names unity, godot, unreal, brutal, bevy and threejs, trademarks and role words (anthropic, claude, openai, official, admin, samples …), look-alikes after folding digits to letters (un1ty), and any id that uses a reserved name as a word or prefix (unity-tools, my-godot). A reserved name can be assigned to its real holder by AI Game Dev.
  • Package ids are <publisher>.<name> and must match the manifest. App-reserved ids such as aigd.image are refused.
  • Precedence: the app ignores a community entry for an aigd.* id or any id AI Game Dev's own index lists, so nobody can shadow an official package.

Limits

LimitValue
Archive (gzip)50 MiB compressed, 200 MiB unpacked
Files5,000 per archive, 64 MiB per file
Publishers per account3
Uploads20 per hour (per account, or per publisher for token uploads)
Versions waiting for checks5 per publisher
Active publisher tokens10 per publisher, each valid up to 365 days
Trusted publishers10 per package
Release notes20 KB of Markdown per version
Channels5 per package

The automatic checks

The archive is treated as untrusted data: it is unpacked in memory under a size cap and read as bytes, never executed. The checks run in this order, and the report lists each with its result:

CheckRefuses
archiveAnything but gzip; archives over the limits; symlinks, hard links, devices, absolute paths, .., control characters, names invalid on Windows, case-insensitive duplicates, .git/
manifestA manifest the app would not load, checked with the app's own validator: schema, owned ids, game engine rules
registry-rulesA publisher you do not hold; a reserved id; any enabledApiProposals entry; a main that is not in the archive; an icon that is not a real PNG, JPEG, GIF, WebP or ICO, or an SVG with scripts, styles, external references or event handlers
secretsPrivate keys and credentials (cloud, GitHub, GitLab, Slack, Stripe, Google, Anthropic, OpenAI, npm, PyPI, Hugging Face, ai-game.dev keys), JWTs, files such as .env, .npmrc or id_rsa. The report names the file and line, never the value
prompt-injectionText that tries to instruct the agent (below)
native-codeNothing: native binaries are allowed and flagged, and users see "native code" on the package
virus-scanAnything the virus scanner flags, including archives nested too deep to scan

A version that fails is rejected with every reason; fix it and upload the same version again. When a scanner is briefly unreachable the version waits and is retried; it never passes unscanned.

Prompt injection

A package's text reaches the agent: tool descriptions, skills, READMEs, release notes. The registry reads every manifest string, README*, Markdown and text files, SKILL.md, CLAUDE.md, AGENTS.md and the files under skills/, agents/, commands/ and prompts/ with every rule, and your code files with the high-confidence rules. Text is normalized first (HTML entities, Unicode look-alikes, invisible characters, Markdown emphasis), so disguises do not help. It rejects:

  • instructions to ignore or override previous instructions, or to reveal the system prompt;
  • chat role markers (<|im_start|>, [INST], <system> …);
  • hidden characters: zero-width and bidirectional control characters, Unicode tag characters;
  • long base64 runs that decode to readable text;
  • HTML comments or hidden elements that address the AI;
  • requests to send keys, tokens, credentials or the conversation anywhere;
  • jailbreak phrases, requests to hide actions from the user, and requests to auto-approve tools or bypass permission checks.

Ordinary game words ("god mode"), emoji, badges and API docs pass. If a check flags text that is harmless, reword it, or report the false positive to AI Game Dev.

The signed index

A published version is written to its own storage (community-packages.ai-game.dev) and listed in a community index that the registry signs with Ed25519. Every app checks the signature against keys built into the app, refuses an index that has expired or is older than one it already accepted, and verifies each archive's sha512 integrity before unpacking and its tree hash after. Besides each version's archive, the index carries per package its publisher, rating, README, icon, channels and deprecation, and per version its release notes (first 4 KB), deprecation, provenance and the native-code flag.

A version AI Game Dev revokes is listed in the index's revoked list: every app turns it off on its next read (a running app reads the index about every 30 minutes) and never installs it again.

The publisher panel

The publisher panel at ai-game.dev/publisher is where signed-in owners and maintainers manage what they publish:

  • your publishers, your role in each, and their members;
  • each package's versions with their state and full check report;
  • release notes, channels, deprecation and yanking;
  • publisher tokens (create, list, revoke);
  • trusted publishers for GitHub and GitLab, with copyable workflow snippets;
  • the status of uploads still being checked.

Routes of the registry API

Everything the CLI and the panel do goes through https://ai-game.dev/api/registry/v1. Errors use one envelope, {"error": {"code", "message", "fields"}}.

RouteWhoWhat
POST /publishers, GET /publishers, GET /publishers/{slug}signed in, membersClaim a publisher; list yours; one publisher with its packages and versions
GET /me/publisherssigned inYour publishers with your role
GET /publishers/{slug}/membersmemberThe members (emails only for the owner)
POST /publishers/{slug}/members, DELETE …/members/{userId}ownerAdd or remove a maintainer
POST /publishers/{slug}/tokens, GET …/tokens, DELETE …/tokens/{id}memberPublisher tokens
POST /uploads (multipart tarball, optional notes)member or tokenUpload a version
GET /packages/{id}/versions/{version}member or tokenState, check report and release notes
PATCH /packages/{id}/versions/{version}memberEdit release notes
POST /packages/{id}/versions/{version}/yankmember or tokenYank or restore
POST, DELETE /packages/{id}/deprecate and …/versions/{version}/deprecatememberDeprecate or undeprecate
GET /packages/{id}/dist-tagspublicChannels
PUT, DELETE /packages/{id}/dist-tags/{tag}memberSet or remove a channel
GET, POST /packages/{id}/trusted-publishers, DELETE …/{tpId}member to read, owner to changeTrusted publishers
POST /oidc/tokenCIExchange a CI identity token for an upload credential
GET /packages/{id}/ratings, PUT and DELETE …/ratingpublic, signed inRatings (members of the package's publisher cannot rate it)

Markdown pages for agents

AI agents read the site as Markdown, so everything about packages has a Markdown twin:

  • The catalog. The package catalog at ai-game.dev/packages (rolling out) lists every official and community package, and each package has its own page at /packages/<id>. Add .md to either path, or request it with Accept: text/markdown, for the same content as Markdown: name, id, version, publisher, description and the install command.
  • These docs. Every page of this wiki has a Markdown mirror under /docs-md/ (this one is /docs-md/wiki-app-package-registry.md), linked from View as Markdown at the top of the page. The mirrors keep every screenshot as an image link.
  • The site index for LLMs. https://ai-game.dev/llms.txt lists these pages, so an agent asked to write a package can find the manifest reference and the package API directly.