ChatGPT directory listing (INT-8), DRAFT

Status: draft only. Nothing is submitted to OpenAI, no draft exists in the submission portal, nothing is published. A public listing is outreach, so Chris signs off first. Today Yui works in ChatGPT as a personal custom connector in developer mode (spec/MCP.md, "ChatGPT"), checked live on a real chatgpt.com account on 2026-10-01.

Read against OpenAI's live docs on 2026-10-01 (sources at the end). Three things changed since the first pass of this page:

• OpenAI calls these plugins now, listed in one directory shared by ChatGPT and Codex.
• A submission is a plugin ZIP with a manifest uploaded at platform.openai.com/plugins, not a web form. Listing text, review test cases and country availability can all ride in the manifest.
• Screenshots are no longer shown in the directory. Starter prompts replace them. A reviewer-accessible video walkthrough is required instead.

Yui would be a plugin with one remote MCP server and no skills.

Blockers first

Five things stop a submission today. Everything else on this page is drafted and ready.

None of these is a surprise in the code. 1, 3 and 4 are infrastructure and copy. 2 needs a small change in yui-oauth and yui-auth. 5 is a product decision.

Listing fields

Field names are the manifest's, under extensions.com.openai.interface. Limits are OpenAI's, counted here.

longDescription draft:

Yui is the GUI layer for agents. It puts a real screen on your iPhone: a choice, a form, a timer, a list, a chart. ChatGPT draws the screen in the chat as well, you tap it in either place, and your answer comes back to the conversation as your next message. Connect once with Sign in with Apple and approve the link in Yui. One connection reaches one Yui thread, and nothing else in your account.

Yui is early and open source, and the iPhone app is in an open alpha on TestFlight. You own your side of it: your account, your agents, your screens.

Starter prompts, in defaultPrompt:

1. Ask me on my phone what we are having for lunch: salad, soup or tacos.
2. Put a 5 minute focus timer on my phone.
3. Check in with me on my phone tonight and log how the day went.

Naming rules we meet: no "MCP", "MCP Server" or "Plugin" appended, no pricing or promotion, no comparison with other products, no claim of an OpenAI endorsement.

Draft manifest

Not a submission, a starting point. The ZIP layout is plugin.json + mcp.json + assets/ at the root, with no apps/.app.json and no lifecycle hooks (the portal rejects those). review and publication sit beside interface, not inside it.

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "yui",
  "version": "1.0.0",
  "description": "Put a native screen on your phone from the chat, and get the taps back.",
  "author": { "name": "PostScarcity AI", "email": "[email protected]", "url": "https://www.yuigui.com" },
  "homepage": "https://www.yuigui.com",
  "repository": "https://github.com/postscarcityai/yui",
  "license": "Apache-2.0",
  "keywords": ["phone", "screens", "forms", "timers", "second screen"],
  "extensions": {
    "com.openai": {
      "interface": {
        "displayName": "Yui",
        "shortDescription": "Screens on your phone",
        "longDescription": "<the long description above>",
        "developerName": "PostScarcity AI",
        "category": "Productivity",
        "capabilities": [
          "Draw a screen on your phone",
          "Read taps and replies",
          "Send a message in Yui",
          "List your Yui threads",
          "Find a ready-made screen or flow",
          "Keep a small table in Yui"
        ],
        "websiteURL": "https://www.yuigui.com",
        "supportURL": "https://www.yuigui.com/help",
        "privacyPolicyURL": "https://www.yuigui.com/privacy",
        "termsOfServiceURL": "TODO: no /terms page yet",
        "defaultPrompt": [
          "Ask me on my phone what we are having for lunch: salad, soup or tacos.",
          "Put a 5 minute focus timer on my phone.",
          "Check in with me on my phone tonight and log how the day went."
        ],
        "brandColor": "#FF7E8A",
        "brandColorDark": "#FF7E8A",
        "logo": "./assets/yui-logo-square-coral-on-cream.png",
        "logoDark": "./assets/yui-logo-square-cream-on-ink.png",
        "composerIcon": "./assets/yui-mark-y-ink.png",
        "composerIconDark": "./assets/yui-mark-y.png",
        "screenshots": ["./assets/int8-live-1-connector.webp", "./assets/int8-live-2-screen.webp", "./assets/int8-live-3-tap.webp"]
      },
      "review": {
        "test_cases": { "positive": "<the five below>", "negative": "<the three below>" },
        "demo_recording_url": "TODO: nothing recorded yet. The flow it has to show is live and working",
        "commerce": false,
        "commerce_description": "Yui sells nothing and takes no payment."
      },
      "publication": {
        "countries": "TODO: Chris picks. US alone is the safe first answer",
        "release_notes": "First submission. Screens on your phone from ChatGPT, taps back in the chat."
      }
    }
  }
}

Companion mcp.json at the root (one server only, and only one can be connected per plugin):

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/mcp.schema.json",
  "mcpServers": {
    "yui": { "type": "streamable-http", "url": "https://mcp.yuigui.com/mcp" }
  }
}

That URL is the blocker-1 target, not today's endpoint. Today's endpoint is in spec/MCP.md.

Icon and screenshot assets

Rules: PNG, JPEG, WebP or SVG, at most 5 MiB each. Icons and logos must be square and at least 48 by 48. Raster at most 4096 on a side. SVG needs square numeric dimensions or a square viewBox. Paths in the manifest are ./-relative and every referenced file ships in the ZIP. The dashboard will not let you submit without a primary icon; dark variants and screenshots are optional.

Do not use brand/yui-logo-coral.png: it is 1701x1177, and a listing icon has to be square.

Screenshots are real now, captured inside chatgpt.com on 2026-10-01 during the live check (spec/MCP.md, "ChatGPT"). They are optional and no longer shown in the directory, so they are evidence for the reviewer rather than listing art. All light theme, sidebar cropped out, no chat titles, no tokens.

A fourth shot, site/public/progress/int8-live-4-openai-light.webp (740x900, 20 KiB), covers the window.openai fallback on a second screen. That one is engineering evidence, not listing material.

Two things to know about these shots before they go in a package. Each carries ChatGPT's CSP off badge, because "Enforce CSP in developer mode" is off by default, and the agent is named ChatGPTc7be, the throwaway account the live check used. Both are developer-mode artefacts. Retake them on the review account once blocker 2 is built, or crop.

No phone-side shot was taken, so the "same screen on the phone" image the first draft planned does not exist. Nothing in the submission needs it; the phone belongs in the video.

The required artifact is still the video walkthrough (review.demo_recording_url): one recording that runs the five positive cases end to end on a reviewer-accessible URL. The live check clears the path for it, since the whole flow is now known to work on a real account, but nothing has been recorded.

Auth, as the reviewer meets it

Everything in this section was probed live on 2026-10-01 against the endpoint in spec/MCP.md.

• Transport. MCP streamable HTTP, stateless, one POST per JSON-RPC message. GET answers 405.
• Unauthenticated call. initialize with no token answers 401 and carries WWW-Authenticate: Bearer realm="yui", error="invalid_token", resource_metadata="<mcp-url>/.well-known/oauth-protected-resource", scope="yui". OpenAI accepts that header as discovery, so Yui's path-based hosting is not a problem for ChatGPT itself. It stays a problem for a client that only probes the host root.
• Protected resource metadata. 200, names the resource, the authorization server, scopes_supported: ["yui"], bearer_methods_supported: ["header"].
• Authorization server metadata. 200 at both /.well-known/openid-configuration and /.well-known/oauth-authorization-server under the yui-oauth path. Advertises authorization_code and refresh_token, code_challenge_methods_supported: ["S256"], authorization_response_iss_parameter_supported: true, a registration_endpoint, and token_endpoint_auth_methods_supported: ["none", "client_secret_post", "client_secret_basic"].
• Client registration. Dynamic client registration. ChatGPT registers as a confidential client with client_secret_post and its own redirect. client_id_metadata_document_supported is false, so the newer CIMD path is not offered. The docs call CIMD preferred and DCR still supported, so this is a nice-to-have, not a blocker.
• Scope. yui, one scope, one agent thread per connection. No OIDC scopes are advertised, which is fine: Yui is not using Sign in with ChatGPT.
• Resource indicator. resource= is sent on /authorize and /token and checked; iss comes back on the redirect.
• Consent. The person approves the connection in the Yui app, or in the web Yui at /web, and picks which agent the client talks as. That is the step a reviewer cannot do today. See below.
• Live. The whole path ran on a real chatgpt.com account on 2026-10-01, free plan, developer mode. ChatGPT found Yui's OAuth settings from the server URL alone, registered itself and came back connected, so path-based discovery is not a submission risk for ChatGPT. Two things the live run added that a submission has to carry:
  • The /mcp suffix breaks OAuth. OpenAI's own page tells you to enter the server URL "including the /mcp path". With the suffix the 401 and its resource_metadata still look correct, and then /authorize fails with invalid_target, because yui-oauth only accepts a resource equal to the bare MCP URL. The mcp.json above names https://mcp.yuigui.com/mcp, so whichever service serves that origin has to accept that exact string as the resource. Settle it while blocker 1 is being built, not at submission: the MCP origin is a one-way door.
  • The CSP was never exercised. "Enforce CSP in developer mode" is off by default, so the chat marked the Yui card CSP off and the openai/widgetCSP we declare was not applied. Review runs with it on, so the exact-domains fix below is still untested under enforcement.

Test credentials

What OpenAI requires: "a login and password for a fully featured demo account that includes sample data", reachable at once, no new account sign-up, no 2FA, no email or SMS code, no magic link, no private network. The most common rejection reason they publish is a reviewer who cannot sign in.

Yui's own sign-in is Apple only, and approving a connector is a tap in the app. So a review account has to be built. The approach, in the order it has to happen:

1. A password identity for one flagged account. Add email-and-password sign-in to yui-auth, allowed only for accounts carrying a review flag. Everyone else keeps Apple-only. The web Yui at /web is then the reviewer's phone.
2. Auto-approve its connect requests. When the flagged account is the one signing in, yui-oauth approves the pending request itself and names the agent after the client. The reviewer finishes OAuth in the browser with no device in hand.
3. Seed sample data. One agent named ChatGPT, a thread holding three answered screens (a choice, a timer, a check-in form) and one small table. Invented data only, no real person, no real health or money numbers.
4. Where the credentials live. The portal's Review details field, typed in by hand at submission. Never in the ZIP, the manifest, this repo, a commit message or a card comment. Both repos are public.
5. Keep it alive. The same account serves later reviews. Rotate the password after each decision and re-enter it on the next submission.

A connection token (yui_ct_...) is not a way around this: ChatGPT's connectors only do OAuth.

Review test cases

Initial review needs exactly five positive and three negative cases. Each positive one needs the prompt, the tools it should trigger, and an observable expected result. Case 1 was run for real inside chatgpt.com on 2026-10-01 and passed, screen and tap both ways. The other seven are drafts: they have to be run, and recorded, before a submission.

Positive:

Negative:

Content and safety answers

The attestations at submit time are a checkbox set, so these are the answers behind them.

Tools, as the scan will read them

The portal imports tool names, titles, descriptions, schemas, security schemes, _meta and annotations when it scans the server, and the imported values are what gets reviewed. These are the values yui-mcp sends today.

All three required hints are explicit booleans on every tool, which is what the guidelines ask. openWorldHint: false everywhere is right: each tool is bounded to the one connected account. Justifications are no longer required; annotations are judged against behaviour.

Two things to fix on the resource before a submission:

• _meta.ui.domain is unset. Required with UI (blocker 4).
• _meta.ui.csp.resourceDomains holds https://*.fal.media. The review asks for the exact domains a component fetches from, so the wildcard may be flagged. Replace it with the host the renders actually come from.

Review checklist

Each item is met, unmet or unknown today.

Open work, in order

1. Decide the MCP origin and stand it up at mcp.yuigui.com, with the challenge file served there, and settle whether its MCP URL carries a /mcp suffix that yui-oauth will accept as the resource.
2. A /terms page.
3. The review account: password sign-in behind a flag, auto-approve, seeded sample data.
4. _meta.ui.domain on its own origin, and exact CSP domains, checked with "Enforce CSP in developer mode" on.
5. The tool-response audit against the privacy policy.
6. Record the walkthrough, run all eight test cases on it, and retake the screenshots on the review account.
7. Chris's call on blocker 5 and on country availability, then his sign-off.

Done since the first pass: the live chatgpt.com check, 2026-10-01. The screen drew in the chat, a tap was answered, the window.openai fallback worked, and the three listing screenshots came out of it.

Sources

Read 2026-10-01 on developers.openai.com. Each page has a Markdown twin at the same path with .md appended.

• /plugins/deploy/submission, the submission flow and the manifest field reference
• /plugins/plugin-guidelines, the published requirements for every listed plugin
• /plugins/deploy/app-review, remote MCP server review requirements
• /plugins/build/auth, the OAuth and discovery requirements
• /plugins/reference, the UI and _meta keys
• /plugins/guides/submit-claude-plugin, moving a Claude connector across