Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

docs(openai-partner): add the partnership and GPT briefing and index the announcement copy

nirholas9d53cc4ecf53
diff --git a/marketing/openai-select-partner/README.md b/marketing/openai-select-partner/README.md
index 452f9944e..dd10d7db1 100644
--- a/marketing/openai-select-partner/README.md
+++ b/marketing/openai-select-partner/README.md
@@ -5,8 +5,12 @@ OpenAI Partner Network, plus the rules that govern how the badge may be used.
| File | What it is | Status |
| --- | --- | --- |
+| [`partnership-and-gpt-briefing.md`](partnership-and-gpt-briefing.md) | Full briefing on the partnership and the live custom GPT: tier facts and dates, the GPT's builder config and QA history, the Actions and MCP contracts, compliance posture, brand rules, live verification, open items | Reference |
| [`press-release.md`](press-release.md) | Draft press release, written against OpenAI's partner template | **Needs OpenAI approval before publishing** (see below) |
| [`social-copy.md`](social-copy.md) | Approved-messaging LinkedIn / X copy, ready to post | Ready |
+| [`founder-thread.md`](founder-thread.md) | Founder reply on the announcement thread: one extended long post plus a threaded version, with fact-check notes on what is and is not claimed | Ready |
+| [`founder-quote-post.md`](founder-quote-post.md) | Quote post from the founder's personal account: same ground as the reply, restructured and voiced differently, with a wider fact set so the two posts do not read as duplicates | Ready |
+| [`community-posts.md`](community-posts.md) | Three community-voice posts (builder, holder, creator) in long and short versions, plus posting notes and the claims that stay out | Ready |
| [`badge-usage.md`](badge-usage.md) | Where the badge assets live and how they may be used | Ready |
| [`cards/social-card.html`](cards/social-card.html) | Layout source for the two announcement graphics; renders with `npm run build:openai-cards` | Ready |
| `openai-press-release-template.docx` | OpenAI's original template, unmodified, kept as the source of truth for the draft | Reference |
diff --git a/marketing/openai-select-partner/partnership-and-gpt-briefing.md b/marketing/openai-select-partner/partnership-and-gpt-briefing.md
new file mode 100644
index 000000000..295505835
--- /dev/null
+++ b/marketing/openai-select-partner/partnership-and-gpt-briefing.md
@@ -0,0 +1,427 @@
+# three.ws and OpenAI: the full briefing
+
+Everything three.ws has with OpenAI, in one place: the partnership itself, what
+the Select tier does and does not grant, the custom GPT that is live on
+chatgpt.com, the free MCP connector behind the ChatGPT app, the contracts both of
+them call, the compliance posture that got them there, the brand rules that
+govern how we talk about it, and the items that are still open.
+
+Written 2026-08-15. Every live claim in the "Live verification" section was
+probed against production the same day; every historical claim cites the artifact
+in this repo that records it.
+
+---
+
+## 1. Executive summary
+
+three.ws is an **OpenAI Select Partner** in the OpenAI Partner Network. The
+application was accepted on **2026-07-14** (welcome email to [email protected]), the
+three.ws organization was verified on platform.openai.com the same day, and the
+status was announced publicly on **2026-07-25** across `/openai`, a blog post, the
+partner directory, the timeline, and the holder changelog.
+
+The partnership sits on top of a real, shipped product surface rather than a
+logo swap. three.ws ships **four live surfaces on OpenAI platforms**, all free,
+all keyless, all open source:
+
+1. **The three.ws 3D Studio connector** (Apps SDK / MCP) at
+ `https://three.ws/api/mcp-studio`, ten tools, no authentication.
+2. **The "three.ws 3D Studio" custom GPT**, published publicly in the GPT Store
+ since 2026-07-14, calling a REST Actions contract at `/api/3d/studio`.
+3. **AR from the conversation**: every generation carries a one-tap
+ place-in-your-room link that routes to Quick Look on iOS, Scene Viewer on
+ Android, and the WebGL viewer on desktop.
+4. **Spatial MCP**, an open CC0 response shape that makes a live 3D scene a
+ first-class tool result instead of a URL in a text blob. three.ws is the
+ reference implementation.
+
+Two things are still open, both human-gated: the **final submit** of the app to
+the OpenAI App Directory (the package is complete and audited), and the **press
+release**, which OpenAI must approve in writing before it can be published.
+
+---
+
+## 2. The partnership itself
+
+**Tier.** Select, the entry tier of the OpenAI Partner Network. The correct way
+to write it in copy is exactly "OpenAI Select Partner". Not "OpenAI partner" as a
+description of status, not "partnered with OpenAI on <product>", and never a tier
+above Select.
+
+**Dates that matter.**
+
+| Date | Event | Source of record |
+| --- | --- | --- |
+| 2026-06-28 | Free 3D Studio connector ships, five tools | `data/changelog.json` |
+| 2026-07-07 | Persona tools land (three more); App Directory submission package assembled | `data/changelog.json`, `prompts/store-submissions/_generated/openai-submission.md` |
+| 2026-07-14 | Accepted into the OpenAI Partner Network; organization verified on platform.openai.com; custom GPT published publicly in the GPT Store | `openai-submission.md` §0, `TRACKER.md` |
+| 2026-07-15 | Partner Network + ChatGPT presence recorded on the public timeline | `data/timeline.json` |
+| 2026-07-25 | Named an OpenAI Select Partner; `/openai` page, blog post, partner card, changelog entry all go live | `data/changelog.json`, `data/timeline.json`, `data/pages.json` |
+| 2026-07-28 | Announcement graphics and the X wording decision ("OpenAI Partner" on X only) | `badge-usage.md`, `social-copy.md` |
+| 2026-07-29 | Custom GPT QA pass; three builder-config defects found and fixed | `docs/chatgpt-3d-studio-gpt.md` |
+| 2026-08-06 | Tier story re-verified live; docs reconciled to the ten-tool surface | `forge-free-tier-evidence.json`, `data/changelog.json` |
+
+**What the Select tier grants.** Use of the OpenAI Partner Network badge under
+the rules in `badge-usage.md`, OpenAI's approved starter messaging and approved
+hashtags (`#OpenAISelectPartner`, `#OpenAIPartnerNetwork`), access to the partner
+portal, and access to PartnerU (role-based enablement and badging) plus the OPN
+Policy Guide.
+
+**What it does not grant.** It is not an endorsement of any three.ws product.
+Every surface that carries the badge also carries an independence line, and that
+line is load-bearing: "three.ws is an independent member of the OpenAI Partner
+Network at the Select tier. This page describes three.ws products that integrate
+with OpenAI platforms; it is not an OpenAI product and is not endorsed by OpenAI
+beyond the partner designation shown." It also does not put three.ws in the
+**Partner Locator**, which is a benefit of the Advanced and Elite tiers only. The
+requirements for progressing tiers are in the OPN Policy Guide in the portal.
+
+---
+
+## 3. The custom GPT on chatgpt.com
+
+**Live URL:**
+`https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio`
+
+Published publicly (Share to Everyone) on 2026-07-14 and returning HTTP 200 to a
+logged-out visitor as of today. The builder profile name attribution is set to
+"three.ws", and domain verification for the clickable byline was completed through
+the Manage Domain flow (the modal rendered blank on the first attempt; reopening
+it fixed that).
+
+**What it does.** The user describes an object and the GPT calls the three.ws
+generation Action, which returns a real, downloadable GLB plus a browser viewer
+link and an AR link. There is no account, no key, and nothing to pay: the
+generation lane is operator-funded.
+
+**Builder configuration** (canonical copy in `docs/chatgpt-3d-studio-gpt.md`,
+because the GPT is configured by hand in the ChatGPT builder UI):
+
+| Setting | Value |
+| --- | --- |
+| Name | three.ws 3D Studio |
+| Web Search | Off |
+| Apps (Beta) | Off |
+| **Image Generation** | **Off. This is load-bearing.** |
+| Code Interpreter & Data Analysis | Off |
+| Action | Imported from URL: `https://three.ws/.well-known/3d-studio-openapi.yaml` |
+| Action auth | None |
+| Privacy policy | `https://three.ws/legal/privacy` |
+| Conversation starters | Four, each naming exactly one concrete subject |
+
+**Why Image Generation being off is load-bearing.** QA on 2026-07-29 ran the four
+suggested prompts and three of them failed. "Make me a low-poly fox for my game"
+produced a DALL-E picture and never called the Action. "Create a dragon
+miniature" produced a picture plus two fabricated links (`three.ws/models/<id>.glb`
+and `three.ws/preview/<id>`), neither of which is a real route; both 404ed. "Create
+a 3D mascot" stalled on clarifying questions. Only "Surprise me" worked end to
+end. Three root causes, all configuration rather than code:
+
+1. With DALL-E available, the model sometimes satisfied a 3D request by painting
+ an image. Instruction text telling it not to return a PNG did not reliably
+ stop it. Only unchecking the capability did.
+2. An instruction reading "Always return GLB file links from https://three.ws"
+ pushed the model to invent three.ws-hosted URLs, because real GLB downloads
+ live on a `pub-*.r2.dev` CDN host. The corrected instructions say the
+ opposite: present URLs exactly as the Action returned them, and the R2 host is
+ correct.
+3. The builder held a hand-pasted, stale copy of the Action schema with no
+ `arUrl`, `previewImageUrl`, `tier`, `etaSeconds`, or title-carrying poll path.
+ The rule now is to re-import from the served URL after any spec change and
+ never hand-edit the schema inline.
+
+**The starter design rule.** Every conversation starter must name one concrete
+subject so the GPT can generate immediately without asking a question. "Create a
+3D mascot for my community" was dropped for stalling. At least one starter should
+surface the AR link.
+
+**The instruction text** (checked in verbatim so the builder can be re-pasted)
+encodes: one iron rule that the Action is the only way to produce a model; never
+show a URL an Action did not return; rewrite vague requests into one concrete
+subject; every generation is a new model (no in-place editing, fold the change
+into the prompt and regenerate); present three labeled links on a finished model
+(download GLB, browser viewer, AR); on a pending response hand the user the
+`watchUrl` immediately and keep polling the returned poll path verbatim; the
+age-13+ safety posture; and a standing rule never to mention pricing, because
+this GPT is simply free.
+
+**Known behaviors we do not control.** Free-tier ChatGPT renders third-party ads
+directly beneath the GPT's replies (Meshy AI was observed during QA). Nothing on
+our side changes that.
+
+**Housekeeping still open.** An earlier draft GPT exists at
+`g-6a5672fbf3f48191b559e482c7fcbf51` and should be deleted from My GPTs so it
+cannot be confused with the published one.
+
+---
+
+## 4. The Actions contract the GPT calls
+
+The GPT imports `https://three.ws/.well-known/3d-studio-openapi.yaml`, which is
+served from `public/.well-known/3d-studio-openapi.yaml` and is the single source
+of truth. The custom-GPT Action file in the submission kit
+(`prompts/store-submissions/_generated/openai-actions.yaml`) is regenerated from
+the served schema by `npm run sync:studio-openapi`, and drift between them fails
+`npm run check:studio-openapi`, which is wired into `npm run gate`.
+
+Submit:
+
+```bash
+curl -s -X POST https://three.ws/api/3d/studio \
+ -H 'content-type: application/json' \
+ -d '{"prompt":"a small ceramic robot figurine"}'
+```
+
+The response is either the finished model or a job to poll, and the design
+decisions behind that are deliberate:
+
+- **Store-clean responses.** The payload carries model URLs and job state only.
+ No pricing, no upgrade hints, no internal identifiers, no wallet or token
+ surface. That is what makes it publishable as a GPT Store action. The
+ agent-facing twin, `POST /api/3d/generate`, returns the same core fields plus
+ `free` and `upgrade`.
+- **The title-carrying poll.** ChatGPT Actions never resend context, so the
+ submit response embeds the prompt in the poll URL as `&title=`. When the poll
+ finally returns done, the `arUrl` and `viewerUrl` arrive labeled with the
+ original prompt without the GPT doing anything but calling the poll path
+ verbatim.
+- **Timeout honesty.** Actions calls time out around 45 seconds and generation
+ takes longer, so the endpoint holds the request as long as it safely can and
+ falls back to `pending` plus a poll handle rather than dying mid-request. A
+ `watchUrl` gives the user a live countdown page, and `previewImageUrl` shows
+ the painted concept art while the mesh is still being sculpted.
+- **Tiers.** `draft`, `standard`, `high`. **Standard is the default** on both the
+ Actions and MCP surfaces, and an attempt to make high the default
+ (`d97e8e94d`) was rolled back (`b39d6e2f7`). Evidence recorded live on
+ 2026-08-06 and read back from the durable `forge_creations` row rather than
+ inferred from response text: a no-tier call took 164s on `trellis_selfhost` and
+ produced a 5.07 MB GLB; an explicit `tier:"high"` call took 144s on the
+ self-hosted `hunyuan3d` worker and produced a 2.69 MB GLB. High degrades to
+ standard on a 402 or a submit timeout rather than failing the conversation.
+- **Safety before GPU.** Every prompt passes an age-13+ appropriateness gate
+ (`api/_mcp-studio/safety.js`) before any provider work. Refusals return
+ `400 prompt_rejected` with a plain-language message, and the GPT is instructed
+ never to reword a prompt to slip past it.
+
+---
+
+## 5. The free MCP connector (the ChatGPT app)
+
+Endpoint `https://three.ws/api/mcp-studio`. Transport is streamable HTTP,
+JSON-RPC 2.0 over POST (GET returns 405). Protocol version `2025-06-18`.
+`serverInfo` is `three-ws-3d-studio-free` 1.0.0. Auth mode: **none**. Because the
+app is anonymous and free, OpenAI's "provide a demo account with test
+credentials" requirement does not apply, and the submission form says so
+explicitly.
+
+**Ten tools**, verified live today: six generation tools (`forge_free`,
+`text_to_avatar`, `mesh_forge`, `rig_mesh`, `forge_avatar`, `refine_model`), the
+read-only `check_job` collector that picks up a generation which outran its tool
+call, and three persona tools (`create_agent_persona`, `get_agent_persona`,
+`persona_say`) that give an assistant a persistent, lip-syncing body in the
+panel.
+
+**The inline widget** is the `<model-viewer>` skybridge component built in
+`api/_mcp-studio/component.js`, resource `ui://widget/three-studio-model.html`,
+with a persona sibling `ui://widget/three-studio-persona.html`. It is linked to
+every generation tool through the tool's `_meta["openai/outputTemplate"]`. That is
+the widget an OpenAI reviewer sees. It renders the GLB, frames it, casts a ground
+shadow, auto-rotates until dragged, and offers Download, Spin, Recenter, and Open
+in three.ws, with a graceful download-or-open fallback where WebGL is
+unavailable.
+
+**Funding and limits.** Generation is operator-funded and routing is free-first
+through a health-aware router, so the marginal cost per generation is normally
+zero and the caller is never asked for anything. Real per-IP abuse protection
+still applies: 4 generations per minute, 30 per hour, 20 persona writes per
+minute, 300 transport requests per minute, plus a platform-wide breaker across
+all free-studio callers. Discovery calls and `check_job` never consume generation
+quota. Per-IP caps fail open on a limiter outage because a Redis blip must not
+dead-end a free feature; the platform-wide breaker deliberately fails closed,
+since a limiter outage is exactly when unbounded global spend would do damage.
+
+**Compliance posture** (full audit in
+`prompts/store-submissions/_generated/openai-submission.md`, 7/7 policy items
+PASS):
+
+- **Zero crypto or payment surface.** The endpoint, its handlers, the widget, and
+ every reviewer-facing JSON artifact contain no coin, token, wallet, x402, or
+ payment string, and `tests/mcp-studio.test.js` pins that with a regex assertion
+ in CI. The paid, wallet-enabled studio is a completely separate endpoint at
+ `/api/mcp-3d` and shares none of this surface.
+- **Data minimization.** The raw internal generation record has fourteen fields;
+ the tool response returns five. Stripped: `creation_id`, `job_id`, `status`,
+ `mode`, `path`, `tier`, `backend`, `durable`, `eta_seconds`,
+ `estimated_credits`, `preview_image_url`, `reference_image_urls`. No session id,
+ trace id, user id, or PII. Inputs are minimal too, with
+ `additionalProperties: false` on every schema.
+- **Correct tool annotations** on all ten tools (read-only, destructive,
+ idempotent, open-world hints), with the rationale for each documented.
+- **Clear utility not native to ChatGPT.** ChatGPT cannot turn language into a
+ manipulable, downloadable 3D asset. That is the whole value proposition.
+
+**Submission state.** Ready to submit. Both historical blockers are resolved and
+re-verified live: the rate-limiter store outage that made every generation return
+429 (fixed permanently by moving to self-hosted Redis) and the `/viewer?src=`
+404 that made the `viewerUrl` in every tool response a dead link. What remains is
+the owner's **final submit in the partner portal**, plus a re-capture of the two
+landscape screenshots against the shipped inline widget (the earlier captures
+were taken against a standalone viewer bundle that has since been deleted and
+must not be reused).
+
+---
+
+## 6. AR and Spatial MCP
+
+Every generation on both surfaces carries an `arUrl` of the form
+`https://three.ws/api/ar?src=<glbUrl>&title=<name>`, built by one shared
+constructor. ChatGPT's only job is to print the link. `GET /api/ar` reads the
+User-Agent server-side and branches: iPhone and iPad get a launch page whose
+button opens Apple Quick Look, with the USDZ generated from the GLB on the fly in
+the page (no server-side USD tooling exists or is needed); Android gets a 302
+straight into Google Scene Viewer through an ARCore intent, with a browser
+fallback; desktop falls back to the WebGL viewer, so the link is never a dead
+end. Rigged avatars carry `kind=avatar`, which swaps the flow to a "Bring it to
+life" handoff into `/irl`, where the avatar walks, animates, and talks through the
+user's camera. Shared AR links unfurl with a real render of that exact model,
+produced by a public GLB-to-PNG renderer at `/api/render/glb`.
+
+**Spatial MCP** is the open, CC0 response shape for returning a live 3D scene as
+a first-class MCP result: scene GLB, camera, environment, animation, AR handoff.
+It is renderer-agnostic, carries no payment or coin surface, and three.ws is the
+reference implementation. It is one of the four things the `/openai` page claims,
+and it is the piece that makes the partnership a standards play rather than a
+single integration.
+
+---
+
+## 7. Brand, badge, and messaging rules
+
+The badge assets are OpenAI's, supplied unmodified, and ship with the site at
+`https://three.ws/partners/openai/openai-select-partner.svg` (preferred) and
+`@3x.png`. The rules that came with them: use the badge as provided (no recolor,
+crop, rotation, effects, or rebuilding from parts; on a dark surface put it on a
+white plate rather than inverting it), keep clear space of roughly half its cap
+height on every side, write the status as "OpenAI Select Partner" in copy, never
+imply OpenAI endorses a three.ws product, and never use the badge as a product
+logo, favicon, or app icon.
+
+For social posts and link previews we attach an announcement card rather than the
+bare badge, because the badge alone reads as OpenAI's mark rather than our
+announcement and carries none of the required independence line. Five cards exist
+(announcement, announcement with the short "OpenAI Partner" phrasing for X, the
+Studio product card with no badge, and two-mark lockups on white and black). All
+are regenerated from `cards/social-card.html` with `npm run build:openai-cards`.
+
+One caveat carries real risk and must not be dropped: the **two-mark lockups use
+OpenAI's logomark, not the partner badge**. OpenAI's brand guidelines govern the
+logomark separately, and a co-branded lockup is the kind of use those guidelines
+normally expect a partner to clear first. This was the owner's call on
+2026-07-28. For a compliant announcement graphic, use
+`social-card-announcement.png`.
+
+---
+
+## 8. Where this is live on the site
+
+| Surface | Path |
+| --- | --- |
+| Partner page | `/openai` (`pages/openai/index.html`) |
+| Announcement post | `/blog/three-ws-openai-select-partner` |
+| Partner directory card | `/partners` |
+| Timeline marker | `/timeline` |
+| Changelog entries | `/changelog` |
+| Press kit (journalist-ready assets) | `/press` |
+
+---
+
+## 9. The press release: the one blocking step
+
+The draft in `press-release.md` is complete and written against OpenAI's own
+partner template. **OpenAI requires written approval before a partner publishes a
+press release about tier status.** The process: confirm the two bracketed fields
+(dateline city, currently Wilmington, Delaware, matching the governing-law
+jurisdiction in our terms; and the spokesperson's name and title on the quote),
+email the full release to **[email protected]** for review, wait for written
+approval, and only then publish or distribute.
+
+Nothing else in the announcement pack is gated. The badge, the site pages, and
+the social copy all use OpenAI's own approved assets and approved messaging.
+
+---
+
+## 10. Live verification, 2026-08-15
+
+Probed against production today:
+
+| Check | Result |
+| --- | --- |
+| GPT Store page, logged out | HTTP 200 |
+| `/openai` partner page | HTTP 200 |
+| `/.well-known/3d-studio-openapi.yaml` | HTTP 200 |
+| Badge SVG | HTTP 200 |
+| MCP `tools/list` | 200, ten tools, exactly the documented set |
+| `GET /api/3d/studio?job=bogus123` | 400 `invalid_job` in 0.43s, correct designed error |
+| `POST /api/3d/studio` (new prompt) | **No response within 100s** |
+| `POST /api/3d/generate` (new prompt) | **No response within 100s** |
+| `POST /api/3d/generate` (previously generated prompt) | 200 in 0.34s with a real GLB and `arUrl` |
+
+**The generation submit path is degraded right now, and it affects the custom
+GPT.** Production logs over the last half hour show the free NVIDIA NIM text-to-3D
+lane returning 504, the paid TRELLIS lane rate-limited and degrading to that same
+NIM lane, and the Vertex reference-image lane returning 403. In
+`api/_mcp-studio/gpt-forge-client.js`, `startForge` waits up to 90 seconds for the
+job to be **accepted**, and on a timeout it retries the accept once, so a fully
+blocked submit path can hold a request up to about 180 seconds before any
+response. ChatGPT Actions give up at roughly 45 seconds, so in this state the GPT
+shows the user a failed action rather than the documented `pending` plus watch
+link. Discovery, polling, the safety gate, cached prompts, and every read path are
+unaffected. The fix belongs in the accept path (fall back to `pending` well inside
+the Actions deadline instead of blocking on an unhealthy upstream) and in the
+lane chain (a healthy rung ahead of the 504ing NIM lane); it is not a partnership
+or configuration issue.
+
+---
+
+## 11. Open items
+
+| Item | Owner | Notes |
+| --- | --- | --- |
+| Final submit of the app in the OpenAI partner portal | Owner | Package complete, 7/7 policy PASS |
+| Re-capture the two landscape widget screenshots | Owner | Must be the shipped inline widget, rendering a real generated GLB |
+| Delete the draft duplicate GPT `g-6a5672fbf3f48191b559e482c7fcbf51` | Owner | Avoids confusion with the published GPT |
+| Press release: fill two fields, email for OpenAI approval | Owner | [email protected], do not publish before written approval |
+| Free generation submit path degraded (see §10) | Engineering | Actions deadline vs 90s-plus-retry accept window |
+| "Nine tools" drift in older copy | Engineering | See below |
+
+**The nine-versus-ten drift.** The connector shipped with five tools, grew to
+nine, and reached ten when `check_job` landed. `/openai` and `docs/mcp-studio.md`
+were reconciled to ten on 2026-08-06, but three announcement-era documents still
+say nine: `press-release.md`, `social-copy.md` (both the X and LinkedIn copy), and
+the OpenAI section of `docs/partners.md`. The press release is the one that
+matters most, because it has not been sent yet and would go out with a number
+that is now wrong. Fix all three before the release is emailed for approval.
+
+---
+
+## 12. File map
+
+| Piece | Path |
+| --- | --- |
+| Announcement pack index | `marketing/openai-select-partner/README.md` |
+| Press release draft | `marketing/openai-select-partner/press-release.md` |
+| Badge rules and assets | `marketing/openai-select-partner/badge-usage.md` |
+| Social copy | `marketing/openai-select-partner/social-copy.md` |
+| Custom GPT builder config | `docs/chatgpt-3d-studio-gpt.md` |
+| AR pipeline | `docs/chatgpt-ar.md` |
+| Free MCP connector docs | `docs/mcp-studio.md` |
+| Paid sibling server | `docs/mcp-3d-studio.md` |
+| Partner ecosystem overview | `docs/partners.md` |
+| App Directory answer sheet and audit | `prompts/store-submissions/_generated/openai-submission.md` |
+| Submission status tracker | `prompts/store-submissions/_generated/TRACKER.md` |
+| Apps SDK handoff pack | `docs/openai-pr/00-START-HERE.md` |
+| Actions endpoint | `api/3d/studio.js` |
+| MCP connector | `api/mcp-studio.js`, `api/_mcp-studio/*` |
+| Inline widget | `api/_mcp-studio/component.js` |
+| Served Action schema | `public/.well-known/3d-studio-openapi.yaml` |
+| Partner page | `pages/openai/index.html` |