06-reference/research

goodnotes connector tool surface listing

2026-10-20·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)

Three unauthenticated render-only tools, and a shelf that is open self-serve on both platforms - the only real gates are Anthropic's AI-media clause and OpenAI's under-13 clause

The question

"What does the Goodnotes Claude/ChatGPT connector actually expose as a tool surface, and what did getting listed on those connector shelves require of Goodnotes?"

Promoted follow-up #1 of [[2026-10-08-goodnotes-generated-visuals-canvas]], which reframed the Goodnotes threat away from the canvas and onto the connector shelf. This brief answers both halves against primary sources, including the live Model Context Protocol (MCP) server itself.

How to read the labels. [DOC] is a documented fact traced to a source opened during this research. [OBS] is an observation from a live endpoint or page I queried directly. [INF] is my inference, which is not a finding. [UNKNOWN] means not publicly documented, and I stopped there rather than reasoning to a plausible answer.

What we already know (from the vault)

What the web says

1. The tool surface, read off the live server

The Claude directory listing publishes the connector endpoint, so the server's own tools/list is the primary source rather than any vendor prose. Both deployments answered an unauthenticated MCP handshake on 2026-10-20.

2. What listing required - Anthropic

Anthropic's connector documentation is first-party material about its own shelf, so treat the framing as promotional even where the requirements are concrete.

3. What listing required - OpenAI

Materially different, and the differences all run in the direction of more friction.

4. What neither platform documents about Goodnotes specifically

Convergences and contradictions

Synthesis for RDCO

The load-bearing finding is that the Goodnotes connector needs no authentication at all, and Anthropic's rules explicitly bless that. Three tools, all readOnlyHint: true, zero read scope, no account credential, no OAuth handshake, and a directory panel that says "SIGN-IN: Not required." The write into the user's library is handed off to an already-signed-in local client through a link the user taps. Read against SW-R18, this is not a constraint Scribble Works has to survive on the shelf; it is the compliant default. Anthropic's checklist offers "no authentication for public data" as a first-class option, and the policy only compels OAuth 2.0 for servers that "connect to a remote service and require authentication." A Scribble Works connector that exposes its catalog and returns a printable can be built with no accounts of any kind on either side of the call, which means no kid accounts, no kid profiles, no identity federation, and nothing to federate. The thing the founder's principle forbids turns out to be the thing the platform does not ask for.

That reframes scribble-works#333 from an access question to a content-policy question, and the two real gates are now named. The first is Anthropic's AI-media clause. Printable kids' activity sheets are image-bearing, and the exclusion of "Software that uses AI models to generate images, video, or audio content" is live. The carve-out permits "other design assets" generated "as part of a design workflow, provided the developer does not offer standalone image generation as a primary service," and Goodnotes shows exactly how to sit inside it: the host model authors SVG or Mermaid source, and the connector renders it deterministically. [INF] The transferable design is that a Scribble Works connector should render from its existing 81-game catalog and from model-authored layout, and should not call an image model at request time. If it does, the primary-service test gets hard to win; if it does not, the connector is a design tool with the same shape Anthropic already approved three times on one listing. That is my read, and it is a judgment about how a reviewer would apply a clause, not a documented ruling.

The second gate is OpenAI's, and it is the first real disqualifier found in either rulebook: "Plugins may not explicitly target children under 13." Scribble Works is explicitly for three-to-eight-year-olds. The clause is about who the plugin targets, and the honest ambiguity is whether a product whose user is a parent and whose subject is a four-year-old "explicitly targets" that child. [UNKNOWN] how OpenAI applies it; no worked example or policy elaboration is published. Against that, Anthropic's Software Directory Policy contains zero instances of "child," "minor," "under 13," "COPPA," "parent," or "teen" - I checked the full policy text, and the age dimension simply is not in Anthropic's rulebook. Combined with no domain verification, no identity verification, no ZIP package, no country declaration, and no EU residency restriction, the sequencing answer is not close: Anthropic first, and OpenAI only after the under-13 reading is resolved. OpenAI's own published conversion guide for Claude connectors makes that order cheap to follow later.

The channel economics point the same way the parent brief's cost analysis does, and that is the part worth carrying into the accepted-page math. Goodnotes' connector consumes the host's model tokens, not Goodnotes' own AI credits, and is free on every Goodnotes plan. A Scribble Works connector lane inherits the same split: the agent pays for authoring, RDCO pays only for rendering what it hands back. Set against the 5-to-8-cent target per accepted page in [[2026-10-01-accepted-spec-v0]], a connector lane where the model composes and RDCO renders is structurally cheaper per page than the site lane, and it arrives with [[2026-08-31-product-model-rulings]]' separate-key, one-ledger billing shape already ruled on. The listing also comes with distribution RDCO cannot buy: every directory entry is automatically eligible for in-chat Suggested Connectors, with usage-based ranking, across claude.ai, Cowork, Desktop, mobile, and Claude Code. [INF] That last point sharpens Isenberg's landlord warning rather than softening it, because usage-based ranking means the shelf rewards whoever is already being used, and a Verified badge is awarded by an automatic escalation RDCO cannot apply for.

Two product mechanics are directly liftable. First, the confirmation boundary. Because read-only tools auto-run, the user-facing gate is not the tool call but the widget's link out, and Anthropic provides an allowed-link-URI list that suppresses the "Open external link" prompt for origins the submitter owns. For Scribble Works that maps to: the tool returns a preview widget, the parent taps through to a PDF on scribbleworks.co, the parent stays the actor, and nothing is written to any account anywhere. Second, the dual-dressing trick. One MCP server, two hostnames, with platform-specific _meta selected off the client identity, is a single codebase serving both shelves. [OBS] I verified that by running the same tools/list twice with different client identity and watching the OpenAI widget vocabulary appear. That is the cheapest possible two-shelf architecture, and it is sitting in the open.

Why this is in the vault

It closes follow-up #1 of [[2026-10-08-goodnotes-generated-visuals-canvas]] with the connector's exact enumerated tool surface from the server itself rather than inference from a help article, and it converts RayDataCo/scribble-works#333 from "can we reach the shelf" into two specific go/no-go questions: whether a printable generator clears Anthropic's standalone-image-generation primary-service test, and whether Scribble Works clears OpenAI's under-13 clause. It also supplies the auth answer that SW-R18 was blocking on - no accounts are required by either platform for a generation-only connector - and it gives [[2026-10-07-ray-plugins-marketplace-verified-eligibility]] the connector-side path to the Verified label that the plugin-side path could not reach.

Open follow-ups

Each of these was checked against existing briefs in 06-reference/research/ via mcp__qmd__query on the rdco-vault collection before being filed, and each is a research question rather than a build task, a live-account test step, or something blocked on credentials.

Related

Sources

Vault

Live endpoints queried 2026-10-20 (read-only MCP metadata calls: initialize, tools/list, resources/list; no tool was invoked)

Web - Anthropic first-party, fetched 2026-10-20. Vendor documentation about its own shelf is promotional material, not neutral.

Web - OpenAI first-party, fetched 2026-10-20. Same promotional caveat.

Web - Goodnotes first-party, fetched 2026-10-20

Flagged - not usable as fact