2026 10 10 connector directory readiness - SyberLabs/RISE GitHub Wiki

RISE in Anthropic's connector directory: readiness review

Date: 2026-10-10. Live build at review time: 5b0996d2. Owner: Mateo (SyberLabs).

Three reviews were run in parallel against main and the live site: Anthropic's published listing requirements (primary sources, quoted and dated 2026-10-09), a security and protocol scan in a reviewer's voice, and a policy, privacy, brand and documentation review. This page is the synthesis: what was fixed today, what is still open, who owns each item, and the exact submission material.

1. Verdict

RISE can be submitted as a Community connector as soon as the two pull requests in flight land (wording and documentation; the five screenshots), the owner settles one decision (whether the Plus voice stays out of the Claude connector for the first submission), and every unchecked must-have in section 7 is done, including the example-prompt verification and the tool testing. Everything the automated scan is known to check is addressed or in flight.

2. How listing works

  • The process is self-serve: a developer portal (claude.ai/directory/manage), an automated scan, then a Community listing by default. Anthropic escalates to Verified on its own. There is no partner program and no interest form.
  • Any paid Claude plan can submit. The listing belongs to the organization it is submitted from (a personal Pro or Max account, or a Team or Enterprise owner).
  • Review contact: [email protected]. MCP App questions: [email protected].
  • Sources: claude.com/docs/connectors/building/submission, claude.com/docs/connectors/building/review-criteria, the Anthropic Software Directory Policy (support.claude.com/en/articles/13145358).

3. Fixed today (merged and live at 5b0996d2)

PR What a reviewer would have flagged What changed
#603, #604 The per-address rate limit (30/min) throttled every Claude user together, since all arrive from Anthropic's egress range; the 429 was not JSON-RPC Anthropic's published range 160.79.104.0/21 is exempt; a tripped limit answers a JSON-RPC error with Retry-After; the limit is charged before the body is read
#603 tools/list was 28 KB; the rise_present description was 11.7 KB and echoed the whole Current back 10.5 KB and 1.7 KB; the accepted call returns a one-line receipt; the long guide lives behind rise_guide
#603 Every /.well-known/* path returned the app page; no security.txt Unknown paths and the OAuth discovery paths answer a JSON 404; security.txt (RFC 9116) is served
#603 The resource declared inline only while the app declared fullscreen and pip Both declare the same modes
#605 Model-written scene code was held in by a denylist; containment rested on the host's CSP The card ships its own Content Security Policy with no unsafe-eval; the worker locks Function, eval, the global object and its message channel; a foreign import() and fetch are proven refused in the browser test

Verified live: security.txt is text/plain; the OAuth discovery path is 404; tools/list is 10,535 bytes; the card's HTML carries the policy meta and declares frameDomains: [] with only https://rise.syberlabs.io for connections and resources.

Also verified clean by the scan, with live evidence: origin validation, no sessions, body caps, clean JSON-RPC errors, tool titles and annotations (readOnlyHint), the SVG figure allowlist, admin routes behind Cloudflare Access, retired routes at 410, no telemetry, no secrets in the tree or history, zero production dependency vulnerabilities (the one Dependabot alert is a development-only chain).

4. In flight (pull requests open at the time of writing)

  • Wording and pages. The server text described rise_guide as something Claude should "read how" from; the policy forbids directing Claude to load behavioural instructions at run time, so the tools now describe a format reference, and the guide's second-person lines become field descriptions. "In the manner of 3Blue1Brown" is dropped. The Sign in link leaves the card (the connector needs no account). The privacy page states that Cloudflare Workers logging is on and what it keeps. Terms say the card shows words the model wrote, and the public-domain claim is scoped to the United States. README gains a Claude section; APPS.md gains the sound bed.
  • Screenshots. Five PNGs of the card alone, at least 1000 px wide, each paired with its prompt, under docs/assets/directory/.

Landed after this page was written: both are merged and done. Wording and pages is #607, live at e3dd1b97; the screenshots are #608, at 608ea99f.

5. Open decisions (owner)

  1. Plus and the AI-audio rule. Policy §4B excludes "software that uses AI models to generate images, video, or audio content" unless Anthropic permits it in writing; the portal asks for an "AI media generation" acknowledgment. The Plus voice is ElevenLabs speech. Two paths: (a) submit with Plus unreachable from the Claude card and absent from all connector text, which is already true today, and ask about Plus afterwards; (b) write to [email protected] first and hold submission. Recommendation: (a). The free voice is the device's own speech synthesis; the visuals are code and SVG the model writes, which is closest to the allowed "diagrams, charts". Decided 2026-10-10: (a). Submit without the Plus voice; it stays out of the Claude card and all connector text, and the AI-media answer says RISE calls no media model.
  2. Terms and Plus pricing. TERMS.md sells Plus at $8.99 while /api/plus/status says available: false and the production payment link is Stripe's test link. A reviewer who reads the terms will notice. Either remove the Plus clause until Plus ships, or make it true. Decided 2026-10-10: correct the terms and remove the test link. The terms and the privacy policy now say Plus is not offered and nothing is for sale, with no price; what Plus would send is kept in conditional form. Production's PLUS_PAYMENT_LINK is empty.
  3. Security reporting channel. security.txt points at [email protected] because no security@ address exists and GitHub private vulnerability reporting is disabled on the repository. Enabling it (gh api -X PUT repos/SyberLabs/RISE/private-vulnerability-reporting) would allow a second, durable channel. Decided 2026-10-10: enable it. Private vulnerability reporting is on (the API answered 204; it now reads {"enabled":true}). SECURITY.md says how to report and what is in scope, and security.txt names the advisory form as a second Contact: and the policy as Policy:.
  4. Workers plan. No limits.cpu_ms is set; admission of a maximum-size Current measured 38 to 112 ms locally. On Workers Free (10 ms CPU) large calls can fail intermittently with a Cloudflare error page. If the account is on Workers Paid, set "limits": { "cpu_ms": 1000 }; if not, that is a reason to upgrade before the listing drives traffic.
  5. Branding. Anthropic's trademark guidelines permit its marks only in approved materials. Keep "Claude" out of the server name, tool names and icon (true today). Whether marketing may say "RISE in Claude" is unknown; ask [email protected].

6. Process risks seen during this review

  • The Codex-feedback bot can race a live author. On #603 the author agent disabled auto-merge to fix a real Codex finding; the workflow answered the thread itself, re-armed auto-merge and merged the unfixed commit. The fix followed as #604. Until the workflow checks for an active author, push the fix commit first and answer second.
  • The voice-provider deploy job fails on every run (the Cloudflare token lacks access to the rise-plus-preview Worker). Production deploys regardless, but the listing must not advertise Plus voices while that is so.
  • The deploy guard refuses any commit that is no longer the head of main. Several merges in quick succession mean only the last one's run deploys. Expected, but check release.txt rather than a single run's status.

7. Submission checklist

Submitted 2026-10-11 from the paid account; live build a7387e6d; Community listing pending the automated scan; the portal had no field for screenshots or allowed link URIs (kept ready in docs/assets/directory/); the attestation was ticked on the Inspector run and the Claude run (the portal's wording: MCP Inspector or a custom connector in Claude).

Must have:

  • Streamable HTTP MCP server, no auth, tools with titles, descriptions, annotations
  • Self-contained card (frameDomains: []); own CSP; display modes agree
  • Rate limit that does not throttle Anthropic's egress; JSON-RPC errors throughout
  • /.well-known/security.txt; OAuth discovery paths 404
  • Privacy policy URL (https://rise.syberlabs.io/privacy), terms, support contact
  • Public documentation URL (https://rise.syberlabs.io/apps)
  • Tool and server text free of run-time instruction fetching, creator names, promotion (#607)
  • 3 to 5 PNG screenshots ≥ 1000 px wide, cropped to the card, prompts separate (#608)
  • Terms without a sale: no price, no Stripe purchase, Plus described as not offered; no test payment link in production (decision 5.2)
  • A security reporting channel: GitHub private vulnerability reporting, SECURITY.md, security.txt Policy: (decision 5.3)
  • At least three working example prompts (drafted below; being verified through Claude by another session; prompt 5's tool calls are in the Claude run)
  • Every tool run in MCP Inspector (the Inspector run, 2026-10-10)
  • Every tool run as a custom connector in Claude (the Claude run, 2026-10-10, from a Claude Code session). The claude.ai chat and phone runs: done by the owner's use during the mobile pass; the five demo prompts on phones remain to be recorded.
  • The attestation itself, ticked by the owner in the portal on the Inspector run and the Claude run (2026-10-11)
  • A paid Claude account in the organization that will own the listing (the owner submitted from the paid account, 2026-10-11)
  • The Plus decision (section 5.1) and the AI-media acknowledgment answered truthfully (the answer is in the submission packet)

Should have:

  • Allowed link URIs in the portal: https://syberlabs.io, https://rise.syberlabs.io (the portal had no field for them on 2026-10-11)
  • Light theme check of the card against the host's tokens; WCAG AA contrast; keyboard control
  • Card assets accept Origin: *.claudemcpcontent.com and never depend on Referer (iOS)
  • structuredContent well under 150,000 characters (true: a Current is capped far below)
  • The reader is told the voice is synthetic (the Voice row names the device voice)

8. Listing copy (draft)

Name: RISE

One-liner: Turns an answer into a short reading you watch and listen to: your device speaks it, the words appear as they are spoken, and a drawn visual plays behind them.

Description:

Ask Claude to explain something and present it in RISE. Claude writes the answer as a reading: a title and its passages (up to 64 beats and 20,000 characters, which is about twenty minutes at the longest; most readings are a minute or two), and, if it likes, a colour theme, a look, and a diagram or figure of its own. RISE shows it in the chat as a card. You press Play. Your device's built-in voice reads it aloud, the words appear as they are spoken, and a visual plays behind them.

You stay in control: play, pause, go back or forward a passage, hear a passage again, change the pace, and in Settings choose the theme, how vivid the imagery is, whether it moves at all, the text size, the sound and the voice. If your device has no voice, the reading still plays, paced, with every word shown.

The imagery is drawn, not generated. RISE draws it live with its own procedural engines: line drawings, a strange attractor, soft light, fractal flames, spectral plates. When Claude adds a diagram or figure, RISE checks it first and runs it apart from the page, with no network and no access to your data. RISE calls no image, video or speech model of its own: the voice is your device's built-in speech.

RISE needs no account and no key, and it costs nothing. It receives only the reading Claude writes for it, plays it, and keeps none of it. It never sees your conversation, memory or files.

What it does not do: it takes no actions, makes no pictures to order, and is not for tables, code, or long reports. Claude answers those itself.

Made by SyberLabs. Privacy: rise.syberlabs.io/privacy

Category: Education (primary); Media or Learning if offered. Before connecting, users need: nothing. Reads or writes: neither; both tools are read-only. Authentication: none. API: our own (Cloudflare Workers at rise.syberlabs.io). Icon: public/android-chrome-512x512.png. Documentation: https://rise.syberlabs.io/apps. Privacy: https://rise.syberlabs.io/privacy. Support: [email protected].

Compliance answers:

  • Financial transactions: none.
  • AI media generation: RISE calls no media model. The voice is the device's own speech synthesis (Web Speech API); diagrams and figures are code or SVG the model writes, rendered by RISE. (If Anthropic reads on-device speech as AI audio, this is the question to ask them; see section 5.1.)
  • Conversation data: only the tool's argument, not stored.
  • Prompt injection: the tool returns a one-line receipt or a clipped validator message; rise_guide returns a format reference, not instructions.

Demo prompts:

  1. "Explain how black holes bend light, and present it in RISE."
  2. "Show me what a derivative is, as a RISE reading in the premium-educational style."
  3. "Give me a RISE reading on why the sky is blue, in the premium-educational style."
  4. "Present a short reflection on Marcus Aurelius and the inner citadel in RISE, in the Garden look."
  5. "Show me the parts of an animal cell in RISE."

Negative checks (Claude should not call RISE): "Book me a table for two tomorrow night." "Generate a picture of a cat." "Write a 10,000-word report and read it in RISE."

9. Questions for Anthropic

  1. Does on-device browser speech synthesis count as "AI models … generate … audio"? Would written permission be granted for a speech model to narrate Claude's own answer?
  2. May the card mention a paid tier or link to a checkout outside the card?
  3. For an authless server, what satisfies "test account credentials"?
  4. May marketing say "RISE in Claude" and show Claude's interface in screenshots?
  5. Is the MCP App UI reviewed by a person at the Community level, and against what list?
  6. Is there a size limit on the ui:// HTML resource?

10. Source documents

The three review documents (requirements with quotes and URLs, the security scan with file and line references, the policy review with the full listing draft) were produced in the session's scratchpad on 2026-10-09/10 and are summarised above. Their findings are recorded in the task records LIVE-004, CC-005 and SND-001 through the roadmap tool.

Published from docs/product/discussions/2026-10-10-connector-directory-readiness.md. Edits made here are overwritten on the next push to main.