The gate is published now, ray-plugins fails two mechanical checks on every plugin, and "Anthropic Verified" is a connector label RDCO cannot earn with a skills-only bundle
The question
"Is ray-plugins (or any brigade-house public plugin) eligible for and worth submitting to the community marketplace plus 'Anthropic Verified' review, what does the security-review gate actually require, and does the leak-gate/manifest-allowlist output already satisfy it?" Promoted follow-up #1 of [[2026-07-21-anthropic-plugin-marketplace-brigade-strategy]], which recommended pursuing the badge without having read any acceptance criteria.
What we already know (from the vault)
- The two-tier directory and the badge claim both entered the vault from secondary press, never from a primary source. [[2026-07-21-anthropic-plugin-marketplace-brigade-strategy]] recorded "stricter review earns an Anthropic Verified badge" on the strength of a Security Boulevard post and recommended RDCO pursue it as "cheap credibility." Nothing primary was read.
- The acceptance criteria were explicitly logged as unread. [[2026-08-24-claude-marketplace-free-skill-paid-cta-policy]] closed the commercial-CTA question but left two named follow-ups open: the full Software Directory Policy text, and "the actual quality and security standards behind the
clau.de/plugin-directory-submissionform." Both are closed below. - The strategic recommendation already exists and is narrower than "list the catalog." [[2026-08-30-anthropic-builder-track-solo-founder-lane]] found there is no builder track, tier, or revenue share to earn, measured four months of zero stars/forks/watchers on two public repos, and prescribed a four-step sequence: fix the public surface, distribute the current build, submit one plugin and instrument installs, then publish one MCP connector. This brief does not relitigate that. It prices step 3.
- The hold is real, deliberate, and now three months deep.
project_rdco_plugin_marketplaceplus the working-context ledger: the founder constrained public sync pending an explicit greenlight; brigade-house has since gone v2, deleting the vendoredrail_adapter.py/walk.py/ mise engines that the public tree still carries.RayDataCo/ray-pluginslast pushed 2026-07-11 (verified today viagh repo view). - The self-hosted surface invokes no review at all, per [[2026-06-10-claude-code-plugin-private-distribution]] and [[2026-08-24-claude-marketplace-free-skill-paid-cta-policy]]: a
/plugin marketplace addcatalog is governed only by the reserved-name check. The gate below exists solely because submission is a different act.
What the web says
- The criteria are published, in two places the prior briefs never reached. [DOC] The submitter-facing gate is the plugin pre-submission checklist, a per-row table of named checks with explicit outcomes: Blocks (cannot submit), Policy hold (a human reads the version before it goes live), Warning, Note. Load-bearing Blocks rows include: a
READMEof at least 40 words in the plugin folder (the directory renders it as the listing description); aLICENSEfile in the plugin folder orlicenseset inplugin.json; one plugin per submission ("Pick one plugin first" when a marketplace repo root is selected); every package launcher pinned to an exact version; no real credentials in any file including examples; no macOS/Windows junk files; component paths inside the plugin folder; valid YAML front matter withdescriptionas a single text value. - The scanner's own policy prompt is committed in the open. [DOC]
anthropics/claude-plugins-official/.github/policy/prompt.mdplusschema.json(read viagh api, 2026-10-07) state the bar as "handles user data responsibly, not merely isn't malicious," and hard-fail on four conditions: malicious/deceptive/exfiltration behavior;has_broad_scope_hooks(anyUserPromptSubmit/PreToolUse/PostToolUsehook without a project-relevance gate);has_undisclosed_telemetry(any outbound call to a non-declared host without disclosure and a documented opt-out); anddescription_matches_behavior=falsewhere the mismatch involves hooks, telemetry, or data access. Credential handling is judged by a cross-service test: using the user's own service-X credential against service X is normal, routing it to a different service is a violation. The reviewer is told to read the whole shipped payload including dotdirs, and prompt-injection text aimed at the reviewer is itself a flag. - The submission path changed and the mechanics now carry real constraints. [DOC]
clau.de/plugin-directory-submission302-redirects to claude.com/docs/directory/publish; the old Claude Console form is unsupported. Anyone on a paid plan (Pro/Max/Team/Enterprise) can submit, with no partner program. A plugin bundle "comes from a GitHub repository, which must be public before the listing goes live." The listing belongs to the submitting organization, and "the first organization to submit a given repository folder holds that listing." Every version gets automated validation plus a security scan, and a person reviews a new listing before it goes live. Review time "isn't fixed." - "Verified" is documented only as an MCP-connector label. [DOC] The same page: connectors are "by default, listed as a Community connector," and "Anthropic may escalate a listing to Verified review, in which a reviewer tests each tool against the connector pre-submission checklist." Plugin bundles get validation, scan, and human review, with no label described. There is no application for Verified. [UNKNOWN] Whether a plugin bundle can carry a Verified label at all is not stated anywhere I could reach, and corroborating evidence is negative: none of the 315 entries in
claude-plugins-official/.claude-plugin/marketplace.json(parsed today) carries any verification field. [PRACT] The "stricter human review earns an Anthropic Verified badge" framing for plugins traces to press coverage (Shopifreaks, TechTimes), not to Anthropic documentation. - The Software Directory Policy says nothing about source availability or provenance, and a lot about the developer. [DOC] Full section walk of the policy (effective 2026-04-15): source code, open source, licensing and provenance are absent. Section 3 Developer Requirements are present and concrete: privacy-policy link (3.A), "verified contact information and support channels" (3.B), functionality/troubleshooting documentation (3.C), a testing account with sample data provided to Anthropic (3.D), three working prompt examples (3.E), verification of API ownership (3.F), and "Developers must maintain their Software and address issues within reasonable timeframes" (3.G). Ongoing review and remediation-to-stay-listed are in the opening paragraph.
- The directory now tracks your branch, and it measures your plugin. [DOC] After the first submission "you keep the listing current by merging to the tracked branch... The directory picks up each commit, scans it, and once it's published serves it to the people who have your plugin." A published plugin's Usage tab reports installs, versions, how often each skill and MCP server runs, and error rates. Corroborated mechanically:
claude-plugins-communityruns a daily SHA-bump sweep and a dailyowner-liveness-sweepthat fails when a recorded owner login resolves to a different account id, and the official repo'sscan-pluginsworkflow caches verdicts per (name, sha) and re-reviews on every bump, with downstream "label automation, delist" hooks named in its comments. - Anthropic now ships an eval runner with a no-plugin baseline arm. [DOC]
claude plugin evalis live in CLI 2.1.292 (run locally today): it executesevals/**/case.yamlorprompt.md+graders/*.mdagainst a plugin "and add[s] a no-plugin baseline arm." The pre-submission checklist instructs submitters to "Compare output with and without the plugin." It also states plainly that "a bundled suite passing is not a security vetting."
Convergences and contradictions
- Convergence, and it is the useful one: the gate is a user-safety gate, and the leak gate is a confidentiality gate. The scanner asks "does this observe or move more of the installer's data than its description justifies." The leak gate asks "did any RDCO client name, employee name, private path, or key escape the center." Those are opposite directions across the same boundary. The single genuine overlap is credentials-in-files, where the checklist's Blocks row and the gate's
AKIA/sk-ant-/private-key patterns coincide. - Correction to the parent brief: the badge RDCO was told to pursue is not available to the thing RDCO ships. [[2026-07-21-anthropic-plugin-marketplace-brigade-strategy]] recommended pursuing "Anthropic Verified" as cheap credibility. Verified is documented as a connector escalation, tested tool-by-tool, and RDCO's public plugins contain zero MCP servers (no
.mcp.jsonanywhere inray-plugins, verified today). There is nothing for a reviewer to escalate. - Correction to the builder-track brief: community listing is not review-free. [[2026-08-30-anthropic-builder-track-solo-founder-lane]] described community-directory listing as "free with no content review beyond a security screen." The publish doc says a person reviews every new listing, and the Software Directory Policy's Developer Requirements bind every listing. The step-3 recommendation survives; its cost estimate was low.
- Partial contradiction with the parent brief's moat framing. The parent concluded Anthropic was structurally not building the efficacy layer, so eval-gated shipping was safe ground. Anthropic now ships a two-arm eval runner in the CLI and tells submitters to measure lift. [INFER] This does not take the moat (per-client onboarding verification is untouched, and Anthropic explicitly disclaims eval passage as vetting), but "we measure two-arm lift and nobody else does" is no longer a differentiator that survives contact with the docs. It is table stakes with a first-mover head start.
- A precision that cuts RDCO's way. The checklist states "People who install the plugin get only the plugin folder," while the open-source CI scanner prompt reasons about a git clone of the entire repo. For a directory-delivered listing the narrower rule governs: submitting one plugin folder from
ray-pluginsdoes not ship the other nine. That makes "one plugin, not the catalog" cheaper than it looked, not just more disciplined.
Synthesis for RDCO
Eligibility: yes, and the generated-public-surface worry is a non-issue. The Software Directory Policy has no source-availability, open-source, reproducibility, or provenance requirement at all, and the review reads the shipped payload, not its build history. Nothing asks where ray-plugins came from, whether distribute.py produced it, or whether a private monorepo sits behind it. The only structural requirements the topology touches are satisfiable and already mostly satisfied: the repo must be public before the listing goes live (it is), the submission is one plugin folder at a time (which maps cleanly onto public.manifest.json's explicit allowlist), and after the first submission the directory follows the tracked branch. The real consequence of the generated topology is not eligibility, it is that a distribute pass stops being an internal hygiene step and becomes a public release, scanned on every commit and served to whoever installed. That is the part of the hold that marketplace submission actually changes.
The gate, concretely, and what the leak gate does not cover. I ran the mechanical layer today. claude plugin validate passes on the marketplace manifest and all ten plugin folders, including --strict on spot checks. ray-plugins ships zero hooks and zero MCP servers, no package launchers in any loaded surface, no lockfiles or .npmrc, no symlinks, submodules, LFS pointers or .gitattributes, no junk files, every file under 256 KiB (largest 69.7 KB), every plugin under 64 files against a 512 cap, and all 52 SKILL.md files carry front matter. That is the entire expensive half of the checklist cleared, and it is cleared because the brigades are skills-and-markdown, which is the shape the scanner likes. Against that, two Blocks classes fail on nearly every plugin: no LICENSE file and no license field in any of the ten plugin.json manifests (10/10 fail), and no README in the plugin folder for eight of ten (only ab-skill-factory at 1,897 words and ab-registrar at 380 clear it; the repo-root README does not count, and the directory renders the plugin-folder README as the listing description, so eight plugins would have no listing copy). Two names draw an almost certain reviewer hold: snowflake-cli and copilot-studio-agent-builder are built on brands that are not RDCO's, which the checklist calls out explicitly ("Name matches a known brand"). [INFER, medium] rdco-example and the opaque ab- prefix on generic discipline words invite the generic-name and look-alike holds. The leak gate and manifest allowlist satisfy one row of all of this (credentials absent from every file) plus the governance discipline of deciding deliberately what ships. They do not and cannot speak to license, README, naming, file-type limits, launcher pinning, hook scope, telemetry disclosure, or description-matches-behavior. Treating a clean leak-gate run as readiness would be the specific mistake this brief exists to prevent.
Worth it: yes, once, for the instrument rather than the reach or the badge. Under the credibility-for-phData-sales frame, where the domain is agents in production and the measure is pipeline rather than subscribers, the directory's most valuable output is not distribution. It is the Usage tab: installs, versions, per-skill run counts, and error rates on a plugin Anthropic reviewed and listed. That is the only artifact in reach that converts "I build agents" into third-party-attested operating numbers a phData buyer recognizes, and it is structurally impossible to produce from a self-hosted marketplace, which is why four months of ray-plugins yielded an uninformative zero rather than a negative result. A listing is also the cheapest available answer to the credibility-surface problem the builder-track brief named: a public repo that still describes itself as a "/hello example" while carrying seven production brigades currently argues against RDCO's competence to anyone who follows a link. The costs are real and should be stated flatly. Review latency is [UNKNOWN] by Anthropic's own wording. The maintenance obligation is contractual, not implied (Policy 3.G, plus verified contact info and a privacy-policy link). Exposure is continuous: every commit to the tracked branch is scanned, a failed version cannot go live, and the directory runs daily owner-liveness and SHA-bump sweeps, so a dormant RDCO is a visibly dormant listing. And the per-plugin multiplication is the quiet cost: ten plugins means ten submissions, ten READMEs, ten license decisions, ten reviews, ten things to maintain.
So: submit exactly one, and not for the badge. The badge recommendation in the parent brief should be retired outright rather than deferred, because Verified is a connector escalation and RDCO ships no connector. If the founder wants Verified, that is a decision to publish an MCP server, which is a different project with a different cost. The single-submission candidate that minimizes work and maximizes signal is ab-skill-factory: it is the only plugin that already clears the README gate, its name is the least brand-collision-prone and the least generic, and "the brigade that manufactures brigades" is the most legible story to a buyer evaluating agents-in-production. snowflake-cli should be excluded on two independent grounds, the brand-name hold and its adjacency to the phData disclosure boundary. The prerequisites are four, all reversible and all cheap except the first, which is the founder's: (1) greenlight a distribute pass, because submitting today would publish v1 artifacts that the center retired three months ago, and because post-submission every distribute pass is a release; (2) add a license (file or manifest field) and a plugin-folder README for whatever is submitted; (3) fix the repo and marketplace descriptions that still call the surface a minimal /hello template; (4) decide which claude.ai organization submits, because the listing belongs to that org and the first org to submit a repository folder holds it permanently. One further piece of free upside worth taking in the same pass: the brigades already carry two-arm eval evidence as prose reports in evals/*.md, and claude plugin eval now runs a native case.yaml + graders/ suite with a no-plugin baseline arm. Porting one station's evidence into the native format makes RDCO's eval claim independently re-runnable by anyone who installs it, which is the closest thing to a credibility artifact in this entire surface.
Why this is in the vault
This closes follow-up #1 of [[2026-07-21-anthropic-plugin-marketplace-brigade-strategy]] and follow-up #2 of [[2026-08-24-claude-marketplace-free-skill-paid-cta-policy]] with primary sources, and it converts step 3 of [[2026-08-30-anthropic-builder-track-solo-founder-lane]] ("submit one plugin and instrument installs") from a recommendation into a priced, prerequisite-complete action with a named candidate plugin. It also retires the standing "pursue Anthropic Verified" item, which was based on press coverage and does not apply to a skills-only bundle, and it supplies the specific blocker list the next ray-plugins distribution review needs so that review does not re-derive it.
Open follow-ups
- Can a plugin bundle carry the Verified label at all, and under what criteria? Primary docs describe Verified only in the MCP-connector review path, and no entry in the 315-entry official marketplace manifest carries a verification field, but the publish page's "no separate application for the Verified label" sentence sits in a general paragraph rather than a connector-scoped one.
- Is the scanner policy prompt committed in
claude-plugins-official/.github/policy/prompt.mdthe same policy the developer portal's scan runs? Thescan-pluginsaction README describes the official repo's prompt as a repo-local file "synced from private source," which leaves open whether the published text is the live bar or a mirror that can drift. - Do Software Directory Policy 3.D (testing account with sample data for Anthropic) and 3.E (three working prompt examples) bind plugin bundles, or only MCP connectors? The policy text does not scope them, and the plugin submission flow as documented never mentions either.
- What is the observed review latency and rejection taxonomy for plugin bundles, from publishers who have been through it? Anthropic states review time "isn't fixed," and no enforcement or delisting precedent for a bundle was found in any source.
- Does
claude plugin eval's native suite format subsume RDCO's two-arm execution-eval evidence, and is a bundled eval suite visible to directory browsers or only to someone who clones and runs it? If it is visible, the efficacy evidence becomes listing copy rather than internal paperwork.
Related
- [[2026-07-21-anthropic-plugin-marketplace-brigade-strategy]]
- [[2026-08-24-claude-marketplace-free-skill-paid-cta-policy]]
- [[2026-06-10-claude-code-plugin-private-distribution]]
- [[2026-08-30-anthropic-builder-track-solo-founder-lane]]
- [[2026-07-09-anthropic-plugin-ecosystem-vs-rdco-brigade-plugins]]
- [[2026-09-28-alphasignal-claude-physics-record-voicestudio-plugin-portal]]
- [[2026-07-18-agent-brigade-v2-simplification-design]]
- [[2026-08-30-data-native-si-anthropic-partner-tier]]
Sources
Vault:
rdco-vault/06-reference/research/2026-07-21-anthropic-plugin-marketplace-brigade-strategy.md- parent brief; source of the question and of the badge recommendation now retiredrdco-vault/06-reference/research/2026-08-24-claude-marketplace-free-skill-paid-cta-policy.md- governance-boundary framing; left the acceptance criteria and full policy text as open follow-upsrdco-vault/06-reference/research/2026-08-30-anthropic-builder-track-solo-founder-lane.md- the four-step sequence this brief prices; its "no content review beyond a security screen" line is corrected hererdco-vault/06-reference/research/2026-06-10-claude-code-plugin-private-distribution.md- self-hosted and private marketplace mechanics (cited at the level the two briefs above summarize)rdco-vault/06-reference/2026-09-28-alphasignal-claude-physics-record-voicestudio-plugin-portal.md- secondhand portal signal that asked for exactly this primary-source pass~/.claude/state/working-context.md- grepped 2026-10-07 for the distribute hold and the v2 deletions; not read whole
Local artifacts, inspected 2026-10-07:
~/Projects/phdata-private/brigade-house/distribution/DISTRIBUTION-SPEC.md,distribute.py,public.manifest.json- leak-gate pattern list, explicit allowlist, stage-not-push discipline~/Projects/ray-plugins/- full audit:.claude-plugin/marketplace.json, all tenplugin.jsonmanifests, per-plugin README/LICENSE/file-count/max-file-size, absence of hooks,.mcp.json, launchers, lockfiles, symlinks, submodules, LFS and junk filesclaude plugin validate(CLI 2.1.292) - run against the marketplace manifest and all ten plugin folders, plus--stricton three; all passedclaude plugin eval --help(CLI 2.1.292) - confirms nativecase.yaml+graders/suites with a no-plugin baseline armgh repo view RayDataCo/ray-plugins- public, licenseInfo null, last push 2026-07-11
Web (primary):
- Plugin pre-submission checklist - the submitter-facing gate; every Blocks/hold row cited above
- Publish to the directory - submission eligibility, public-repo requirement, org ownership and first-to-submit rule, review stages, Verified as a connector escalation, Usage tab, branch tracking. Reached via the 302 from
clau.de/plugin-directory-submission - Anthropic Software Directory Policy - full section structure; Developer Requirements 3.A-3.H; source-availability and provenance absent
anthropics/claude-plugins-officialviagh api:.github/policy/prompt.md,.github/policy/schema.json,.github/workflows/scan-plugins.yml,validate-plugins.yml,validate-licenses.yml,validate-frontmatter.yml,.claude-plugin/marketplace.json(315 entries parsed), READMEanthropics/claude-plugins-communityviagh api: README (read-only mirror, nightly sync from an internal review pipeline),.github/actions/validate-plugins/README.md(invariants I1-I11),.github/actions/scan-plugins/README.mdand its bundledpolicy/prompt.md,owner-liveness-sweep.yml,bump-plugin-shas.yml
Web (secondary, used only to attribute the badge claim):
- Shopifreaks and TechTimes - the "Anthropic Verified badge for plugins" framing; not corroborated by any Anthropic documentation reached here