Anthropic's plugin/MCP marketplace direction and how RDCO's brigade-house strategy must adapt
The question
What is Anthropic's current strategic direction for its plugin/MCP marketplace (based on July 2026 signals), and how should RDCO's brigade-house plugin strategy adapt if Anthropic ships a native plugin discovery catalog? Context: surfaced from an orphan analysis doc whose premise (brigade value partly rests on RDCO controlling skill discovery/routing) needed a web reality-check.
What we already know (from the vault)
- [[2026-07-09-anthropic-plugin-ecosystem-vs-rdco-brigade-plugins]] — Anthropic already runs two plugin worlds:
claude-plugins-official(255 entries surveyed, vendor-shaped — one entry per external system, dominated by bare MCP connectors) andknowledge-work-plugins(17 first-party discipline/role packs, "Plugins = Connectors + Skills"). The role packs are structurally the same bet as RDCO'sab-brigades. Key finding: zero eval evidence across the entire public corpus (1 trigger-accuracy fixture in 54 inspected official plugins; none in the 141 knowledge-work skills), and endemic README↔reality drift with no freshness mechanism. - [[2026-06-10-claude-code-plugin-private-distribution]] — Private/enterprise distribution (Bitbucket, self-hosted git,
strictKnownMarketplaceslockdown,CLAUDE_CODE_PLUGIN_SEED_DIR) is a separate, first-class surface Anthropic explicitly supports; the CAF-on-phData-Bitbucket plan lives here, orthogonal to any public catalog. - [[project_rdco_plugin_marketplace]] — RDCO topology: private center
RayDataCo/brigade-houseis the only dev home (13 plugins, cellar evals);ray-pluginsis a generated public surface via distribute.py + leak gate + manifest allowlist. House fully v2 since 2026-07-18; public surface undistributed since (drift until greenlight). - [[2026-07-15-sivulka-million-bad-employees-evals-as-moat]] and [[2026-07-14-gbrain-adopt-port-build-tradeoff]] — independent convergence on the same moat: generic skills are "a million bad employees"; the differentiator is firm-specific, verified efficacy, not distribution reach.
What the web says
- The native discovery catalog already shipped. Anthropic launched the official Claude Code plugin directory on 2026-05-22; it is pre-configured in Claude Code v2.1.143+ and browsable in-app via
/plugin→ Discover with no setup. Theclaude-plugins-officialrepo hit 20k stars in 4 days and lists 200+ plugins by July 2026 (~20 first-party). (Claude Camp, designrevision) - Two-tier trust, not two-tier quality. Official = Anthropic-curated, every listing passes automated security review; stricter review earns an "Anthropic Verified" badge. Community (
claude-plugins-community) = automated safety screening only, SHA-pinned. Verification certifies security/safety, not efficacy. (Security Boulevard, Claude Camp) - The MCP Registry is a "root of trust," deliberately NOT a discovery portal. It verifies namespaces, enforces uniqueness, provides canonical metadata, and leaves discovery to others — explicitly DNS-like. ~9,650 latest server records / ~29k versions as of 2026-05-24; Anthropic cited 10,000+ active public MCP servers in its Dec 2025 update. (Gentoro, MCP adoption stats)
- Direction = verticalization + server-side execution, not marketplace scale. May/June 2026: 20+ legal-vertical MCP connectors + 12 practice-area plugins via Managed Agents (Tokyo event) — a shift toward industry-specific packs running in customer-controlled sandboxes with private MCP connectivity, not generic-marketplace scale. (Totalum, BigGo)
- No monetization/billing layer as of mid-2026 — free distribution; competitors (Vercel Pro, Smithery hosted, x402 pay-per-call) are ahead on billing. High Anthropic runtime lock-in; skill portability to Codex/Cline is poor. (Totalum)
- Anthropic is actively warning about unverified MCP risk — the security-review gate is being positioned as the reason to prefer the official directory over raw third-party MCP servers. (Security Boulevard)
Convergences and contradictions
- Convergence — the moat was never discovery. Vault (zero-eval corpus) and web (verification = security not efficacy; registry = trust not discovery) agree: Anthropic is building the distribution + trust rail, not a quality/efficacy-proof rail or a per-client onboarding-verification rail. RDCO's two gates (evals = "works at all", mise = "works HERE") sit in layers Anthropic is structurally not building.
- Convergence — verticalization validates the
ab-brigade shape. Anthropic's own move to role/practice-area packs is the same discipline-shaped bet as RDCO's brigades; the difference remains proof, not structure. - Premise correction (the load-bearing one). The question's conditional — "if Anthropic ships a native discovery catalog" — has already resolved to yes (shipped 2026-05-22). So this is not a hypothetical to plan against; it is a live condition RDCO has been operating under for two months, and the orphan doc was written after it shipped. The strategic question is no longer "what if" but "given it exists, what did/didn't it erode."
Synthesis for RDCO
The framing that "the brigade house's value partially depends on RDCO controlling skill discovery and routing" is the part of the orphan doc's premise that the web reality-check retires. A native catalog does not, and structurally cannot on current evidence, erode the brigade-house moat — because the moat was misdescribed. Anthropic's catalog answers "which plugin exists and is it safe to install" (catalog-level discovery + security verification). It does not answer "does this plugin actually work, and does it work in this client's environment against their systems." Those two questions are exactly the brigade house's two gates: eval-gated shipping with per-fixture lift evidence (works at all) and mise declared-deps/reachability/version-stamp onboarding gates (works HERE). Neither is a discovery feature, so a discovery catalog leaves both untouched. The 2026-07-09 doc's own "two halves of trust" synthesis already had this right; the web just confirms Anthropic is not competing on either half.
Where the catalog should change RDCO behavior is at the top of the funnel, not the moat. RDCO should stop treating a self-owned public discovery surface as a differentiator — that game is over, Anthropic ships 200+ plugins pre-configured out of the box, and no third-party catalog will out-distribute the in-app /plugin Discover screen. The correct adaptation is to ride Anthropic's rails for reach and keep the private center for delivery-grade proof: (1) list eligible public ray-plugins in the community marketplace and pursue the "Anthropic Verified" badge — it is cheap credibility and a security-trust signal, but explicitly not the moat, so market it as table stakes, not as the sell; (2) keep RayDataCo/brigade-house as the private dev home and the manufacturing floor where evals + mise gates produce delivery-grade artifacts (architecture documented in [[2026-07-18-agent-brigade-v2-simplification-design]]); (3) preserve intra-plugin routing (MENU + router skill) as RDCO-owned — Anthropic's discovery is catalog-level (pick a plugin), its only router (smb-router) is nascent and hand-authored, and situation→skill routing inside a delivery-grade pack is a layer it is not building.
For the [[2026-06-09-caf-restructure-proposal|CAF]]/phData topology, the native catalog is essentially a non-event: enterprise distribution via private Bitbucket + strictKnownMarketplaces lockbox + seed-dir container pre-population is a separate surface Anthropic first-classes and does not fold into the public catalog. CAF's value story runs entirely on the onboarding gate (mise proving connectivity to the client's actual Snowflake/systems) and eval evidence, both of which the catalog does not provide. If anything, Anthropic's own verticalization move (legal packs, Managed Agents in customer sandboxes with private MCP) is a tailwind: it normalizes the exact "discipline pack + private connectors, run in the client's environment" shape RDCO sells — and RDCO ships it with the two gates Anthropic omits. The one genuine watch item is monetization: Anthropic has no billing layer yet, but a future paid/verified tier or a Managed-Agents-plus-vertical-plugins bundle is the plausible next move, and that is where a well-capitalized platform could eventually compress a generic-plugin business. RDCO's insulation there is the same as everywhere: sell delivery-grade proof and per-client onboarding, which is a services-and-verification moat, not a distribution moat that a platform can undercut.
Net: adapt by conceding public discovery to Anthropic entirely, riding its rails for top-of-funnel reach and its verified badge for security-trust, while doubling down on the two gates it structurally does not build (measured efficacy + per-client onboarding verification) as the actual brigade-house differentiation. The orphan doc's "adopt their surface, keep our teeth" line is the right posture; the web evidence just tightens it from prediction to fact.
Why this is in the vault
This brief closes the citation gap on the orphan 2026-07-09 plugin-ecosystem analysis by grounding its central premise against July 2026 reality, and it directly informs the brigade-house public-distribution decision (whether/how to surface ray-plugins on Anthropic's catalog + pursue the Verified badge) and the CAF/phData plugin-topology positioning — both are live, undelivered decisions.
Open follow-ups
- Is
ray-plugins(or any brigade-house public plugin) eligible for and worth submitting to the community marketplace + "Anthropic Verified" review — what does the security-review gate actually require, and does the leak-gate/manifest-allowlist output already satisfy it? - Does Anthropic's Managed-Agents "customer-controlled sandbox + private MCP" model overlap or collide with the CAF-on-Bitbucket private-plugin plan — is Managed Agents a distribution channel RDCO should ride for CAF, or a competitor to it?
- What would an "Anthropic ships a paid/verified tier or billing layer" scenario do to RDCO's public-plugin economics — is there a monetization move (x402-style pay-per-call on an RDCO MCP connector) worth prototyping before the platform closes it?
- Should RDCO publish its eval-lift evidence format as a public artifact (a "delivery-grade" badge of its own) to make the efficacy-vs-security distinction legible to buyers, given Anthropic's badge only certifies security?
- Does the in-plugin router (MENU + situation→skill routing) remain defensible if Anthropic promotes
smb-routerinto a first-class routing primitive across the official directory — track that plugin as the leading indicator.
Related
- [[2026-07-09-anthropic-plugin-ecosystem-vs-rdco-brigade-plugins]]
- [[2026-06-10-claude-code-plugin-private-distribution]]
- [[2026-07-15-sivulka-million-bad-employees-evals-as-moat]]
- [[2026-07-14-gbrain-adopt-port-build-tradeoff]]
- [[project_rdco_plugin_marketplace]]
- [[2026-06-09-caf-restructure-proposal]]
Sources
Vault:
rdco-vault/08-tooling/2026-07-09-anthropic-plugin-ecosystem-vs-rdco-brigade-plugins.mdrdco-vault/06-reference/research/2026-06-10-claude-code-plugin-private-distribution.mdrdco-vault/06-reference/2026-07-15-sivulka-million-bad-employees-evals-as-moat.mdrdco-vault/06-reference/research/2026-07-14-gbrain-adopt-port-build-tradeoff.mdrdco-vault/.claude/projects/-Users-ray/memory/project_rdco_plugin_marketplace.mdrdco-vault/01-projects/phdata/2026-06-09-caf-restructure-proposal.md
Web:
- Claude Code Plugins: Inside Anthropic's Official Directory (2026 Guide) — Claude Camp
- Official Claude Code Plugins: The Complete List (2026) — designrevision
- 7 Claude Code Plugins From the Marketplace Worth Your Time — Security Boulevard
- What Is Anthropic's New MCP Registry? — Gentoro
- MCP Adoption Statistics 2026 — Digital Applied
- Agent Skills Marketplaces in 2026: Anthropic vs Vercel vs OpenAI vs Cline vs MCP — Totalum
- Anthropic launches 12 legal AI plugins and 20+ connectors — BigGo Finance