Phase D Batch 26 — Official Vendor Skill/Plugin Marketplaces
Theme: First-party vendor distributions (Anthropic, AWS, IBM, Microsoft, Google)
Roster (10)
| slug | stars | distribution | cli_binary | local_ui | orchestration | multi_model | tier |
|---|---|---|---|---|---|---|---|
| ccplugins-awesome | 806 | claude-plugin | no | no | sequential | yes | A |
| asyncope-agentskills-io | 19,405 | methodology-doc | no | no | none | no | A |
| aws-aidlc-workflows | 2,443 | methodology-doc | no | no | sequential | no | A |
| agentcore-samples | 2,868 | standalone-repo | yes (agentcore) | no | hierarchical | yes | A |
| google-agents-cli | 2,589 | claude-plugin | yes (agents-cli) | no | task-decomposition-tree | yes | A |
| ms-mcp-servers | 3,209 | mcp-server | no | no | none | no | A |
| ms-learn-mcp | 1,661 | mcp-server | yes (mslearn) | no | none | no | A |
| ibm-mcp | 379 | standalone-repo (catalog) | no | no | none | no | A |
| anthropics-skills-ip-guard | — | delta-report | no | no | none | no | delta |
| anthropics-plugins-official | — | delta-report | no | no | none | no | delta |
Intra-batch Patterns
The vendor distribution strategies split cleanly into three tiers: (1) MCP server catalogs (ms-mcp-servers, ms-learn-mcp, ibm-mcp) where the vendor provides tool servers that work with any MCP client — zero proprietary runtime lock-in; (2) workflow methodology documents (asyncope-agentskills-io, aws-aidlc-workflows) that are IDE-agnostic specifications rather than implementations; and (3) CLI-anchored full-stack platforms (agentcore-samples, google-agents-cli) where the vendor provides both the agent runtime and the developer tooling. A striking pattern across all three tiers: every vendor has independently landed on MCP as the tool integration protocol — even ccplugins-awesome (Claude Code native plugins) exposes tools through the MCP transport. The multi-model frameworks (agentcore-samples, google-agents-cli, ccplugins-awesome) all support framework-agnostic agent execution (LangGraph, CrewAI, Strands, etc.), while the MCP-catalog vendors (IBM, Microsoft) are deliberately model-agnostic at the catalog level. Star count spread is extreme — asyncope-agentskills-io at 19,405 stars dwarfs ibm-mcp at 379, reflecting the difference between a community-adopted open specification and an enterprise product catalog.
Most Interesting Finds
agentcore-samples (2,868 stars): The only managed serverless agent runtime in the entire corpus — not a framework you self-host, but AWS infrastructure you deploy to. Cedar policy language for fine-grained agent access control (e.g., deny delete operations on S3 objects tagged classification=sensitive) is the most sophisticated authorization model in the entire batch. The Gateway component automatically wraps any REST API as an MCP server, compressing large API surfaces into tool-count budgets appropriate for LLMs — a pattern IBM independently discovered (621 APIs → 4 tools for Guardium).
aws-aidlc-workflows (2,443 stars): The only framework in the corpus supporting 6 IDEs simultaneously (Kiro, Amazon Q, Cursor, Cline, Claude Code, GitHub Copilot) from a single specification with no per-IDE branching. The mandatory audit.md output artifact and A-E structured Q&A files are the most rigorous governance mechanism in the vendor tier. Blocking gates on extension violations (security/baseline, testing/property-based opt-in) are the only example of user-activated feature flags in the workflow tier.
ibm-mcp (379 stars): Unique in the entire corpus for mainframe coverage — IBM Developer for z/OS MCP, IBM i MCP, and z/OS Connect give AI agents access to COBOL/PL/I/JCL environments. MAMMAL-MCP for protein-protein interaction prediction is the only biomedical AI server in the entire 33-batch catalog. The 4-tool meta-API compression pattern (Guardium DP: 621 APIs → 4 tools; Security Verify: 210 APIs → 4 tools) is an independently discovered architectural principle that matches AgentCore Gateway's approach.
Items Written as Delta Reports (not Tier C)
anthropics-skills-ip-guard — The anticipated
ip-guardskill does not exist inanthropics/skills. Closest analog isbrand-guidelines. Canonical coverage:spec-system/anthropics-skills/. The source-available dual-license (CC-BY-NC-4.0 for commercial, MIT for non-commercial) is the actual IP governance mechanism, not a discrete skill.anthropics-plugins-official —
anthropics/claude-pluginsreturns HTTP 404; repo does not exist publicly. Canonical coverage:spec-system/anthropics-knowledge-work-plugins/(Phase B Batch 13). No Tier C because the subject repo itself is absent, not just underdocumented.
Cross-References Discovered
- agentcore-samples ↔ ibm-mcp: Both independently converge on API-surface compression as a core LLM-friendliness technique (AgentCore Gateway: REST→MCP; IBM Guardium: 621→4 tools). Neither references the other.
- asyncope-agentskills-io ↔ ms-learn-mcp: MicrosoftDocs extends the agentskills.io SKILL.md spec with a
context: forkfield for context isolation. This is the only documented extension of the agentskills.io spec found in the corpus. - asyncope-agentskills-io ↔ google-agents-cli: Google ADK skills follow the agentskills.io SKILL.md format (YAML front-matter + progressive disclosure) but don't cite it. The ADK's
metadata.requirespattern matches the spec exactly. - aws-aidlc-workflows ↔ ccplugins-awesome: Both support Claude Code as one target tool among many; neither is CC-exclusive. The ccplugins-awesome
agent-sdk-verifier-py.mdusesmodel: sonnetper-agent model locking — a pattern absent from AWS's multi-IDE workflow. - ms-mcp-servers ↔ ms-learn-mcp: Two complementary Microsoft MCP strategies in the same batch — ms-mcp-servers covers Azure infrastructure (40+ Azure services) while ms-learn-mcp covers documentation retrieval. Both use Streamable HTTP transport. Neither references the other in their READMEs.
- agentcore-samples: References Strands SDK (aws/strands-agents, covered in Phase D Batch 29) as a first-class framework-agnostic partner alongside LangGraph/CrewAI.