Enterprise architecture knowledge MCP server for the Agentic POC demo
Server defines 7 tools with complete input schemas and descriptions. Naming follows verb_noun convention (arch_pattern_lookup, tool_selection_lookup, risk_policy_lookup, brand_context_lookup, brand_search). Descriptions are substantive (100-200 chars), explaining WHAT each tool does and WHEN to use it. However, output schemas are not explicitly documented in the source code, only inferred from test assertions. Parameters have types and descriptions, meeting baseline requirements. Error handling is minimal; graceful degradation for missing API keys is present but recovery guidance is absent. No tool annotations (readOnlyHint, destructiveHint) despite all tools being READ_ONLY. Composition is sound: each tool has one responsibility, and tool outputs chain naturally (pattern_id from arch_pattern_lookup feeds into arch_diagram and arch_pattern_references).
Returns full diagram_data (components, connections, boundaries) for a matched architecture pattern
Deterministic architecture pattern recommender: scores pattern profiles against the input's industry, data stack, cloud, and constraints, then emits a recommendation with confidence score and core fields only (no diagram_data or source_references)
Returns source_references (citations) for a matched architecture pattern with industry, data_stack, cloud, and constraints context
Brand context lookup: returns company name, industry hint, confidence, and tags for a domain. Identity-only response with no logo_url, positioning, or brand fields. Gracefully unavailable without BRANDFETCH_API_KEY
Brand search: returns ranked brand candidates matching a company name query with name, domain, and logo_url. Gracefully unavailable without LOGO_DEV_SECRET_KEY
Output schemas not documented in source code. Tool descriptions state what is returned (e.g., 'core fields only', 'full diagram_data'), but formal JSON Schema output definitions are absent. LLMs cannot plan downstream tool calls without knowing response structure.
No tool annotations despite all tools being READ_ONLY. MCP spec (2026-07-28) supports readOnlyHint, destructiveHint, and idempotentHint. Adding readOnlyHint=true to all 7 tools would signal safety to agents and enable optimizations.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Risk and policy control recommender: returns required controls, HITL review flag, and review reason based on industry, data classification, region, and deployment type
Deterministic platform recommender: scores platform profiles against the input's data stack, constraints, latency, and use-case keywords, then emits a recommendation with cloud fit, constraint-aware reasoning, and alternatives with when-to-prefer rationale
Error handling lacks recovery guidance. brand_search and brand_context_lookup gracefully degrade when API keys are missing (available=false), but descriptions do not explain what the LLM should do next. Should state: 'If unavailable, the tool cannot be used; consider alternative approaches.'
Parameter descriptions lack format/constraint details. E.g., 'strategy' enum in brand_search lists values but does not explain when to prefer 'suggest' vs 'match'. 'latency' in tool_selection_lookup accepts 'batch | real-time' but does not clarify impact on recommendations.
No pagination or result-limiting documented. arch_pattern_lookup, tool_selection_lookup, and risk_policy_lookup return single recommendations, but brand_search returns 'candidates' array with no documented limit. Large result sets could exhaust context windows.