MCP Health Status Comparative Report - TerrenceMcGuinness-NOAA/global-workflow GitHub Wiki

MCP Server Health & Status — Comparative Report

Date: July 2, 2026 Servers compared: agentcore-mcp-rag (AWS-native, Python/FastMCP) vs eib-mcp-gateway (Node.js, Docker) Method: Ran the six health/status utility tools on each server at the highest detail level the tools expose, then compared the results side by side.

This report is a live operational health/status snapshot. It complements the earlier design-level comparison in MCP-Parity-Assessment-SoC-Topology (which focused on response quality, data-gap analysis, and separation-of-concerns topology) by focusing strictly on what the two servers' own diagnostic tools report about themselves right now.

Each section presents a three-column table: AgentCore (agentcore-mcp-rag) | EIB Gateway (eib-mcp-gateway) | Comparative Analysis.


Tools Invoked (highest detail)

Tool AgentCore call EIB Gateway call
mcp_health_check deep=true, detailed=true, functional=true deep=true, detailed=true, functional=true
get_server_info include_capabilities=true include_capabilities=true
get_knowledge_base_status include_graph=true, include_vector=true include_graph=true, include_vector=true
get_health_trend limit=10 limit=10
get_quality_metrics compare=true compare=true
check_knowledge_integrity sample_size=50 sample_size=50

1. Server Identity & Capabilities (get_server_info)

Attribute AgentCore (agentcore-mcp-rag) EIB Gateway (eib-mcp-gateway) Comparative Analysis
Server name / version MDC MCP/RAG Server v1.0.0 global-workflow-unified-mcp v3.6.2 Different version lineages. Gateway is the mature, long-lived Node.js line (v3.x); AgentCore is the freshly-cut AWS-native rewrite (v1.0.0). The low AgentCore version reflects a new codebase, not lesser maturity.
Runtime Python / FastMCP (AWS Bedrock AgentCore Runtime) Node.js (Docker) Fundamentally different stacks. AgentCore is AWS-managed serverless microVM; Gateway is a container.
Architecture label 9 active modules Week 2 Consolidated + v3.6.0 SOC Refactor (7 modules) Gateway advertises an explicit consolidation/SoC refactor (8 duplicate tools eliminated, 21 unique tools). AgentCore reports a flat 9-module layout — the SoC-refactor topology recommendations from the parity assessment appear applied on the Gateway side.
Total tools 52 51 Near-identical surface. AgentCore's extra tool aligns with the 9th module split. Functionally the two expose the same capability families.
Multi-tenancy 5 tenants (default gw) Single tenant (not tenant-aware) Major differentiator. AgentCore is multi-tenant (gw, gw_sfs, gw_jedi_gfs, gw_v17, gw_gefs_v12), enabling branch-isolated knowledge bases. Gateway serves a single knowledge base.
Capabilities Data Access connected · Vector available · Graph available RAG enabled · GitHub enabled (no auth) · Workflow root auto-detected Both fully wired to their data planes. Gateway explicitly reports GitHub is unauthenticated; AgentCore does not surface a GitHub-auth flag here but its GitHub functional probe passed.

2. Infrastructure Health (mcp_health_check, deep+detailed+functional)

Dimension AgentCore (agentcore-mcp-rag) EIB Gateway (eib-mcp-gateway) Comparative Analysis
Overall status HEALTHY (4/4 components) HEALTHY (9/9 components) Both healthy. The component counts differ only because each server groups its self-checks at a different granularity (AgentCore rolls up to 4 buckets; Gateway itemizes 9). Not a quality difference.
Component breakdown Base (FastMCP), Utility (4 tools), Vector (21 indices), Graph (105,891 nodes / 4,729,093 rels) Base (51 tools), Workflow Info (3), Code Analysis (4), Vector (15 indices), Graph (19,689 nodes), Semantic (15), Operational (3), GitHub (4), SDD (9) AgentCore's health snapshot reports far more graph nodes/rels and more vector indices. Gateway's per-module itemization is more granular for spotting which subsystem is degraded.
Vector store OpenSearch — 21 indices ChromaDB — 15 indices Different vector engines. AgentCore's OpenSearch carries more indices (dual embedding families + registry); Gateway's ChromaDB is leaner.
Graph store Neptune (openCypher / SigV4 IAM) Neo4j Different graph engines. AgentCore = AWS-managed Neptune; Gateway = self-hosted Neo4j.
Data-plane probe (rolled into component health) ChromaDB heartbeat OK, 220,538 docs, sample query returned 10 results from global-workflow-docs-v8-0-0 Gateway surfaces an explicit end-to-end query probe in the health output. AgentCore validates the equivalent via its functional module tests (below).
Functional validation 9/10 passed, 1 skipped, 0 failed 6/6 passed, 0 failed Both green. AgentCore runs more functional probes (10 modules incl. branch_isolation); its one skip is workflow_info because the workflow filesystem is not mounted. Gateway runs 6 tool-level probes, all pass.
Notable skip / caveat workflow_info SKIP — /mnt/workflow not mounted none AgentCore's workflow-filesystem-backed tools are degraded in this environment (EFS not mounted to the runtime). Gateway auto-detects its workflow root locally and has no equivalent gap.
Per-module latency (AgentCore) semantic 198ms · code_analysis 43ms · graph_rag 62ms · ee2 192ms · operational 184ms · github 214ms · branch_isolation 407ms (latency not itemized per functional test) AgentCore exposes latency per functional probe, useful for spotting slow paths (branch_isolation is the slowest at 407ms — expected, it touches multiple tenants). Gateway reports pass/fail without per-test latency here (latency lives in get_quality_metrics).

3. Knowledge Base — Vector Store (get_knowledge_base_status)

Metric AgentCore (OpenSearch) EIB Gateway (ChromaDB) Comparative Analysis
Collections 16 15 Comparable collection counts.
Total documents 252,013 220,538 AgentCore holds ~31.5K more documents (~14% larger corpus), largely due to dual embedding families and the SHA registry.
Embedding families Titan 1024-dim + MPNet 768-dim (dual) MPNet 768-dim (single) Key differentiator. AgentCore maintains parallel Titan-1024 and MPNet-768 embeddings of the same content, enabling higher-dimensional semantic retrieval. Gateway is MPNet-768 only.
Largest code collection mdc-code-context-titan1024: 90,135 · mdc-code-context-mpnet768: 60,576 code-with-context-v8-0-0 / mdc-code-context-mpnet768: 60,574 The MPNet code corpora match almost exactly (60,576 vs 60,574) — strong evidence both ingested the same source tree. AgentCore adds a 90K-doc Titan code corpus on top.
Docs collections workflow-docs titan1024 20,155 · mpnet768 22,498 · nova1024 150 versioned: v8-0-0 22,498 · v8-1-0 20,511 · v8-2-0 23,624 Gateway keeps multiple versioned doc snapshots (v8-0-0 → v8-2-0); AgentCore keeps one live set per embedding family. Gateway's versioning aids rollback; AgentCore's is leaner.
Community summaries titan1024 2,113 · mpnet768 2,113 community-summaries 2,113 Identical count (2,113) — the hierarchical GraphRAG community layer is in parity across both servers.
Empty / placeholder collections Several nova1024 collections at 0 docs (jjobs, community, ee2, code-context) none reported empty AgentCore has provisioned-but-unpopulated nova1024 collections (future embedding family, not yet ingested). Cosmetic; does not affect current retrieval.
Registry mdc-content-sha-registry: 52,754 (not exposed as a collection) AgentCore surfaces a content-dedup SHA registry as a first-class collection; Gateway manages dedup internally.

4. Knowledge Base — Graph Store (get_knowledge_base_status)

Metric AgentCore (Neptune) EIB Gateway (Neo4j) Comparative Analysis
Files 17,273 17,273 Exact match — both graphs index the identical file set.
Functions 95,996 2,012 Largest single divergence. AgentCore's function-level graph is ~48× denser. AgentCore materializes Fortran subroutines/functions and Python functions as first-class nodes; Gateway's Neo4j does not expand the Fortran call graph to the same depth.
Classes 281 54 AgentCore captures more class-like entities, consistent with its deeper code-graph ingestion.
Total nodes 148,976 19,689 (from health probe; KB-status omits a total) AgentCore's graph is ~7.5× larger by node count. This is the core "richness" gap the parity assessment flagged.
Total relationships 4,555,408 4,220,211 Closer than node counts suggest (~8% more on AgentCore). Both are multi-million-edge graphs dominated by CALLS.
CALLS 3,407,104 3,306,540 Within ~3% — the call graph is near-parity.
USES 997,616 679,698 AgentCore has ~47% more USES edges (deeper Fortran USE module resolution).
Relationship vocabulary CALLS, USES, DEFINES, DEPENDS_ON_ENV, IMPORTS, EXPORTS, DEPENDS_ON, INVOKES, SOURCES, EXECUTES CALLS, USES, MEMBER_OF, CONTAINS, DEPENDS_ON_ENV, SETS_ENV, DEFINES, IMPORTS, USES_ENV, EXPORTS Overlapping but distinct edge taxonomies. Gateway has MEMBER_OF (91,815) + CONTAINS (35,109) + SETS_ENV/USES_ENV (community + env modeling). AgentCore has INVOKES/SOURCES/EXECUTES (cross-language shell→Fortran bridge). Each models environment and containment differently.
Node label breakdown Function 87,610 · FortranSubroutine 27,941 · File 17,273 · FortranFunction 5,744 · FortranModule 4,800 · PythonFunction 2,642 · ShellScript 315 (not label-broken-down; Shell graph reported separately) AgentCore exposes a rich label taxonomy incl. 27,941 Fortran subroutines — absent from Gateway's summary. This is why AgentCore answers Fortran code-structure questions in more depth.
Shell graph detail (folded into node labels: ShellScript 315) Explicit: 628 scripts (178 J-Jobs, 12 ex, 12 ush), 2,760 env vars, SOURCES 1,237, INVOKES 2,115 Gateway provides a dedicated Shell-graph breakdown (Phase 27B) with J-Job/ex/ush categorization; AgentCore reports 315 ShellScript nodes without the same operational breakdown. Gateway is stronger for shell-script operational queries.

Note on internal count drift: AgentCore's mcp_health_check snapshot (105,891 nodes / 4,729,093 rels) differs from its get_knowledge_base_status (148,976 nodes / 4,555,408 rels). The two tools count over different scopes/tenants — health rolls up a snapshot while KB-status queries the default gw tenant graph directly. Worth reconciling but not a data-integrity failure.


5. Health Trend (get_health_trend)

Aspect AgentCore EIB Gateway Comparative Analysis
Trend data Empty — "No health history found" Empty — "No health history found" Parity (both empty). Neither server has accumulated a persisted health-snapshot series yet. Each mcp_health_check run persists a snapshot to health_history.jsonl, so trends will populate on subsequent runs. This run seeds the first snapshot on both.

6. RAG Quality Metrics (get_quality_metrics, compare=true)

Metric AgentCore EIB Gateway Comparative Analysis
Benchmark available No — "No benchmark results found" (quality_metrics.jsonl absent) Yes — 2026-03-11, corpus v1.0.0 (60 queries) Biggest observability gap. AgentCore has not run its benchmark harness on the current runtime, so it reports no quality data. Gateway has a full benchmarked baseline.
Precision@5 n/a 0.71 Gateway baseline: 0.71. AgentCore un-benchmarked — cannot compare retrieval quality quantitatively yet.
Recall@5 n/a 0.71 "
MRR n/a 0.93 Gateway's mean-reciprocal-rank is strong (0.93).
Coverage n/a 93% "
Latency P50 / P95 n/a 42ms / 177ms Gateway reports fast median retrieval.
Weakest category n/a Code Structure (P@5 0.40, MRR 0.70) Gateway's known soft spot is code-structure retrieval — ironically the area AgentCore's denser graph should win, but AgentCore lacks a benchmark to prove it.
Strongest categories n/a EE2 (0.89), Semantic (0.88), Operational (0.83) "
Regression vs prior n/a All categories IMPROVED except EE2 P50 (+6% DEGRADED) Gateway trend is healthy; only a minor EE2 latency uptick.

Action item: Run the AgentCore benchmark harness so the two servers can be compared on retrieval quality, not just infrastructure. Until then, quality parity is unproven for the AWS-native server.


7. Knowledge Base Integrity (check_knowledge_integrity, sample_size=50)

Check AgentCore EIB Gateway Comparative Analysis
Overall Issues detected All checks passed Gateway is clean; AgentCore surfaces two warnings.
Path Consistency [WARN] 2/34 sampled docs have checkout-specific prefix [OK] 0/0 sampled AgentCore found a couple of docs with environment-specific path prefixes (a portability nit). Gateway sampled 0 (nothing flagged).
Orphaned Graph Nodes [OK] 17,273 File nodes, 0/20 lack identity [OK] 17,273 File nodes, 0/20 lack identity Exact parity — both graphs are free of orphaned file nodes.
Stale Embeddings [WARN] 12/12 sampled older than source (30-day threshold) [OK] 0/0 current (git comparison unavailable → 30-day threshold) AgentCore flags all 12 sampled embeddings as older than their source — a re-ingest/refresh signal. Gateway reports clean, but note it sampled 0/0 and fell back to a 30-day age heuristic (git comparison unavailable), so its "pass" is weaker evidence than it looks.
Coverage Gap [SKIP] no Fortran files in /supported_repos/global-workflow [SKIP] no Fortran files in supported_repos/global-workflow Parity — both skip Fortran coverage because the on-disk checkout used for the probe has no Fortran (the Fortran lives in submodules not present at the probe path).

Overall Comparative Summary

Theme AgentCore (agentcore-mcp-rag) EIB Gateway (eib-mcp-gateway) Verdict
Platform AWS-native (Neptune + OpenSearch + Bedrock Titan), serverless microVM Node.js + Neo4j + ChromaDB, Docker container Different generations of the same system; AgentCore is the strategic AWS target.
Health HEALTHY 4/4, 9/10 functional HEALTHY 9/9, 6/6 functional Both healthy. Tie.
Multi-tenancy 5 tenants, branch-isolated Single tenant AgentCore wins — branch isolation is a capability Gateway lacks.
Graph richness 148,976 nodes, 95,996 functions, 27,941 Fortran subroutines 19,689 nodes, 2,012 functions, no Fortran expansion AgentCore wins on code-graph depth.
Shell-graph detail 315 ShellScript nodes, no operational breakdown 628 scripts w/ J-Job/ex/ush categorization Gateway wins on shell operational modeling.
Vector corpus 252K docs, dual Titan+MPNet 220K docs, MPNet only + versioned snapshots AgentCore larger + dual-embedding; Gateway keeps version history. Mixed.
Quality benchmark Not generated Full baseline (P@5 0.71, MRR 0.93) Gateway wins — AgentCore quality is un-benchmarked.
Integrity 2 warnings (path prefix, stale embeddings) All pass (but weaker sampling) Gateway cleaner on paper; AgentCore's warnings are actionable but its checks sampled more.
Workflow FS /mnt/workflow not mounted → 1 module degraded Auto-detected, no gap Gateway wins in this environment.
Observability Per-module latency, tenant table, functional matrix Data-validation probe, quality benchmark, regression deltas Complementary. AgentCore is stronger on live runtime introspection; Gateway on quality trend.

Bottom line

Both servers are operationally healthy and expose the same ~51–52 tool surface over the same underlying Global Workflow source tree (file counts match exactly, and the MPNet code corpora are nearly identical). The meaningful differences are architectural and maturity-driven:

  • AgentCore is the richer, multi-tenant, AWS-native platform with a far deeper code graph (esp. Fortran) and dual embeddings — but it has two open gaps in this environment: no quality benchmark generated and the workflow filesystem not mounted, plus minor integrity warnings (path prefixes, stale embeddings signalling a re-ingest is due).
  • Gateway is the mature, single-tenant, benchmarked reference with clean integrity and better shell-operational graph detail — but a smaller graph, single embedding family, and no multi-tenancy.

Recommended follow-ups: (1) run the AgentCore benchmark harness to close the quality-comparison gap; (2) mount /mnt/workflow (EFS) to the AgentCore runtime to restore workflow_info; (3) trigger a re-ingest/embedding refresh on AgentCore to clear the stale-embedding warning; (4) reconcile the node/relationship count drift between AgentCore's mcp_health_check and get_knowledge_base_status.


Generated from live mcp_health_check, get_server_info, get_knowledge_base_status, get_health_trend, get_quality_metrics, and check_knowledge_integrity calls against both servers on July 2, 2026. See also MCP-Parity-Assessment-SoC-Topology for the design-level comparison and MDC-MCP-RAG-AWS-Architecture-v3 for the AgentCore deployment architecture.