MCP server architectures,
evidence-backed.
Stop guessing what good MCP server design looks like. Production-grade architecture from a corpus of 13 scored reference servers, 29 patterns, and 11 antipatterns — synthesized into 8 reusable archetypes.
// 30-day refund · lifetime updates 12mo · cross-AI portable
// problem_statement
Every MCP tutorial says "use atomic tools".
None of them tell you which ones for your domain.
01
Maturity ≠ design quality
GitHub's official MCP scored 49/70 in the corpus — held back by god-tool dispatch and conditional required params. The simpler reference Filesystem MCP scored 59/70. The biggest project isn't always the best teacher.
02
Safety is non-negotiable, design is hard
Restricting an MCP server to "read-only" by string-matching SQL prefixes is safety theater (`; DROP TABLE` bypasses it). The corpus maps the AP-03 antipattern as a mitigation spectrum: same tool shape, different safeguards, different verdicts — from lethal (`mcp-bash`, no sandbox) to sanctioned (Cloudflare's V8-isolated Code Mode, 99.6% token reduction over 2,594 endpoints).
03
Patterns transfer, but only with citations
Stripe + Sentry + Cloudflare all converged on hosted-remote OAuth (P-21) — three independent domains, same pattern. Sentry's resource-resolver tool (P-26) showed how to collapse five detail handlers into one. Brodels pioneered episodic audit logs (canonical for D10). Every recommendation cites its evidence — so you can verify and adapt.
// what_you_get[]
9 skills.
One bundle.
The router auto-classifies your domain. The 8 archetype skills supply the prescriptive content. Each archetype includes a spec.md loaded on-demand — tool surface, auth, safety rails, antipatterns, worked skeleton, README requirements.
router · auto-invoked first
mcp-architecture-router
Maps the buyer's domain to one or more of 8 archetypes via a decision tree. Synthesizes hybrid architectures when multiple archetypes apply (the common case). Escalates to premium tier when high-risk patterns are present (money movement, prod infra, code merges, PII).
Read-Only Data Source
Databases, files, web content, search APIs, KBs. The LLM only reads — never mutates.
Postgres, S3, knowledge bases, REST APIs
Writable System
Create / update / delete in external systems. HIGH-RISK — always recommended with safety rails.
Stripe, Salesforce, Slack, internal CRMs
Search & Fetch
Two-step retrieval: search ranks summaries, fetch drills into chosen IDs.
Memory graphs, docs search, ticket search
Workflow Orchestrator
Multi-phase pipelines (CI/CD, AI agents, ETL). Cost + audit + safety per phase.
AI engineering pipelines, CI/CD, ETL
State / Memory
Cross-session memory: knowledge graphs, agent learning, shared team context.
Knowledge graphs, agent KBs, personal memory
UI / Browser Automation
Drive a browser or GUI. Accessibility-tree-first; pixel-vision opt-in.
Playwright, E2E tests, admin UIs
Observability / Telemetry
Investigate errors, logs, traces, metrics. Time-bounded, cost-aware queries.
Sentry, Datadog, PagerDuty, log aggregators
Code / Dev-Tools
Repos, PRs, CI, deploys. Inherits Writable rails + draft-by-default, branch protection.
GitHub, GitLab, Vercel, deploy pipelines
// how_it_works.sh
From "I want to build an MCP for X" to a production-grade spec in minutes.
Install
unzip mcp-architecture-skills-v0.1.zip -d ~/.claude/skills/
Drop the 9 skills into your Claude skills directory. Or commit to your project's .claude/skills/ for team-scoped use.
Describe
"I want Claude to interact with our Postgres + Stripe."
In any Claude conversation, describe the system you want an MCP server for. The router auto-invokes and classifies.
Deliver
→ archetype 01 + 02 hybrid → 14-section spec → tool surface · auth · rails
Get a 2,000-6,000 word architecture spec: tool list with annotations, auth, safety rails, README requirements, antipatterns to avoid, first tool to implement.
// evidence_base
Every recommendation, cited.
No "best practices" floating in space. Every pattern, antipattern, and architectural choice in the bundle traces back to a specific server in the corpus — which is public and auditable.
Sentry 65 · Filesystem 59 · Stripe 58 · Cloudflare 57 · ...
P-01 through P-29 with corpus citations
AP-01 through AP-11 (god-tools, safety theater, etc.)
Plus 1 router skill that picks for you
Sample scoring
// pricing
Four ways in.
Pick the one that fits.
The free corpus is genuinely useful by itself. The skill stack is for self- serve buyers. Audits are for buyers who want a tailored, human-reviewed spec. Consulting is for deep architectural review.
Free Corpus
The evidence behind everything. 13 scored servers, rubric, patterns, antipatterns. MIT-licensed.
- ▸13 reference servers scored on 14-dim rubric
- ▸29 design patterns with citations
- ▸11 antipatterns with corpus examples
- ▸Free forever, public, auditable
Skill Stack
The 8 prescriptive archetype specs (not in the free corpus) + auto-invoking router. Describe your domain → get a 14-section spec back. Install into ~/.claude/skills/.
- ▸Router skill + 8 archetype skills
- ▸Each with prescriptive spec.md loaded on-demand
- ▸12 months of updates
- ▸Cross-AI portable (Codex, Gemini, Cursor)
- ▸30-day refund if not useful
Audit (Baseline)
Tailored architecture spec for YOUR domain. Within 24 hours. For read-mostly, internal, low-stakes systems.
- ▸Domain-specific tool surface
- ▸Auth, pagination, safety rails
- ▸Antipatterns to avoid for your domain
- ▸Corpus citations for every claim
- ▸Delivered within 24 hours
Audit (Premium)
Baseline + blast-radius + threat model + recovery plan. For money, prod infra, PII, code merges.
- ▸Everything in Baseline
- ▸Per-tool blast-radius analysis
- ▸Per-tool prompt-injection threat model
- ▸Recovery plan for LLM mistakes
- ▸Human-in-the-loop policy recs
- ▸Delivered within 48 hours
Concierge consulting
1-hour deep-dive — $750
Zoom call + written follow-up reviewing your existing or planned MCP server. For architectural review of a real system, tradeoff discussions, or post-audit implementation questions.
// faq
Honest answers.
// who_built_this
Built by an engineer who's been
designing systems for 30+ years.
Gunnard Engebreth — independent software engineer focused on AI engineering, two Apress books on PHP 8, 9 live SaaS products. The MCP corpus was built by reading every reference server's source code, scoring them on a 14-dimension rubric, and synthesizing the patterns and antipatterns into reusable archetypes.