Skip to content
MCP AI Agent Guidelines
Architecture diagram showing the MCP hub connected to instruction tools and skill domains Architecture diagram showing the MCP hub connected to instruction tools and skill domains

MCP AI Agent Guidelines

20 tools. 102 skills. 18 domains. Physics-inspired AI workflows.
Get started Browse tools Browse skills GitHub

What is this?

Section titled “What is this?”

MCP AI Agent Guidelines is a Model Context Protocol server whose public surface is built from 20 instruction tools — high-level mission verbs (design, implement, review, debug, govern…) — each backed by richly orchestrated skill pipelines underneath. You invoke one tool; the server orchestrates dozens of skills internally.

// Claude Desktop / VS Code — add to your MCP config
{
  "mcp": {
    "servers": {
      "ai-agent-guidelines": {
        "command": "npx",
        "args": ["-y", "mcp-ai-agent-guidelines@latest"]
      }
    }
  }
}

Six Pillars

Section titled “Six Pillars”

Instruction-First Architecture

20 public tools act as mission verbs. Each instruction orchestrates multiple internal skills — consumers never call skills directly. Learn more →

Physics-Inspired Analysis

30 QM + GR skills provide bias-free code analysis through quantum-mechanics and general-relativity metaphors. Coupling, entropy, gravitational debt — all objectively measurable. Learn more →

Bio-Inspired Routing

5 adapt-* skills — Ant Colony Optimization, Hebbian reinforcement, Physarum network pruning, quorum sensing, simulated annealing — make multi-agent workflows self-optimizing over time. Learn more →

Governance-First

7 gov-* skills, criticalSkillGuard env-var gating, HMAC session integrity, and prompt-injection hardening are built-in — not optional add-ons. Learn more →

Model-Agnostic Orchestration

Role-class routing resolved from orchestration.toml at runtime. Physical model IDs are never hardcoded. 5 orchestration patterns documented. Learn more →

Zero Runtime LLM Calls

Advisory outputs only — this is a routing and orchestration layer, not an LLM executor. Works with any MCP-compatible host. Architecture →

Instruction Tools

Section titled “Instruction Tools”
Zero-Costfree
REASONING
Basic
SPEED
Fastest
PRICE
Free
$0 / 1M tokens
CONTEXT
128–192K
prompt window
Efficientcheap
REASONING
Moderate
SPEED
Fast
PRICE
~$0.30
per 1M tokens
CONTEXT
~160K
prompt window
Advancedstrong
REASONING
Highest
SPEED
Medium
PRICE
$2.5–$15
per 1M tokens
CONTEXT
160–400K
prompt window
Cross-Modelreviewer
REASONING
Highest
SPEED
Sequential
PRICE
$1–$15
per 1M tokens
CONTEXT
173K+
prompt window
system-design Architecture and system design — Advanced
feature-implement Build new features or tools — Advanced
evidence-research Synthesis, comparison, recommendations — Advanced
code-review Code, quality, and security review — Cross-Model
strategy-plan Strategy, roadmap, and sprint planning — Advanced
issue-debug Diagnose and fix problems — Efficient
code-refactor Improve existing code safely — Efficient
test-verify Write, run, and verify tests — Efficient
docs-generate Generate documentation artifacts — Efficient
quality-evaluate Benchmark and assess quality — Cross-Model
prompt-engineering Build, evaluate, and optimize prompts — Efficient
agent-orchestrate Compose multi-agent workflows — Advanced
enterprise-strategy Leadership and enterprise scale — Advanced
policy-govern Safety, compliance, and guardrails — Advanced
fault-resilience Self-healing and fault tolerance — Advanced
routing-adapt Bio-inspired adaptive routing — Advanced
physics-analysis QM and GR code metaphors — Advanced

Skill Domains

Section titled “Skill Domains”
Requirements (4) req-* — Analysis, scope, ambiguity detection, acceptance criteria
Architecture (4) arch-* — System, reliability, scalability, security design
Quality (5) qual-* — Code analysis, review, performance, security, refactoring priority
Debugging (4) debug-* — Assistant, root cause, reproduction, postmortem
Documentation (4) doc-* — Generator, README, API reference, runbook
Evaluation (5) eval-* — Prompt, output grading, variance, design, benchmark
Benchmarking (3) bench-* — Analyzer, blind comparison, eval suite
Workflows (3) flow-* — Orchestrator, context handoff, mode switching
Governance (7) gov-* — Policy, compliance, guardrails, prompt-injection hardening
Orchestration (4) orch-* — Agent orchestrator, delegation, multi-agent, synthesis
Prompting (4) prompt-* — Engineering, chaining, refinement, hierarchy
Research (4) synth-* — Comparative, research, recommendation, engine
Strategy (4) strat-* — Advisor, roadmap, prioritization, tradeoff
Resilience (5) resil-* — Homeostatic, membrane, voter, replay, clone-mutate
Adaptive (5) adapt-* — ACO router, annealing, Hebbian, Physarum, quorum
Leadership (7) lead-* — Capability mapping, roadmap, exec briefing, mentoring
Quantum Mechanics (15) qm-* — Physics-metaphor code analysis using QM principles
General Relativity (15) gr-* — Physics-metaphor code analysis using GR principles

Orchestration Patterns

Section titled “Orchestration Patterns”
Parallel Critique → Synthesis

Advanced model generates plan → independent critic reviews → model reconciles. Used for architecture decisions and high-risk design.

Read more →
Draft → Review Chain

Zero-cost model drafts → advanced model reviews → zero-cost model applies notes. Cuts advanced-model token cost ~60%.

Majority Vote

3 zero-cost/efficient lanes vote → tiebreak escalation → final arbitration. Used for evaluation and classification tasks.

Cascade with Fallback

Try efficient → fallback to advanced → emergency escalation. Implements homeostatic PID setpoint concept.

Free Triple Parallel + Synthesis

3 zero-cost lanes run concurrently → single advanced synthesis pass. ~80% cheaper than running advanced end-to-end.

Read full orchestration docs →