DEVELOPED BY RAPHAEL MORAES
MODIFIED BY INDEST LABS
MODIFIED BY INDEST LABS
GenAI Technology Stack Setup Canvas
Opinionated, layered AI-native engineering stack for Vibe Coders · v1.1
★ Personal Standard Depends on: Claude RulesDepends on: Claude SkillsDepends on: Local SetupDepends on: Local & Remote Infrastructure
Architecture Overview
1
Client Layer
How you interact with AI — the tools, IDE, and configuration that shape every session. This is where Rules and Skills are defined and where Claude Code runs.
2
Standards & Frameworks Layer
How you build everything — MCP servers, agents, security controls, and input/output contracts. Defines the engineering standards that govern every artifact you produce.
3
AI Gateway Layer
The runtime control plane — routes LLM calls, proxies MCP servers, publishes agents as endpoints, and enforces budget and rate controls. Powered by LiteLLM.
4
Observability Layer
Full visibility across every layer — traces, logs, metrics, and LLM evaluation. Built on OpenTelemetry and Arize Phoenix. Nothing runs as a black box.
Layer 1 — Client Layer
1
AI Model Tool
Claude Code (primary) + Claude Desktop
- Claude Code: terminal-first agentic coding — runs bash, edits files, manages git
- Claude Desktop: conversational AI, file & image analysis, document processing
- Both use the same underlying model but serve different interaction patterns
Claude CodeClaude Desktop
Depends on: Claude RulesDepends on: Claude Skills
2
IDE Setup
VSCode
- VSCode: primary editor with Claude Code integration
- Extensions: GitLens, Docker, YAML, REST Client, Thunder Client
- Settings: auto-save, format on save, linting integration
VSCodeGitLensDocker Ext
3
Configuration Layer
Rules & Skills
- CLAUDE.md: project rules, coding standards, tool permissions
- Skills: reusable slash-command workflows for common tasks
- MCP config: server list, tool permissions, resource access
CLAUDE.mdSkillsMCP Config
Depends on: Claude RulesDepends on: Claude Skills
4
MCP Servers
Model Context Protocol — tool layer for Claude
Local
Filesystem MCP
Read/write project files
Depends on: Local Setup
Git MCP
Branch, diff, commit, log
Depends on: Local Setup
Docker MCP
Container management
Depends on: Local Setup
Postgres MCP
Query + schema inspection
Depends on: Local Setup
Remote
Prefect MCP
Workflow orchestration
Depends on: Local & Remote Infrastructure
Databricks MCP
Data platform integration
Depends on: Local & Remote Infrastructure
Kubernetes MCP
EKS / GKE cluster ops
Depends on: Local & Remote Infrastructure
ArgoCD MCP
GitOps deployment
Depends on: Local & Remote Infrastructure
5
Agents
Specialized AI agents in the stack
Local
Claude Code CLI
★ Personal Standard Agentic coder — primary dev agent
Depends on: Claude RulesDepends on: Claude Skills
HolmesGPT
Incident investigator agent
Depends on: Local Setup
Remote / CI
PR-Agent (Qodo)
Automated code review
Depends on: Local & Remote Infrastructure
GitLab Duo
CI-integrated AI coder
Depends on: Local & Remote Infrastructure
GitHub Copilot
IDE + PR AI assistance
Depends on: Local & Remote Infrastructure
Metis (Arm)
Security agent for CI
Depends on: Local & Remote Infrastructure
Layer 2 — Standards & Frameworks Layer
6
MCP Server Standard
FastMCP (Python)
- FastMCP as the only MCP server framework
- Poetry for dependency management
- Pytest + coverage for all servers
- Typed inputs/outputs — no untyped tools
FastMCPPoetryPytest
7
Agent Standard
LangGraph + CrewAI
- LangGraph for stateful multi-step agent graphs
- CrewAI for multi-agent crew orchestration
- All agents emit OTEL traces
LangGraphCrewAI
8
Code Quality Standard
Linting + Security
- Megalinter: unified linting for all languages
- Trivy: container + IaC vulnerability scanning
- Pre-commit hooks enforce before every push
MegalinterTrivy
9
Security Standard
OWASP MCP Top 10 + 1Password
- OWASP MCP Top 10 as the security checklist for every MCP server
- 1Password Connect for secrets — no .env files in repos
- Chainguard / Docker Hardened Images for containers
OWASP MCP1PasswordChainguard
10
Data Standard
PostgreSQL + pgvector
- PostgreSQL as the standard relational + vector database
- pgvector for embedding storage and similarity search
- Databricks for large-scale data processing
PostgreSQLpgvectorDatabricks
11
Standard Inputs
What every MCP server / agent accepts
- Typed Pydantic models for all tool parameters
- JSON Schema validation at boundary
- Structured prompts with role + task + constraints
- Context injection via MCP resources (not raw file paths)
12
Standard Outputs
What every MCP server / agent returns
- Typed Pydantic response models
- Structured JSON — never raw string blobs
- OTEL span attached to every tool call
- Error objects with code + message + trace_id
Layer 3 — AI Gateway Layer (LiteLLM)
LLM Router
★ Personal Standard Route calls to any provider
- Anthropic, OpenAI, Gemini, local Ollama
- Model fallbacks + load balancing
- Cost tracking per model
MCP Proxy
Remote MCP over SSE
- Expose local MCP servers remotely
- Auth + rate limiting
- Server-Sent Events transport
Agent Endpoint
Publish agents as APIs
- LangGraph / CrewAI agents as REST endpoints
- OpenAI-compatible API surface
- Session management
Budget Controls
★ Personal Standard Cost + rate enforcement
- Per-user / per-key spend limits
- Token rate limits
- Budget alerts
OTEL Export
★ Personal Standard Traces to observability layer
- All LLM calls emit OTEL spans
- Token usage as metrics
- Latency + error tracking
Layer 4 — Observability Layer
Arize Phoenix
★ Personal Standard LLM tracing & evaluation
- Trace every LLM call end-to-end
- Evaluate responses for quality + safety
- Dataset curation from traces
OpenTelemetry
★ Personal Standard Universal trace/metric standard
- OTEL SDK in every MCP server and agent
- Spans propagated across service boundaries
- Export to Phoenix + any OTLP backend
Metrics & Logs
Structured observability
- Prometheus-compatible metrics from LiteLLM
- Structured JSON logs (no raw print statements)
- Log correlation via trace_id
Alerting
Proactive incident detection
- Cost spike alerts via LiteLLM budget controls
- Error rate thresholds in Phoenix
- HolmesGPT for AI-driven incident investigation
Tools Reference
AI Model
Claude Code
claude.ai/code
AI Model
Claude Desktop
claude.ai/download
IDE
VSCode
code.visualstudio.com
Configuration
CLAUDE.md Reference
docs.anthropic.com/en/docs/claude-code/memory
MCP Framework
FastMCP
github.com/jlowin/fastmcp
Package Mgmt
Poetry
python-poetry.org
Testing
Pytest
docs.pytest.org
Linting
Megalinter
megalinter.io
Security Scan
Trivy
trivy.dev
Security
OWASP MCP Top 10
owasp.org/www-project-mcp-top-10
Secrets
1Password Connect
developer.1password.com/docs/connect
Orchestration
LangGraph
github.com/langchain-ai/langgraph
Agent Framework
CrewAI
crewai.com
Database
PostgreSQL + pgvector
github.com/pgvector/pgvector
Containers
Chainguard Images
chainguard.dev
Containers
Docker Hardened Images
hub.docker.com/search?q=&type=image&image_filter=official
AI Gateway
LiteLLM
litellm.ai
Observability
Arize Phoenix
phoenix.arize.com
Observability
OpenTelemetry
opentelemetry.io
Agent — Investigator
HolmesGPT
holmesgpt.dev
Agent — Coder
Claude Code CLI
docs.anthropic.com/en/docs/claude-code
Agent — Reviewer
PR-Agent (Qodo)
github.com/qodo-ai/pr-agent
Agent — Security
Metis (Arm)
github.com/arm/metis
Agent — CI Coder
GitLab Duo
about.gitlab.com/gitlab-duo
Agent — CI Coder
GitHub Copilot
github.com/features/copilot
MCP — Orchestration
Prefect
prefect.io
MCP — Data
Databricks
databricks.com
MCP — GitOps
ArgoCD MCP
argo-cd.readthedocs.io
MCP — Infra
Kubernetes EKS/GKE
kubernetes.io
Indest Labs · GenAI Engineering Standard v1.0 · 2026