--- title: "Our AI Stack, StudioMeyer Academy" description: "Which models, memory systems and MCP servers we run ourselves. Transparent, tool-agnostic, no proprietary lock-in." canonical: "https://studiomeyer.academy/en/ai-stack" language: en image: "https://studiomeyer.academy/api/og/page/share?title=Our+AI+Stack%2C+StudioMeyer+Academy§ion=StudioMeyer+Academy&locale=en" markdown_versions: ["https://studiomeyer.academy/ai-stack.md", "https://studiomeyer.academy/en/ai-stack.md", "https://studiomeyer.academy/es/ai-stack.md"] publisher: "StudioMeyer Academy, https://studiomeyer.academy (llms.txt: https://studiomeyer.academy/llms.txt)" --- Transparent # Our *AI stack* . No black-box course. Every component we teach you, we run ourselves. When you reach what we are at Level 4 to 6, the same code is running as on this page. Tool-agnostic, memory-first, no proprietary lock-in. 3 Claude models · Sonnet, Haiku, Opus 50+ StudioMeyer Memory MCP tools 14 codebase-memory-mcp tools SM-2 spaced-repetition algorithm 16 Postgres tables · EU host OSS stack open-sourced component by component Models ## Which AI we use when. Different tasks need different models. The tutor needs depth. The support chatbot needs speed. Reviews need both plus a wide context window. AI tutor Claude Sonnet Answers lesson questions right in the chat. 80 questions per day per user. 1M context window. AnthropicFree Support chatbot Claude Haiku Answers Academy questions at the bottom right. Faster, cheaper, enough for orientation questions. AnthropicFree Review agents Claude Opus Content reviews, security audits, architecture analyses. Orchestrated through the nex-hq agent framework. internalResearch Code explorer Claude Code (CLI) The development workflow behind Academy itself. Taught as real tooling in Level 5 and 6. Anthropic CLI Memory ## Why no system starts from zero. The difference between using AI and working with AI: persistent context. Every session should build on the last, every mistake should happen once and then live in the system. That is the core of Level 4. Persistent memory StudioMeyer Memory (mcp-nex) 56 tools for sessions, learnings, decisions, entity graph. Stores what we do so no session starts from zero. OSSMulti-tenant Knowledge graph codebase-memory-mcp Indexes the Academy code as a graph. 14 tools. 740 nodes, 912 edges. Enables targeted impact analysis before changes. internal Spaced repetition SM-2 algorithm The Academy review deck uses the proven SuperMemo-2 algorithm. No AI, just math. publicdeterministic Platform ## What keeps Academy running. Boring-technology philosophy. Proven open-source components, clean separation of concerns, idempotent migrations, and Docker for everything. No SaaS dependency except Stripe and email. Framework Next.js 16.2 App Router, standalone output. React 19.2 as the engine. Server components for performance. VercelMIT Database PostgreSQL + Prisma 7.8 16 tables: Users, Enrollments, Progress, Achievements, Quizzes, ReviewItems, Certificates. Source of truth: idempotent init.sql. OSSEU host Auth Magic link (Resend SMTP) Passwordless. 15-min token, race-condition protected, HMAC session cookie. No OAuth, no passwords. EU sending Payments Stripe Checkout Subscriptions and one-time payments. Webhook idempotency. We never see card data. Dublin Hosting Hetzner dedicated Docker Compose on your own server. Nginx reverse proxy, Let's Encrypt auto-SSL, Cloudflare in front. Germany Distribution mcp-academy (npm) 21 tools as an stdio MCP for Claude Code / Cursor / Claude Desktop. Bearer token auth. Public on GitHub. MITOSS Agents ## How we review the platform itself. Three specialised agents, parallel, with their own tool kits. The same pattern you learn in Level 5, Matthias runs them on the Academy code itself before new features go live. Conductor Agent orchestrator Starts three parallel specialist agents for code review, content review, and research synthesis. nex-hq Research Research agent Combines SearXNG + Brave + Anthropic search + optional Tavily. No single-source bias. free-first Critic Critic agent Looks for mistakes, hallucinations, and contradictions. Built for 7 security fixes in session 855. internal Analyst Analyst agent Contextualises findings against existing docs and decision history. Prevents reinventing old answers. internal Honest note ## This is a stack, not the recipe. These components do not solve business problems on their own. They give you the building blocks so you can build your own solutions. What we teach in Level 4 to 6 is not which buttons to press, but why, in what order, and what you do when a model hallucinates, the context window fills up, or the memory layer contains contradictions. The components are interchangeable. Claude can be replaced by GPT or Gemini tomorrow. Stripe can be replaced by Paddle. The memory layer can be Mem0 or Zep. What stays: the discipline of knowing why which component is in there and what for. If that interests you, the best place to start is [Level 1](https://studiomeyer.academy/levels/1) . ## Want to work with this? Level 4 is the first step in. One lesson to get started. [Open Level 4](https://studiomeyer.academy/levels/4) [Read again](https://studiomeyer.academy/ai-stack#top) ## Sitemap Public pages of this site as Markdown: [sitemap](https://studiomeyer.academy/sitemap.md). Curated entry point: [llms.txt](https://studiomeyer.academy/llms.txt), full text: [llms-full.txt](https://studiomeyer.academy/llms-full.txt).