The ONLY AI Tech Stack You Need in 2026

Key Concepts

• AI-First Tech Stack: A technology stack designed with Artificial Intelligence as the primary consideration.
• Capabilities Over Tools: Prioritizing the desired functionality and problem-solving over the specific tools used.
• Core Infrastructure: Foundational technologies applicable to all software development.
• AI Agents: Software programs that can perform tasks autonomously, often leveraging LLMs.
• RAG (Retrieval-Augmented Generation): A technique that enhances LLM responses by retrieving relevant information from external knowledge bases.
• Web Automation Agents: AI agents designed to interact with and automate tasks on the web.
• Full Stack Development: Building both the front-end and back-end of an application.
• Deployment & Infrastructure: Processes and tools for getting code into production and managing its environment.
• Self-Hosting: Running software on one's own hardware or servers.
• Open Source (OSS): Software whose source code is available for modification and redistribution.
• Vector Database: A database optimized for storing and querying high-dimensional vectors, commonly used in AI applications.
• Knowledge Graph: A structured representation of information that describes entities and their relationships.
• Observability: The ability to understand the internal state of a system based on external outputs.
• CI/CD (Continuous Integration/Continuous Deployment): Practices for automating the integration and deployment of code changes.
• LLM (Large Language Model): A type of AI model trained on vast amounts of text data, capable of generating human-like text and performing various language tasks.
• MCP (Machine Communication Protocol): A protocol for communication between machines, often used in AI agent interactions.
• OAuth: An open standard for access delegation, commonly used for secure authorization.

Core Infrastructure

This section covers the foundational technologies that underpin all software development.

Database

• Primary Choice: PostgreSQL (Postgres)
• Hosting Platforms: Neon and Superbase.
  • Neon: Noted for its scalability and is a more recent experiment.
  • Superbase: Used for a longer duration and also offers Postgres as part of the local AI package.
• Local AI Integration: Postgres is included in the local AI package for self-hosting.
• Alternatives:
  • MongoDB: A NoSQL database.
  • Firestore: Another NoSQL database.
• Rationale for Postgres:
  • Industry standard for building AI agents.
  • LLMs are perceived to understand SQL queries better than NoSQL queries.
  • Advantages in pricing and scaling have led to widespread adoption.

Caching

• Primary Choice: Redis
  • Key Feature: Blazing fast performance.
• Open-Source Alternative: Valkey
  • Compatibility: Completely compatible with Redis.
  • Integration: Available in the local AI package for local use.

AI Coding Assistant

• Primary Choice: Claude Code
  • Usage: The tool most frequently open on the computer.
  • Enhancement: Used in conjunction with Archon (an open-source project for AI coding assistance knowledge and task management).
• Alternatives:
  • Cursor (2.0): Mentioned as a strong contender.
  • CodeX: Catching up to Claude Code.
• Rationale for Claude Code:
  • Generally considered the best AI coding assistant.
  • Features like slash commands, sub-agents, and Claude skills are highly valued.
  • Despite being pricier and having rate limit issues, it remains the daily driver.

Prototyping & Workflow Automation

• Primary Choice: N8N
  • Usage: For quickly prototyping ideas, validating tools for agents, and system prompts.
  • Transition: Usually moves to a coded solution after prototyping.
  • Key Strengths: Extensive app integrations, AI-focused features, continuous development of agent-creation tools, open-source, and self-hostable.
• Alternatives:
  • Langflow: A visual tool for building LLM applications.
  • Flowise: Another visual tool, previously covered on the channel and included in the local AI package.

AI Agents

This section details the technologies used for building various types of AI agents.

AI Agent Framework

• Primary Choice: Pydantic AI
  • Rationale: Offers a balance between ease of agent building and flexibility, comparable to raw LLM calls.
  • Advantages: Simplifies agent creation, especially when switching between LLM providers. Maintains control and flexibility.
  • Protocol Support: Actively supports new protocols like MCP, A2A, and AGUI.
• Alternatives:
  • Raw LLM Calls: A common approach, respected but can be more complex for some.
  • Other Frameworks: Mentioned as numerous, but Pydantic AI is preferred for avoiding "abstraction distractions."

Multi-Agent Framework

• Primary Choice: LangGraph
  • Usage: Connecting individual agents created with Pydantic AI into complex workflows.
  • Application: Used for use cases that genuinely require multi-agent systems, avoiding over-engineering.
  • Key Features: State management, human-in-the-loop capabilities, agent routing, graph persistence, and a UI for graph visualization.
• Alternatives:
  • CrewAI: Popular for building multi-agent systems.
  • Pydantic AI Graphs: Allows building multi-agent systems within the same framework used for single agents.
• Rationale for LangGraph: Considered the most mature for handling core components of complex workflows, such as human-in-the-loop.

Agent Authorization & Tool Security

• Primary Choice: Arcade
  • Functionality: Handles agent authorization and tool security, which frameworks like Pydantic AI do not.
  • Use Case: Granting agents permission to access user accounts (e.g., Gmail, Slack).
  • Key Feature: Enables secure OAuth flows for agent authorization.
  • New Development: MCP server SDK for building secure MCP servers with integrated tool authorization.
  • Example: Demonstrates an agent requiring OAuth to access a user's Reddit account.
• Alternatives: No direct alternatives mentioned that replicate Arcade's functionality.

Agent Observability

• Primary Choice: Langfuse
  • Functionality: Essential for monitoring AI agents in production.
  • Key Metrics Tracked: Token usage, cost, latency, tool calls.
  • LangGraph Integration: Integrates with LangGraph to visualize multi-agent system decisions and routing.
  • Importance: Crucial for setting up evaluations, A/B testing, and system prompt optimization.
• Alternatives:
  • Langsmith: Popular but not fully open-source or self-hostable.
  • Helicone: Another option for observability.
• Rationale for Langfuse: Most feature-rich and open-source/self-hostable, included in the local AI package.

RAG Agents

This section focuses on tools specifically for Retrieval-Augmented Generation.

Data Extraction

• For Complex Documents (PDFs, Excel, diagrams):
  • Primary Choice: Dockling
    • Strengths: Handles complex documents, easy to use with self-hosted models, open-source.
    • New Addition: A recent addition to the tech stack, simplifying manual extraction.
  • Alternatives: LlamaIndex (agent framework for RAG), Unstructured.
• For Website Data:
  • Primary Choice: Crawl4AI
    • Strengths: Fast, efficient, automatic junk cleaning, LLM integrations for specific text extraction.
    • Decision Process: Use Dockling for files, Crawl4AI for websites.

Data Storage (for RAG)

• Primary Choice: PostgreSQL (with PGVector)
  • Functionality: Acts as a vector database in addition to a regular SQL database.
  • Trade-off: Not as fast as dedicated vector databases.
  • Rationale: Many RAG strategies require both SQL and vector storage for document and user data. Scales extremely well.
• Dedicated Vector Database Alternatives: Quadrant, Pinecone.

Long-Term Memory (RAG Implementation)

• Primary Choice: Memzero
  • Integration: Integrates with any database, including PGVector.
  • Ease of Use: Simple to add to any AI agent, allowing for memory injection into system prompts and extraction.
• Alternative: Zep (not open-source, hence less preferred).
• Note: Langchain's memory features are mentioned as more of an AI agent framework focused on long-term memory.

Knowledge Graphs

• Graph Database (Engine):
  • Primary Choice: Neo4j
    • Strengths: Beautiful UI, easy querying, high-speed, scalable, supported by most knowledge graph libraries.
    • Consideration: Licensing for commercial use.
  • Alternatives: Memgraph, Folklore DB.
• Library for Data Insertion & Search:
  • Primary Choice: Graffiti
    • Functionality: Intelligent entity and relationship extraction from raw text using LLMs.
    • Process: Stores data, formats it for storage, and facilitates searching.
  • Alternative: Lightrag (more of a vector database and knowledge graph hybrid).
• Rationale for Separation: Prefers separating vector database and knowledge graph functionalities.

Evaluation

• Primary Choice: Regos (pronunciation uncertain)
  • Functionality: Setting up evaluation pipelines with specialized RAG metrics (faithfulness, relevance).
  • Capabilities: Automated test dataset generation, works with any LLM provider.
• Note: Langfuse is used for general agent evaluations, focusing on tool calling.

Web Search (for RAG)

• Primary Choice: Brave
  • Strengths: High-level, faster implementation for AI agents, privacy-focused, no tracking, independent index, built-in AI search.
• Alternative: Perplexity (more detailed but slower).
• Rationale: Essential for accessing general knowledge beyond a private knowledge base.

Web Automation Agents

This section covers tools for agents that interact with the web.

Web Data Extraction (Live)

• Primary Choice: Crawl4AI
  • Usage: Providing agents with tools to extract information from specific URLs in real-time.
  • Strengths: Open-source, fast, feature-rich.
• Alternatives: Firecrawl.
• Note: This differs from RAG data extraction, which is done ahead of time.

Social Platform Automation

• Tools for LinkedIn, X, Instagram:
  • Choices: Ampify, Bright Data.
  • Reason: Crawl4AI does not handle social platforms well.

Browser Automation

• Primary Choice: Playwright
  • Usage: Simpler browser automations, web testing, visual validation of website changes by AI coding assistants.
  • Key Feature: Playwright MCP server is a "golden nugget" for AI coding assistants working on frontends.
  • Strengths: Great multi-browser support, considered the "deterministic web automation king."
• Alternatives: Puppeteer, Selenium (used for years but switched to Playwright).

Advanced Browser Control

• Primary Choice: Browserbase
  • Functionality: Enables agents to control a browser live.
  • Key Features: Managed infrastructure, session recording and storage, anti-bot detection, secure.
  • Components:
    • Stagehand MCP Server: Spins up sessions with natural language requests.
    • Director: AI on top of Playwright (and other tools) for browser control. Allows natural language requests for web tasks, shows step-by-step execution, and provides code snippets.
• Underlying Technology: Stagehand and Director are built on Playwright.

Full Stack Development

This section covers technologies for building backends and frontends, often to support AI agents.

APIs

• Primary Choice: FastAPI
  • Rationale: Python-based, aligning with Python-based AI agent development. More feature-rich than Flask.
• Alternative (TypeScript): Express.

Database

• Primary Choice: PostgreSQL (as covered in Core Infrastructure).

Authentication

• Simple Authentication: Superbase
  • Usage: Go-to for straightforward authentication needs.
• Enterprise Authentication: Auth0
  • Use Cases: MFA, universal login, enterprise SSO (SAML, Active Directory).
  • Integration: Can be used with Superbase for database needs while leveraging Auth0 for advanced authentication.
• Other Enterprise Alternatives: Clerk, Okta.
• Transparency: Auth0 is the chosen solution, and other options haven't been extensively explored.

Front-end Library

• Primary Choice: React
  • Rationale: Simple, preferred by AI coding assistants and front-end builders.
  • Build Tool: Vite (leads to snappy, quick, lightweight applications).
• Alternatives:
  • Next.js: Respected but recent version changes have caused compatibility issues.
  • Vue: Another good option.

Component & Styling Library

• Component Library: Shadcn UI
• Styling: Tailwind CSS
  • Note: These are considered standard and work well.

Agentic Front-end Builders

• Purpose: Creating beautiful UIs, which AI coding assistants may not excel at.
• Primary Choice: Lovable
  • Strengths: System prompting for UI creation, great integrations, new agent mode.
• Alternatives: Bolt New, Bolt.diy (open-source version).

UI Prototyping

• Primary Choice: Streamlit
  • Functionality: Easiest way to create user interfaces directly in Python.
  • Use Case: Prototyping AI agents with a nice UI without building a full React application.
  • Benefit: No need for separate backend/frontend connections during prototyping.

App Monitoring & Analytics

• Primary Choice: Sentry
  • Strengths: Great real-time analytics, native AI features in development, extensive integrations.
• Alternatives: PostHog, Google Analytics.

Payments

• Primary Choice: Stripe
  • Rationale: Best developer experience and fantastic documentation.
• Alternatives: Lemon Squeezy, Paddle.

Deployment & Infrastructure

This section covers getting code into production, CI/CD, and testing.

Deployment Platforms

• Primary Choice: Render
  • Strengths: Simplicity, infrastructure as code (YAML), Git-based deployments, free hosting for frontends, CDN, background workers, cron jobs.
• Alternatives: Fly.io, Netlify.
• Enterprise/Cloud:
  • Primary Choice: Google Cloud Platform (GCP)
    • Usage: For enterprise requirements, SLAs, compliance.
    • Feature: Serverless functions.
• GPU-Heavy Workloads:
  • Primary Choice: RunPod
    • Strengths: Cheapest reliable GPU hosting, spot instances for lower cost, instant GPU availability (no queues).
  • Alternatives: TensorDock (cheaper but less reliable), Lambda Labs.

Virtual Machines

• Primary Choice: DigitalOcean
  • Usage: Owning and managing machines, hosting local AI packages in the cloud.
  • Strengths: Reliable, predictable pricing, AI integrations (App Platform for managed databases, local LLMs, RAG).
• Alternatives: Hostinger (KVM offering), Hetzner (affordable).

Containerization

• Primary Choice: Docker
  • Functionality: Industry standard for deploying applications, creating isolated environments, solving "works on my machine" problems.
• Alternative: Podman
  • Reason for Consideration: Licensing concerns with Docker.
  • Note: While a good alternative, Docker is considered more robust.

CI/CD

• Primary Choice: GitHub Actions
  • Strengths: Simple, integrated with GitHub repositories, free for public repos, generous pricing for private repos, large marketplace of actions.
  • AI Assistance: AI coding assistants are good at generating YAML for workflows.

Testing

• Python Testing: Pytest
• TypeScript Testing: Jest
  • Note: Both are standard, easy to use, and support mocking and fixtures for reliable testing.

AI Code Review

• Primary Choice: CodeRabbit
  • Functionality: Thoroughly reviews pull requests, including security vulnerability detection.
  • Integration: Used with Archon for automatic pull request reviews.
  • Key Benefit: Completely free for open-source repositories.

Self-Hosting & Local AI

This section covers tools for running AI models and services locally.

Local LLM Chat Platform

• Primary Choice: Open WebUI
  • Functionality: ChatGPT-like interface running locally.
  • Features: Custom agents via functions/pipelines, RAG integration.
• Alternative: AnythingLLM.

Local Web Search

• Primary Choice: CRXNG

Local LLM Serving

• Primary Choice: Ollama
  • Functionality: Serves open-source LLMs locally.
  • Strengths: Easiest to use, auto-leverages multiple GPUs, supports quantization, configurable context limits.
• Alternatives: VLM, LiteLLM.

HTTPS/TLS

• Primary Choice: Caddy
  • Strengths: Simplest option for managing domains for self-hosted services.
• Alternatives: Traefik, Nginx.

Synthesis & Conclusion

The presented tech stack is deeply integrated with an "AI-first" philosophy, prioritizing capabilities over specific tools. The core infrastructure relies on PostgreSQL for databases and Redis for caching, with Claude Code serving as the primary AI coding assistant. For AI agent development, Pydantic AI is used for individual agents, and LangGraph for multi-agent orchestration, complemented by Arcade for crucial authorization and Langfuse for essential observability.

RAG agents leverage Dockling and Crawl4AI for data extraction, PostgreSQL with PGVector for storage, Memzero for long-term memory, Neo4j and Graffiti for knowledge graphs, Regos for evaluation, and Brave for web search. Web automation agents utilize Crawl4AI for general web scraping, Ampify/Bright Data for social platforms, Playwright for browser automation, and Browserbase for advanced browser control.

Full-stack development employs FastAPI for APIs, Superbase for authentication (with Auth0 for enterprise needs), React with Vite for frontends, and Shadcn UI/Tailwind CSS for styling. Agentic front-end builders like Lovable and prototyping tools like Streamlit are also key. Deployment is handled by Render for simplicity, GCP for enterprise, and RunPod for GPU workloads, with Docker for containerization and GitHub Actions for CI/CD. Testing is done with Pytest and Jest, and AI code review is powered by the free CodeRabbit for open-source projects. For local AI, Open WebUI, Ollama, and Caddy are highlighted.

The overarching message is to use these recommendations as a guide to fill gaps in one's own tech stack, emphasizing adaptability and problem-solving rather than tool expertise. The creator acknowledges working with some of these teams and thanks them for their collaboration in showcasing their tools.