See AI First — The Opinionated AI Stack Guide

Curated directory of 66 AI developer tools across 13 categories.

Protocols — Standardized Communication Layer

MCP — Model Context Protocol — Anthropic's connection standard — lets AI talk to any tool, database, or API through one unified protocol. Often called 'USB-C for AI'.
A2A — Agent2Agent Protocol — Google's protocol — lets AI agents communicate and collaborate with each other, regardless of their underlying framework.

Agent Frameworks — Building AI Agents

LangGraph — Stateful agent orchestration framework from the LangChain ecosystem. The most widely adopted today.
CrewAI — Build role-based AI agent teams with minimal code. Each agent has a role, they coordinate automatically.
AutoGen — Microsoft's framework for multi-agent conversations. Flexible architecture, easy to extend.
OpenAI Agents SDK — OpenAI's lightweight Python SDK — build agents fast with built-in safety guardrails.
Google ADK — Google's agent building toolkit — with built-in Gemini, Vertex AI, and A2A protocol integration.
LlamaIndex — The RAG specialist — connects AI to your private data for accurate answers from internal documents.
Rasa — Chatbot and voice AI platform — runs on your own server with full data control.
Browser-Use — Browser automation framework for AI agents — control the web with Python using any LLM
Pydantic AI — Type-safe agent framework in the style of FastAPI — built by the Pydantic team, supports 40+ model providers
Mastra — TypeScript-first agent framework from the Gatsby team — workflows, RAG, evals, and MCP built in

Platforms & Config — AI Platforms & System Files

Claude.ai (Web) — Anthropic's chat workspace — with memory, web search, Claude Cowork, and team plugins/connectors for enterprise workflows.
ChatGPT (Web) — OpenAI's most popular AI chat — with GPTs, search, tools, and a built-in memory system.

Memory Systems — Cross-Session Memory

Claude-Mem — Plugin that gives Claude Code persistent long-term memory — automatically records experience, compresses intelligently, and remembers across sessions.
Other Memory Systems — Beyond Claude-Mem, several other memory systems exist: MemGPT/Letta — self-manages context windows like an OS manages RAM, with a paging mechanism.
MemOS — Unified memory operating system for LLMs — manages memory via MemCube with short-term and long-term support

Skills Ecosystem — "Bigger than MCP"

Universal Agent Skills — Skills are .md files with instructions for AI agents — simple yet effective, runnable on any AI coding tool.
Notable Skill Categories — Key categories: Engineering & DevOps (API testing, Docker, CI/CD), Web Development (React, Tailwind, animations), AI/ML Research (77 skills for paper reading, fine-tuning), Cross-Agent Orchestration (handoff between Claude Code ↔ Codex ↔ Cursor), Claude Code Ecosystem (Task Master, Claude Swarm, session management).
Superpowers — Agent-style skills framework for Claude Code — automatically activates the right skills based on context.
Agents (wshobson) — Collection of 112 agents + 146 skills + 73 specialized plugins for Claude Code — the largest to date.
Planning with Files — Markdown-file-based planning skill inspired by Manus — maintains context across multiple work sessions.

Orchestration — Multi-Agent Coordination

Claude-Flow — Agent orchestration platform with a Queen + Workers architecture — manages Claude, GPT, and Gemini simultaneously via MCP.
OpenClaw — Evolution: Clawdbot → Moltbot → OpenClaw
Claude Squad — Terminal app for running multiple Claude Code and Codex sessions simultaneously, each in its own workspace.

AI Coding Agents — AI-Powered Coding Tools

Claude Code — The leading CLI agent — SWE-bench, outperforming all rivals.
Cursor — Smart IDE based on VS Code — features Composer and Agent mode, best for daily coding.
Codex CLI — OpenAI's CLI agent — runs in terminal or cloud, handles tasks sequentially and precisely.
Windsurf — AI IDE with Cascade — a multi-step autonomous agent: reads code → analyzes → edits → tests → commits.
GitHub Copilot — Microsoft's code assistant built into VS Code — features Agent mode, strong for enterprise.
Gemini CLI — Google's CLI agent — free for individuals, context (the largest).
Cline / RooCode — Open-source VS Code extension — supports multiple AI models and MCP, community-driven.
Kiro (AWS) — AWS's spec-first AI IDE — from requirements → design → plan → automated code generation.
Replit Agent — Cloud-based agent — builds full-stack apps fast, no setup needed. Best for beginners.
OpenCode — Open-source coding agent running in terminal — supports all LLM providers, MIT license
Goose — Open-source coding agent from Block (Square) — doesn't just suggest but also installs, runs, and debugs with any LLM

AI Trends 2026 — Trends Shaping the Future

Multi-Agent Systems — Gartner reports queries about multi-agent systems surged from Q1/2024 to Q2/2025.
Mechanistic Interpretability — Research into how AI 'thinks' internally — enabling better understanding and control of AI behavior. 2026 breakthrough per MIT.
World Models — AI simulating the real world — AMI Labs, Google, and startups are racing to develop it. Gaming is a major driver.
Generative Coding
Chinese Open-Source AI — DeepSeek R1 shook the AI world — open-source reasoning model from China. Technology gap shrinking from months to weeks.
Post-Training > Pre-Training — Trend toward post-training refinement instead of building larger models — RL and fine-tuning become the key.
AI Drug Discovery — AI-discovered drugs entering mid/late-stage clinical trials. Focus: cancer and rare diseases.
AI Infrastructure — Massive data centers, even using nuclear power for AI. A core battleground in the AI race.
AI Governance — AI agents monitoring other AI agents — setting safety guardrails and keeping humans in the loop.

Observability & Evaluation — AI Monitoring & Assessment

Langfuse — Open-source observability platform for LLMs — tracing, evals, prompt management
DeepEval — LLM evaluation framework like Pytest — 50+ metrics, CI/CD integration, red-teaming
Promptfoo — CLI tool for prompt testing, model comparison, and LLM red-teaming/security scanning
LangSmith — LangChain's observability platform — tracing, evaluation, prompt hub for LangChain/LangGraph
Arize Phoenix — Open-source AI observability platform — tracing, evaluation, RAG debugging, embeddings analysis

AI Infrastructure — Model Serving & Deployment

Ollama — The easiest way to run LLMs locally — supports 100+ models, 3 commands to get started
vLLM — High-performance production inference engine — PagedAttention, 2-4× throughput over baseline
LocalAI — Self-hosted OpenAI alternative — run LLMs, image gen, audio on consumer hardware, no GPU required
LiteLLM — Unified AI Gateway for 100+ LLM providers — cost tracking, load balancing, unified API

RAG Systems — Retrieval-Augmented Generation

Dify — Open-source platform for building AI apps with visual workflow builder and built-in RAG pipeline
RAGFlow — Open-source RAG engine specializing in deep document understanding — smart chunking, grounded citations
Flowise — No-code tool for building RAG pipelines and LLM apps with drag-and-drop UI — built on LangChain
LightRAG — Lightweight RAG framework based on knowledge graphs — GraphRAG approach from HKU research (EMNLP 2025)

Vector Databases — Vector Storage & Search

Milvus — Largest open-source vector database — cloud-native, scales to billions of vectors, backed by LF AI Foundation
Qdrant — Vector database written in Rust — fast, memory-efficient, best-in-class payload filtering
Chroma — Simplest vector database for AI apps — embedded mode, zero-config, ideal for getting started with RAG
pgvector — Extension that turns PostgreSQL into a vector database — no separate DB needed, uses familiar SQL

Security & Guardrails – Protecting AI Applications

garak — LLM vulnerability scanner – probes security flaws in AI models
NeMo Guardrails — Programmable guardrails for LLMs with Colang DSL – dialog control and safety
Presidio — PII detection & anonymization for text and images – protecting personal data
Guardrails AI — I/O validation framework for LLMs – structured output + 100+ validators