A curated index of the tools developers actually use to build with AI. No models — just the tooling around them: frameworks, vector databases, inference, observability, agents and more.
Frameworks & SDKs
The de facto framework for building LLM applications.
LangChain gives you composable primitives — prompt templates, chains, agents, retrievers and hundreds of integrations — to build anything from a chatbot to a multi-step agent pipeline, in Python or JavaScript. It shines when many components need to talk to each other.
Frameworks & SDKs
The data framework for connecting your data to LLMs.
LlamaIndex specializes in ingestion and retrieval: it reads documents, chunks and indexes them, and gives you query engines with citations out of the box. If you are building RAG, it is the fastest path from raw files to a working question-answering system.
Frameworks & SDKs
The TypeScript toolkit for shipping AI-powered products.
The AI SDK provides one streaming interface to every major provider — OpenAI, Anthropic, Google, Mistral, Ollama and more — plus React hooks like useChat that make streaming UIs trivial. It has become the standard choice for AI features in web apps.
Frameworks & SDKs
Structured, validated outputs from any LLM.
Instructor patches popular model clients so calls return typed objects — Pydantic models or Zod schemas — instead of raw JSON strings, with automatic retries when validation fails. It removes most of the pain of getting reliable structured output.
Inference & Routing
Run open models on your own machine with one command.
Ollama bundles model weights, a runtime and an OpenAI-compatible REST API behind a single CLI. It is the easiest way to develop against local models — no API keys, no rate limits, full privacy — which makes it a favorite for local development and testing.
Inference & Routing
High-throughput inference server for production.
vLLM is the go-to engine for serving open models at scale. PagedAttention and continuous batching deliver many times the throughput of naive serving, behind a drop-in OpenAI-compatible API. Used across the industry for self-hosted inference.
Inference & Routing
One API client and proxy for 100+ providers.
LiteLLM translates the OpenAI completion format to every major provider, so switching models is a one-line change instead of a rewrite. Its proxy server adds centralized API keys, budgets, rate limits and spend tracking for your whole team.
Inference & Routing
A single account and API for every model.
OpenRouter routes requests to OpenAI, Anthropic, Google, Meta, Mistral and hundreds of other models behind one OpenAI-compatible API, with unified pay-as-you-go billing and automatic fallback between providers when one is down.
Models & Datasets
The GitHub of models, datasets and demos.
The Hub hosts over a million open models and datasets plus Spaces for hosting live demos. The transformers and datasets libraries pull straight from it, making it the backbone of the open-source ML ecosystem and the first place to look for any open model.
Vector Databases
Fully managed vector database for production RAG.
Pinecone handles the operational side of vector search — sharding, replication, backups, updates — so you can focus on retrieval quality. Its serverless indexes scale to billions of vectors with millisecond queries and zero ops.
Vector Databases
Open-source vector database built in Rust.
Qdrant delivers fast similarity search with rich payload filtering — the exact combination production RAG pipelines usually need. Self-host it with Docker or use the managed cloud, with clients for Python, JS/TS, Rust and Go.
Coding Tools
The AI-first code editor.
Cursor is a VS Code fork rebuilt around AI: it indexes your whole codebase, edits across multiple files at once, and runs agentic tasks that can read code, run commands and fix their own errors. The most polished AI coding experience available today.
Coding Tools
AI pair programming in your terminal.
Aider edits your git-tracked files directly from the CLI, writes clean commits, and works with any model — from Claude and GPT to local Ollama models. Perfect for developers who live in the terminal and want AI without leaving it.
Coding Tools
Open-source AI code assistant for VS Code and JetBrains.
Continue brings tab-autocomplete, chat and edit modes into your IDE with total control: choose any provider and any model — including fully local ones via Ollama — and decide exactly which files and context your assistant sees.
Observability & Evals
Open-source LLM observability and prompt management.
Langfuse records every trace of your LLM application — prompts, latencies, costs, scores — so you can debug and improve what you shipped. It also versions prompts and runs dataset-based evaluations, covering the loop from development to production.
Observability & Evals
Experiment tracking and LLM tracing for ML teams.
W&B is the industry standard for tracking ML experiments, and its Weave module extends that to LLM apps with tracing, prompt experiments and evals. One dashboard for datasets, runs, metrics, prompts and costs.
Observability & Evals
Test and evaluate prompts like code.
promptfoo runs your prompts across a matrix of models and test cases with assertions on the output — then fails your CI when something regresses. The quickest way to make prompt and model changes safe instead of a leap of faith.
Data & Agents
Crawl any website into clean, LLM-ready data.
Firecrawl scrapes single pages or entire sites, handles JavaScript rendering, and returns markdown or structured data that is ideal for RAG ingestion — no headless-browser babysitting. It turns the messy web into model-ready input.
Data & Agents
Web search API designed for AI agents.
Tavily returns concise, relevant, LLM-friendly search results instead of raw listings, saving tokens and making agents more reliable. It has become the default search tool in many agent frameworks.
Data & Agents
Validate and repair LLM outputs in production.
Guardrails wraps model calls with typed validators — structure, tone, safety, custom Python checks — and automatically re-asks or rephrases when a constraint fails. It keeps LLM features from shipping embarrassing outputs.