Build an AI Agent / RAG App

A retrieval-augmented assistant that ingests your docs, retrieves over a vector store, and runs a model with tool use. Local-first if you want privacy, or cloud LLMs if you want speed.

Repos

Layers

Build time

About 2 weeks

Outcome

See below

You will ship

An agent that answers questions over your data with citations, runs locally or in the cloud.

How to feed this to your AI agent Browse the full directory

Vector storage

4 repos

bypgvector

pgvector

Postgres extension. If you already have Postgres, do not add another DB.

Vector search inside Postgres. ~12k stars and dominant in 2026 as the boring-but-correct default. If you already have Postgres (Supabase, Ne…

bychroma-core

chroma

Embedded or server. The easiest first vector DB before scale.

AI-native embedding database designed for rapid prototyping. Easy to run locally, tight LangChain integration. Better for hackathons and int…

byweaviate

weaviate

When you need hybrid search and module-rich pipelines.

Vector DB with strong hybrid search (vector + keyword) and built-in modules for common embedding workflows. Slightly higher learning curve t…

bymilvus-io

milvus

Battle-tested at billions of vectors. Overkill for v1.

Open-source vector DB with the most mature features for huge scale. ~30k stars. Steeper operational burden than Qdrant. Used in production b…

Local LLM

3 repos

byollama

ollama

Run Llama 3, Qwen, Mistral on your machine. One curl install, full local privacy.

Run open-source LLMs locally with one command. ~95k stars. Essential for AI development — test prompts against Llama, Mistral, Phi without b…

byvllm-project

vllm

Production LLM serving — 10-24× faster than HF Transformers for batched inference. Used at scale by Mistral, Together AI.

Production LLM serving. 10-24× faster than naive Hugging Face Transformers for batched inference. Used by Mistral, Together AI, and most ent…

byopen-webui

open-webui

ChatGPT-clone UI for your Ollama. Multi-user, RAG, plugins. The default frontend for local LLMs.

Voice AI (optional)

2 repos

bypipecat-ai

pipecat

Open-source voice AI pipelines. Build phone agents that call APIs + speak responses.

Open-source voice AI orchestration. Build voice agents that call APIs, transcribe, generate, speak — all in pipelines. Pairs with LiveKit fo…

bylivekit

livekit

WebRTC infrastructure for real-time voice/video. Powers OpenAI Realtime API.

Open-source WebRTC infrastructure. Build voice/video apps without renting Twilio. Used by OpenAI for Realtime API. Self-host the SFU or use …

9 more layers · 23 more repos · members only

AI coding agents (developer use)2 repos
Framework + UI3 repos
Data layer1 repo
Agent frameworks8 repos
Embeddings + reranking2 repos
Observability (LLM-specific)2 repos
Frontend chat UI2 repos
Data ingestion2 repos
Validation + types1 repo

23 more curated repos · unlock full access · members only

Unlock with lifetime membership.

Pay once. Full directory unlocked forever. No renewals, no surprise charges.

See pricing

How to build build an ai agent / rag app with AI

The 4-step AI workflow

The AI agents are good at code. They're bad at deciding what stack to use. This bundle does the second part. You bring the agent.

1
Ideate with ChatGPT or Claude.ai (web)
Paste your idea: “I'm building build an ai agent / rag app. Help me sharpen the product spec — features, edge cases, MVP scope.” Iterate for 10-15 minutes until you have a clear one-page brief.
2
Pick your coding agent
For this kind of bundle, we recommend Claude Code — Sonnet 4.6/4.7 handles full-stack multi-file reasoning best. See the install guide → Cursor and Codex are also great; pick the one you already pay for.

Feed this bundle to the agent

Open Claude Code / Cursor / Codex in an empty folder, then paste:

I'm building build an ai agent / rag app. Use this bundle as the source of truth for the stack:
https://stackpicks.dev/build/ai-agent

Brief from my product spec:
[paste your brief from step 1]

Follow the bundle order strictly:
  1. Vector storage
  2. Local LLM
  3. Voice AI (optional)
  4. AI coding agents (developer use)
  ...

Stop and confirm with me after each layer.

4
Wire one layer at a time, commit between each
Don't let the agent install everything before the first git commit. One layer = one commit. Catches drift early, easy rollback.

Beyond the bundle

1Ship the boring version first. The bundle above is the maximalist list. For an MVP, start with 60% of these and add the rest when real users ask.
2Deploy early. Push to Railway / Vercel after layer 02 (auth) — not after layer 09. Production breaks differently than localhost.
3Read CLAUDE.md / .cursor/rules in this repo for the project conventions your AI agent should follow.
4Iterate on the take. If a repo here doesn't fit your specific use case, tell us — contact — and we'll add a better one within 60 minutes.