Repo Radar

An AI assistant that keeps your data in your own claws

OpenClaw is a TypeScript-built LLM assistant designed around one hard rule: your data stays yours, not the vendor's. It handles the usual assistant work — chat, task help, tool use — but stores and syncs everything locally first, so you're never locked into someone else's cloud. Worth a look if you want an AI helper without handing over the keys to your own information.

Desktop AI chat that runs on anything

GPT4All was among the first projects to make local models genuinely easy: download an app, pick a model, chat — no GPU required. Still one of the gentlest introductions to private AI.

The CLI behind LM Studio's local model magic

LM Studio is the friendliest desktop app for running models locally; lms is its command-line companion for loading, serving, and scripting those models. Great bridge from 'clicking around' to automating your local AI.

Visual maps for every developer career path

Interactive roadmaps showing exactly what to learn — and in what order — for frontend, backend, DevOps, AI engineering, and more. One of the most-starred repos on GitHub because it answers the eternal question: 'what next?'

Spin up thousands of AI agents, watch society emerge

OASIS simulates large-scale social platforms where thousands to a million LLM-driven agents post, follow, and react to each other, letting researchers study how trends, rumors, and group behavior actually spread. It's a sandbox for the messy dynamics of social networks — useful for anyone studying misinformation, group polarization, or emergent agent behavior without touching real users.

Anthropic's agentic coding tool, in your terminal

Claude Code puts a frontier model directly in your terminal where it can read your codebase, run commands, and ship real changes. The docs double as a masterclass in how to prompt an agent well — worth reading even if you use something else.

Pair programming with AI in your terminal

Aider is the beloved open-source AI coding companion: it maps your git repo, makes surgical edits, and commits with clean messages. Its benchmark leaderboards are also required reading for anyone comparing coding models.

Agents that argue until the citations check out

AutoResearchClaw runs a team of LLM agents that debate a research question, cross-check each other's claims, and refuse to cite anything they can't verify. It's a solid answer to the sloppiest part of AI research tools: made-up sources. 14k+ stars and a live, actively maintained multi-agent architecture worth reading even if you never run it.

Skills that make Claude actually think like you

Superpowers is a skill library that gives Claude structured playbooks for brainstorming, planning, and shipping code, instead of winging it every session. Built by obra, it's become one of the most-starred ways to turn a generic coding agent into one with real SDLC habits.

An open-source AI software engineer

OpenHands (formerly OpenDevin) gives an AI a terminal, browser, and code editor, then lets it fix issues and ship pull requests like a junior dev. Watching it work is a preview of how software gets built next.

Serve AI models at production speed

vLLM is what companies use when they need to serve language models to thousands of users without melting their GPUs. Its PagedAttention trick massively increased throughput and became an industry standard overnight.

Drive Chrome like a puppet, script the whole web

Puppeteer is Chrome's own team giving you a remote control for the browser: a Node/TypeScript API that opens pages, clicks, fills forms, and grabs screenshots or PDFs without a human at the keyboard. It's the backbone of countless scraping, testing, and automation pipelines, and with 95k+ stars it's the default answer whenever 'just script the browser' comes up.

One-click deploys, minus the vendor lock-in

Dokploy is a self-hosted alternative to Heroku or Vercel — spin up apps, databases, and automated backups on your own server with a clean UI instead of a wall of YAML. With 36.5k stars, it's become the default answer for teams who want PaaS convenience without handing infrastructure to a third party.

Agentic AI, curated in Chinese

A sprawling, well-organized index of agentic AI frameworks, papers, and tools, written for the Chinese-speaking dev community. With 7.4k stars it's become a go-to map for anyone navigating the fast-moving agent ecosystem without relying on English-only sources.

Microsoft's free 21-lesson generative AI course

A structured, genuinely beginner-friendly path through generative AI: prompting, RAG, agents, fine-tuning — with videos and code in multiple languages. The syllabus your 'how do I start with AI?' friend is looking for.

A local multi-agent that browses, codes, and thinks for itself

agenticSeek is a fully autonomous agent framework built to run on your own machine with local LLMs like DeepSeek R1 — no cloud API required. It routes tasks across specialized agents for web browsing, coding, and file ops, so you get agentic behavior without shipping your data out. With 27k+ stars, it's become a go-to reference for anyone building privacy-first autonomous agents in Python.

The API client that keeps your data on your machine

Bruno is the open-source Postman alternative that stores collections as plain files in your git repo — no cloud, no accounts, no sync fees. A model example of local-first developer tooling done right.

The virtual whiteboard everything sketches on

Excalidraw is the hand-drawn-style diagramming tool that became the internet's shared whiteboard — open-source, end-to-end encrypted collaboration, zero accounts. Perfect for sketching how your agents talk to each other.

The free textbook of prompt engineering

DAIR.AI's guide covers every major prompting technique — few-shot, chain-of-thought, self-consistency — with papers and examples to back each one. If you want the theory behind why good prompts work, it's here, free, in ten languages.

Ask your codebase a question, get a real answer

Deepwiki-open points an AI at any GitHub repo and generates a readable wiki out of it — architecture notes, diagrams, and a chat interface for asking how the code actually works. With 18k+ stars and support for Ollama, Gemini, and Grok alongside the usual suspects, it's become a go-to for onboarding onto unfamiliar codebases fast.

Build AI apps on a visual canvas

Dify is an open-source platform where you assemble chatbots, agents, and workflows by dragging blocks instead of writing glue code — then ship them with monitoring included. A favorite for teams who want AI apps without an ML team.

A command-line arsenal of battle-tested prompts

Fabric collects hundreds of curated 'patterns' — extract wisdom, summarize, analyze a paper — and pipes them through any AI from your terminal. It treats prompts as reusable tools, which is exactly the right way to think about them.

Run powerful AI models on your own machine with one command

Ollama made local AI boring in the best way: 'ollama run llama3' and you're chatting with a model that never leaves your laptop. It's the de-facto standard for private, offline, zero-subscription AI. If you try one local-AI tool, try this.

Chat with your documents, privately

AnythingLLM turns your PDFs, notes, and websites into a private AI knowledge base you can chat with — local models welcome. One of the smoothest 'RAG in a box' experiences going.

The research agent that fixes real GitHub issues

From Princeton researchers: SWE-agent takes an actual GitHub issue and autonomously produces the fix, and its papers taught the field that the *interface* you give an AI matters as much as the model. Where a lot of agent best practices were born.

The repo that started the autonomous-agent craze

AutoGPT was the first project to ask 'what if GPT could give itself its own to-do list?' — and the internet lost its mind. Today it's a full platform for building agents that plan, act, and self-correct. Historically important, still evolving fast.

How the big AI tools really prompt their models

The extracted system prompts of Cursor, v0, Devin, and other famous AI products. Reading how professional teams structure instructions, tools, and guardrails is the fastest masterclass in serious prompt architecture.

Assemble a crew of role-playing AI agents

CrewAI's big idea is casting: give each agent a role, a goal, and a backstory, then let the crew collaborate like coworkers. It made multi-agent workflows feel like writing a job description instead of code. One of the fastest-growing agent frameworks.

Leaked system prompts from thousands of custom GPTs

A giant collection of the hidden instructions behind popular custom GPTs. Ethically gray at the edges, but as prompt education it's unmatched — you learn more from reading fifty real system prompts than from any tutorial.

Your codebase, drawn as an architecture diagram

Archify is an agent skill for Claude and Codex that turns a codebase into an architecture diagram, so you can see how the pieces fit without drawing boxes by hand. It follows the architecture-as-code idea, where a diagram is generated instead of drawn, which makes it easier to redo when the code changes. With nearly 69,000 stars, a lot of developers clearly want a faster way to explain their own systems.

One agent, every chat app, zero glue code

nanobot is a lightweight framework for building AI agents that plug straight into Discord, chatops, and other chat surfaces without you wiring up the boilerplate yourself. It handles the agent-framework plumbing so you can focus on behavior, not transport code. With 46k+ stars, it's become a default starting point for anyone shipping a chatbot that needs to feel less like a bot and more like a teammate.

Turn any codebase into a tutorial someone actually reads

This tool points an LLM at your repo and has it write a beginner-friendly tutorial explaining how the code actually works — abstractions, relationships, the works. With 12.6k stars, it's clearly filling a real gap: onboarding docs nobody has time to write by hand.

Claude Code prompts, cut with a stone knife

caveman is a prompt library for Claude Code that strips instructions down to blunt, unambiguous commands — no fluff, no hedging, just what the model needs to do the job. It's caught on because terse prompts turn out to work better than polite ones, and 95k+ stars say a lot of people needed that reminder.

The engine that made local AI possible

This is the legendary C++ project that squeezed billion-parameter models onto laptops and phones, spawning the entire local-AI ecosystem — Ollama included. Deeply technical, endlessly optimized, and quietly running inside half the AI apps you've heard of.

The legendary Stable Diffusion interface

The scrappy web UI that put image generation in everyone's hands and accumulated a decade of features in two years. ComfyUI has the momentum now, but A1111 remains the most approachable on-ramp to local image AI.

The open-source Firebase that AI apps love

Postgres, auth, storage, and realtime APIs in one open-source package. Supabase became the default backend of the AI-app boom for a simple reason: it turns weekend projects into products. (This very site runs on it.)

Microsoft's framework for AI agents that talk to each other

AutoGen lets you build teams of AI agents that debate, delegate, and check each other's work before answering you. It pioneered the multi-agent conversation pattern most frameworks now copy. Great starting point if one AI isn't enough.

Build your own AI code assistant

Continue is the open-source alternative to Copilot that lets you wire ANY model — local or cloud — into VS Code and JetBrains, with custom rules, docs, and prompts. The escape hatch from assistant lock-in.

Stop guessing, start A/B testing your prompts

Prompt Optimizer is a free, open-source toolkit that iteratively rewrites and scores your prompts side-by-side so you can see what actually works before you ship it. With 37.7k stars and support for most major LLMs, it's become the go-to way to turn a vague prompt into a sharp one without the usual trial-and-error slog.

Agent skills and Cursor rules, tied together

Ponytail is a dev-tools repo for wiring up Claude Code agent skills and Cursor rules so both live in one shared setup instead of two separate configs. It's the kind of plumbing that saves you from maintaining duplicate prompt files across editors and agents. Handy if you're juggling Claude Code plugins alongside Cursor in the same project.

Multi-agent muscle for Claude Code

ECC is a developer toolkit for orchestrating multiple Claude Code agents so they can tackle bigger coding tasks together instead of one at a time. It's built for anyone wiring up AI-driven dev workflows on top of Anthropic's tooling and wants agents that coordinate, not just chat.

Anthropic's own hands-on prompting course

A nine-chapter interactive tutorial from the people who make Claude, teaching prompt construction with exercises and answer keys. Straight-from-the-source technique — most of it transfers to any model.

Your second brain, self-hosted and yours

Khoj is an open-source AI assistant that reads your notes, files, and even your Emacs buffers, then chats with you about them using whatever model you point it at. It's for people who want an assistant wired into their own knowledge base instead of a search box that forgets everything by tomorrow.

The Swiss-army framework that defined LLM apps

LangChain coined half the vocabulary of modern AI development — chains, retrievers, memory — and connects language models to basically everything. Sprawling and opinionated, but its cookbook of patterns is a genuine education.

The editor the world writes code in

VS Code is the most-used code editor on earth and one of the most successful open-source projects ever. Every AI coding tool you've heard of is either an extension of it or a fork of it — understanding it is understanding the terrain.

Claude Code's context, finally on a diet

Headroom is a compression layer for Claude Code agent sessions, trimming bloated context so long-running coding agents stay fast and cheap instead of choking on their own history. With 64k+ stars, it's become the default fix for anyone running serious agent workloads on Anthropic's stack.

Fine-tune big models on a gaming GPU

Unsloth is an open-source toolkit that makes training and fine-tuning language models like Gemma and DeepSeek faster and much lighter on memory, so you can do serious work on hardware you already own. You don't need a rented cluster to teach a model your own data, and that puts custom local AI within reach of solo builders and small teams. It's a big reason fine-tuning stopped being something only labs with deep pockets could do.

Real system prompts, ripped straight from the source

A growing archive of leaked and extracted system prompts from major AI assistants like Claude, ChatGPT, and more — the instructions vendors don't publish. Handy for anyone reverse-engineering behavior, comparing prompt design across products, or just curious what's really steering these bots.

Turns your writing back into you

humanizer is a Claude Code / Cursor skill that strips the AI cadence out of AI-drafted text, so it reads like someone actually wrote it. With 44.8k stars it's become the default fix for the em-dash, three-part-list, 'it's not just X, it's Y' tell that gives AI writing away. If you ship words as well as code, this belongs in your prompt library.

Enterprise search that actually finds the answer

Onyx is an open-source AI platform that connects chat to your company's real knowledge — docs, wikis, tickets, code — instead of the open web. With 31.5k stars, it's become a go-to for teams who want ChatGPT-style answers grounded in their own data, not hallucinated guesses.

Turn your old devices into an AI cluster

Exo lets you run big models across whatever hardware you have lying around — MacBooks, phones, Raspberry Pis — automatically splitting the work between them. Frontier-lab ambition, garage-sale hardware.

The most-starred repo on GitHub — free coding education

freeCodeCamp has taught millions to code with free certifications in web development, Python, and machine learning. It tops GitHub's star charts because it might be the highest-impact educational project on the internet.

Turns Claude into a research co-author

A Python toolkit that wires Claude Code into the academic writing pipeline — literature review, citation handling, and drafting, stitched into one workflow instead of ten browser tabs. With 46k+ stars it's become a default starting point for researchers automating the parts of the process that eat time without needing a PhD in prompt engineering.

One man's public quest to become a Google engineer

John Washam's complete self-study computer science curriculum, built while grinding toward a big-tech job and shared with the world. Equal parts study plan and motivation — the repo that proves the path is walkable.

A gorgeous ChatGPT-style interface for your local models

Open WebUI gives your self-hosted models the polished chat experience they deserve — conversations, documents, voice, image generation, multi-user — all offline-capable. Pair it with Ollama and you've replaced a subscription.

An autonomous coding agent inside VS Code

Cline plans, edits files, runs commands, and uses your browser — with a human-approval gate on every step. It's one of the best showcases of what agentic coding feels like when you stay in the loop.

Agents as flowcharts you can actually debug

LangGraph models an AI agent as a graph — nodes think, edges decide where to go next, and you can checkpoint or rewind any step. It's what production teams reach for when 'a loop around an LLM' stops being reliable enough.

Free GPT/Claude/Gemini API keys, no card

A pooled proxy that hands you working API keys for GPT, Claude, Gemini, and DeepSeek without signing up or paying anyone. It's the shortcut for prototyping across models when you don't want five separate billing accounts just to test an idea.

Copilot customizations, crowdsourced and curated

A massive, community-driven stash of custom instructions, prompts, and agent skills for GitHub Copilot — nearly 40k stars of people sharing what actually works instead of guessing. If you're trying to get real mileage out of Copilot agents, this is the cheat sheet.

Let your AI drive a real web browser

Browser-use gives language models eyes and hands on the web: it reads pages, clicks buttons, fills forms, and completes tasks on real sites. It's the missing piece between 'AI that talks about the web' and 'AI that uses it.'

Learn by rebuilding the tools you use

A master list of tutorials for building your own Redis, Git, Docker, database, or programming language from scratch. The single best repository for turning 'I use it' into 'I understand it.'

The self-taught LLM engineer's curriculum

A complete course for going from developer to LLM engineer: fundamentals, fine-tuning, quantization, deployment — with notebooks for everything. Regularly updated, ruthlessly practical, entirely free.

Claude Code skills, versioned and testable

autoharness is a Python toolkit for building and running Claude Code agent skills as a proper plugin — define them, harness them, test them like real software. With ~7,930 stars it's become a go-to reference for anyone wiring LLM agents into real workflows instead of one-off scripts.

PewDiePie's self-hosted AI workspace — yes, really

Felix built a full AI workspace — chat, agents, deep research, email, notes — that runs entirely on your own machine, then open-sourced it. 77k stars in three weeks. The 'own your AI instead of renting it' philosophy, shipped as working code.

One API for every AI model

LiteLLM lets you call 100+ models — OpenAI, Anthropic, Google, local Ollama — through one consistent interface, with budgets, fallbacks, and spend tracking built in. The plumbing every multi-model app ends up needing.

The framework behind half the modern web

Next.js is the React framework that powers a staggering share of new websites — including this one. Server components, instant deploys, and enough momentum that AI tools generate Next.js code by default.

Take your photos back from the cloud

PhotoPrism is a self-hosted photo app that runs your own AI-powered search and organization over your library, no subscription or third-party server required. It recognizes faces, objects, and places locally, so you get Google Photos-style search while keeping every image on hardware you control.

Andrej Karpathy's tiny, readable GPT

The simplest serious GPT training code in existence — a few hundred lines you can actually read, from one of the field's great teachers. Pair it with his YouTube lectures and you'll understand transformers better than most engineers.

The original prompt collection

The repo that proved prompts were worth collecting: hundreds of 'Act as a…' personas that defined early prompt sharing. A time capsule and a starting point — the craft has moved on (that's where we come in), but this is where it began.

Build a language model from first principles

Sebastian Raschka's book-as-repo walks you through implementing a GPT-style model line by line in PyTorch — pretraining, fine-tuning, everything. The deepest 'no magic' understanding of LLMs you can get.

Node-based image generation for power users

ComfyUI turns AI image generation into a visual circuit board: wire nodes together to control every step of the pipeline. Steeper than a prompt box, but it's how the most impressive AI art workflows actually get built.

Workflows that actually talk to your other tools

n8n is a self-hostable automation platform where you wire up APIs, data, and AI models on a visual canvas instead of gluing together scripts. It's the local-AI crowd's favorite because you can run it on your own hardware, drop in an LLM node next to a database call, and own the whole pipeline. Nearly 200k stars and counting from people tired of paying per-task for Zapier.

An AI agent from the team behind the Hermes models

Hermes Agent is an open-source Python framework from Nous Research, the lab known for its Hermes family of language models. It lets you set an AI model up as an agent that works on tasks for you, and it isn't tied to one provider, so it works with Claude, ChatGPT and others. With around 247,000 stars, it's among the most-watched agent projects on GitHub, and it's a good place to see how people are building agents in the open.