Briefing
GitHub hosts thousands of AI repositories, and most of them will never matter to your work. The hard part isn't finding projects to look at. It's working out which handful are worth your attention. So we picked ten that we think every AI developer should have starred, forked, or at least bookmarked, plus a few honourable mentions worth a look.
A quick word on the numbers below: open-source projects move fast, and star counts climb every week. The figures here were accurate when the article was first compiled, but several have already crept higher. Treat them as a rough sense of scale, not a live scoreboard. Each entry links straight to the repo so you can check the current count yourself.
Why bother with a list like this at all? Because the tools a team chooses early on tend to shape how it builds for years. Pick the project with a real community and active maintainers, and you inherit thousands of other people's bug fixes and ideas. Pick the abandoned one with a slick README, and you end up maintaining it yourself. The ten below are, for the most part, in the first camp.
1. OpenClaw (345,000 stars), github.com/openclaw/openclaw
One of the most-starred AI projects on GitHub, sitting somewhere north of 345,000 stars. MIT license, Node.js, more than 100 built-in skills, and the ClawHub marketplace for adding more. Worth noting that the project's own README pitches it as a personal AI assistant you run on your own devices and reach through messaging apps like WhatsApp, Telegram, Slack and Discord, rather than a developer agent framework as such. Either way, the skills it ships for browser automation, code execution and API orchestration are solid reference implementations, and if you build with agents you should know what's in here.
2. Langflow (146,000 stars), github.com/langflow-ai/langflow
A visual agent builder with a big following at roughly 146,000 stars. It's MIT licensed (not Apache 2.0, as you'll sometimes see written). You drag and drop your way through a workflow using more than 100 components, then export the whole thing as code when you're ready. Good for prototyping fast and for poking at ideas without writing much, and it's seen real enterprise pickup.
3. Dify (136,000 stars), github.com/langgenius/dify
A full platform for building LLM apps, with around 136,000 stars. You get a RAG pipeline, prompt management, visual orchestration and one-click deployment out of the box. If you're shipping production LLM applications, Dify hands you a chunk of infrastructure you'd otherwise have to build and maintain yourself.
4. Firecrawl (130,000+ stars), github.com/mendableai/firecrawl
The web-context API that a lot of agents use to read the internet, with 130,000+ stars (there's a hosted version too). It takes any website and hands back clean Markdown an LLM can actually use, and the MCP server integration makes it easy to wire into an agent that needs to browse. The crawler is AGPL-3.0; the SDKs are MIT.
5. Browser-use (86,000 stars), github.com/browser-use/browser-use
Natural-language browser automation for agents, built on Playwright, with around 86,000 stars. You tell it what you want in plain English and it drives the browser, with some visual understanding of the page. It's one of the cleaner ways to give an agent real web-browsing ability, which makes it handy for research, scraping and anything that involves clicking around a site.
6. AutoGen, github.com/microsoft/autogen
Microsoft's multi-agent orchestration framework: conversational agents that can run code, with human-in-the-loop support and flexible conversation patterns, plus deep Azure integration. One caveat worth knowing before you commit: the project is now in maintenance mode, and Microsoft is pointing new users toward its newer Microsoft Agent Framework. Still a useful codebase to learn from, but check where active development has moved before you build on it.
7. Mem0 (52,000 stars), github.com/mem0ai/mem0
Memory for agents, with about 52,000 stars. It offers layered storage across short-term, long-term and episodic memory, works with any model, and aims to give agents something better than a blank slate every session. The project advertises very fast retrieval; its own published benchmarks put single-pass retrieval closer to a second than the millisecond figures sometimes quoted, so test it against your own latency budget rather than taking the marketing number at face value. If your agents keep forgetting what happened five minutes ago, this is the kind of thing that fixes it.
8. nanochat (55,000 stars), github.com/karpathy/nanochat
Andrej Karpathy's minimal LLM training stack, sitting at roughly 55,000 stars. The headline: you can train a GPT-2 class model for about $48 in compute (the README clocks it at roughly two hours on an 8xH100 node). More than that, it's one of the best ways to actually understand how transformers work. The code is small and well commented, and reading it teaches you something.
9. LocalAI (44,000 stars), github.com/mudler/LocalAI
Run models on your own hardware, no GPU required, with around 44,000 stars. It's MIT licensed and exposes an OpenAI-compatible API, so you can point existing code at a local backend without rewriting it. It handles LLMs, vision models, embeddings, diffusion and audio. If you're moving work off the cloud for cost or privacy reasons, this is a sensible foundation.
10. Hermes Agent (22,000 stars), github.com/nousresearch/hermes
Nous Research's learning agent, built around what they call dialectic memory via the Honcho system, with 40+ tools. One correction before you go looking: the repo lives at NousResearch/hermes-agent, and its star count is far higher than the 22,000 originally listed here. The pitch is an agent that actually adapts to the person using it over time, which puts it closer to the research edge than the production mainstream. Reportedly built by around 142 contributors, though we couldn't confirm that figure.
Honourable Mentions
These didn't make the top 10 but are worth a look:
- CrewAI: probably the most approachable multi-agent framework if you're starting out
- MetaGPT: multi-agent teams that take a brief and build software end to end
- OpenHuman: a desktop-first personal AI (tinyhumansai/openhuman) with 118+ integrations and persistent local memory
- Vercel agent-browser (27,000 stars): serverless browser automation, at vercel-labs/agent-browser
- awesome-claude-skills: the broader community has put together collections of 1,000+ skills for Claude Code, though any single repo of that exact name tends to be smaller
How to Use This List
Star the ones that look relevant, read their docs, and run a quickstart or two. Even the projects you never adopt are worth opening, because the design choices their maintainers made will rub off on how you build. And since this space turns over fast, it's worth coming back every few months to see what's changed.
The thread running through all ten is that they're genuinely open and genuinely maintained. None of them are README-and-nothing-else projects. They're worked on in the open, depended on by a lot of people, and that's most of the reason they earned a place here.