Last week a fresh batch of open-source AI projects hit GitHub, and one of them racked up 27,000 stars in about seven days. That kind of speed usually means something real is happening, not just hype. So when Matthew Berman rounded up six of these projects in one video, I stopped what I was doing and watched the whole thing. The creator did the digging, tested each tool, and lined them up so you can grab whatever fits your workflow. I pulled out the good stuff so you can skim it fast.
Here’s what makes this roundup interesting: most of these tools run completely local. No cloud bill, no data leaving your machine, no vendor lock-in. And a few of them are built “agent-native,” meaning your AI agent is treated as a real user, not a bolted-on afterthought. That’s the shift worth paying attention to.
What’s new: the six projects
The original poster walked through each one. Here’s the quick map.
- Unsloth 🛠️ Started as a way to make local LLM fine-tuning easier for regular people. Now it’s a full powerhouse. You can run and train the latest open models (Kimi K2, MiniMax, Qwen, Gemma, DeepSeek and more), and it ships a full agent UI that looks like ChatGPT. Web search, tools, MCP, memory, remote access, image and video models. All local, all free, Windows, Mac and Linux.
- Diagram Design Gives your agent the power to make diagrams that actually look clean. No broken arrows or overlapping boxes. Flowcharts, architecture charts, state machines, timelines, quadrants. It installs as a plugin or skill in Claude Code, Codex, Hermes and more.
- Obsidian Skills Connects your agent to your Obsidian notes. Since Obsidian runs on markdown and agents love markdown, they work together cleanly. People use it as a wiki, a knowledge base, basically the brain for their agent. Local-first, syncable across devices.
- Buzz An open-source Slack alternative from Block, the company behind Jack Dorsey’s Square. Built AI-first, so your agents are first-class users working in the same channels as your human team. Self-hostable, privacy-focused, with threads, channels, emojis, plus deep workflow automations baked in.
- Ego Lite Calls itself the fastest browser for AI agents. It shares your logged-in browser state with agents like Codex or Cloud Code without disturbing you. Zero cost, zero config. You invoke it with a simple “/browser” command.
- Modly Local, open-source, image-to-3D mesh generation. Feed it any image and it spits out a 3D mesh for printing, games or asset work. Runs on a normal desktop GPU, not compute-heavy, and supports Windows, Linux and Mac.
The twist
The part that surprised me most was Buzz. The mind behind it didn’t just clone Slack. Every message, reaction, workflow step, review and approval gets logged as a signed event in one shared log. Same identity model, same audit trail, whether a human or an agent did it. So there’s no separate “bot” lane. A person and a process look identical in the record. For teams that want agents doing real work alongside people, that’s a genuinely different foundation.
Unsloth is the other quiet twist. The creator pointed out how intimidating fine-tuning usually is. So many settings, so many ways to break things. Unsloth turns it into a point-and-click flow that walks you through it. No code required. That lowers the barrier for anyone curious about training their own model.
A quick way to try one
Want to test Diagram Design without leaving your machine on all day? The original poster showed a clean path using Hermes hosted in the cloud. Here’s the gist. 🚀
- Host your agent in the cloud so it runs 24/7 without your computer staying on.
- Tell the agent “install this” and paste the GitHub URL for the tool.
- Wait about a minute while it installs as a skill.
- Ask it to build something, like “Create a diagram for my agent to show the architecture.”
- Watch it pick the right skill automatically and output a clean diagram.
In the video, that whole install took roughly one minute and the diagram came out looking sharp. The nice thing about hosting in the cloud is you can run background processes around the clock and step away from your desk.
Pro tips
- Start local, then sync. Tools like Obsidian Skills and Unsloth are local-first. Get comfortable on one machine, then turn on sync across your devices when you’re ready.
- Match the tool to the hardware. Modly and Unsloth both run on a normal desktop GPU. You don’t need a server farm to start experimenting with 3D meshes or model training.
- Check the star count as a signal, not a verdict. Buzz at 27,000 stars is buzzy, but Ego Lite at 10,000 and Modly near 6,000 are more under the radar. Sometimes the smaller ones fit your exact need better.
- Use agent-native tools where agents do real work. If you’re running Codex, Claude Code or Hermes daily, the plugin-style installs (Diagram Design, Obsidian Skills, Ego Lite) slot right in with a single command.
One small note. This roundup came with a sponsor plug for cloud hosting, so treat that part as an ad, not a neutral pick. The six projects themselves are all open source and free to grab from GitHub.
The full video shows each tool running live, including the Unsloth training interface and the Buzz workspace, which are worth seeing in motion. Go watch it, then pick one project and install it this weekend. 👇