Agent Infrastructure and On-Device AI Dominate GitHub Trending
Hello everyone, I'm Ivan.
It's Friday again, and I'm here to share this week's trending open-source projects on GitHub. This week, the top projects are mostly related to AI and Agents, and several of them have grown very quickly. diagram-design gained over 11,000 stars in a week, and MoneyPrinterTurbo also added 9,712 stars.
However, GitHub Trending's ranking isn't sorted directly by new stars added this week, so this article follows the order currently displayed on the weekly trending page. As of August 21, 2026, the top eight projects are as follows.
1. diagram-design: Let Agents Generate Production-Ready Technical Diagrams
Repository: https://github.com/cathrynlavery/diagram-design
Ranked first this week is diagram-design, which gained 11,325 stars in a week and currently has over 24,000 stars.
This project is a set of drawing Skills for Claude Code, Codex, and Pi, offering 38 graphic styles, including architecture diagrams, flowcharts, sequence diagrams, state machines, ER diagrams, and Gantt charts. The output is HTML and SVG, which can be opened directly in a browser and easily embedded into articles and project documentation.
Getting an Agent to draw a diagram isn't hard now, but the results are often just a few boxes and arrows. While the content is understandable, the layout still needs manual adjustment. This Skill encodes rules for how different diagrams should be laid out, how much content to include, and what colors to use. It solves a very specific problem, which explains why it suddenly surged in popularity this week.
2. OpenViking: Unifying Agent Memory, Knowledge, and Skills
Repository: https://github.com/volcengine/OpenViking
The second project is OpenViking, open-sourced by Volcano Engine. It's a Context Database specifically designed for AI Agents, currently with over 31,000 stars.
When we build Agents, memory might be in one database, the RAG knowledge base in another, and Skills placed directly in the project directory. This is fine when you have little content to store, but once an Agent needs to run long-term, managing where this context is stored and how to find it becomes increasingly troublesome.
OpenViking's approach is to place memory, knowledge, and Skills into a unified virtual file system, managed via viking:// paths. After content is written, it generates three layers of information: L0, L1, and L2. Queries can first look at summaries and then read the full content when needed, avoiding stuffing all the content into the context upfront.
It also logs every directory that was searched during retrieval. If the Agent looks up the wrong information, you can see exactly where it went to search, rather than just getting a few similarity scores. I think this direction is quite close to Agent Infra, Memory, and Skill management, and we might discuss it further later on.
3. Omarchy: A Linux Development Environment by DHH
Repository: https://github.com/basecamp/omarchy
Ranked third is Omarchy, a Linux distribution created by DHH. It gained 2,395 stars this week and currently has over 27,000 stars.
The project's characteristic is its fairly complete configuration: the terminal, Neovim, development tools, browser, AI tools, keyboard shortcuts, and themes are all pre-configured. It retains many of DHH's personal choices, so you can use this entire development environment right after installation.
If you already want to use Linux for development but don't want to spend a lot of time configuring the desktop and various tools, you can check out this project. However, its configuration has a distinct personal style, and people accustomed to Windows or macOS might need some time to adapt initially.
4. Needle 2: An On-Device Model That's Only 14MB
Repository: https://github.com/cactus-compute/needle
Needle 2 is a small model that can run on phones, wearables, smart home devices, and robots. It has only 45 million parameters, and after compression, it's a 14MB binary file. A full conversation requires about 28MB of memory.
This model primarily handles tool calling, device control, and structured information extraction. Developers tell it the tool names and parameter formats, and the model returns corresponding JSON with a confidence score. For example, if a user says "turn on the living room lights," this kind of task doesn't necessarily require calling a large model. Processing it locally on the device is faster and doesn't require sending content to the cloud.
However, the small parameter count also means its ability to handle complex problems is limited. This project is suitable for relatively fixed on-device tasks and shouldn't be directly compared to general-purpose large models.
5. semantica: Making Agent Results Traceable
Repository: https://github.com/semantica-agi/semantica
semantica is a graph-structured Context infrastructure. It gained 3,674 stars this week and currently has fewer than 10,000 stars, but it has already reached fifth place on the weekly chart.
It extracts entities and relationships from data and saves them into a Context Graph and knowledge graph. Later, when an Agent obtains a result based on this data, it can trace back to see which information was used and which relationships were traversed in between.
Vector databases are good for finding semantically similar content, but many business scenarios also require preserving explicit relationships. For example, which document a conclusion cited, which version of the data was used, and what processing steps were involved—these things are not well represented by similarity alone. semantica aims to fill this gap.
In its project description, it calls itself the open-source Palantir, which is a rather ambitious positioning, and the project is still in its early stages. I think you can look at its ideas first; it's still hard to say how much work remains before it can be put into a real business environment.
6. Switchyard: Adding a Routing Layer for Different Models
Repository: https://github.com/NVIDIA-NeMo/Switchyard
Switchyard is an LLM traffic proxy and routing library open-sourced by NVIDIA NeMo, developed in Rust. It can translate protocols between OpenAI Chat, OpenAI Responses, and Anthropic Messages, then forward requests to vLLM, NVIDIA NIM, Ollama, or other services compatible with the OpenAI API.
For instance, if an Agent originally uses the Anthropic API but later wants to connect to a local model, modifying every call site would be cumbersome. With Switchyard sitting in the middle, the Agent still sends requests using the original API, and Switchyard selects the backend model and translates the format.
It also supports random routing, model-classification routing, and phased routing, which can be used for A/B testing or for routing simple tasks to cheaper models and complex tasks to more capable ones. However, the repository explicitly states that the project is still in a pre-alpha stage, and both the API and algorithms may continue to change. It's not suitable for production use yet, so be careful.
7. Modular: Mojo and MAX in a Single Repository
Repository: https://github.com/modular/modular
The modular repository contains the open-sourced parts of the Modular Platform, primarily the Mojo language and the MAX Framework. The repository currently has over 27,000 stars, gaining 744 this week.
Mojo is a programming language for AI and high-performance computing, while MAX provides model inference, acceleration libraries, model pipelines, and an OpenAI-compatible inference service. Now, the Mojo compiler, standard library, parts of the MAX Kernel, inference services, and example code are all in this repository. If you want to see the current state of Mojo's progress, you can look directly here.
However, the licenses in this repository are not entirely uniform. The open-source code mainly uses Apache 2.0 with LLVM Exceptions, but the use and distribution of MAX are also subject to the Modular Community License. You need to check this separately before incorporating it into your own projects.
8. MoneyPrinterTurbo: Generate Short Videos from a Topic
Repository: https://github.com/harry0703/MoneyPrinterTurbo
The eighth project is MoneyPrinterTurbo. This project has been open-source for two years, but it gained another 9,712 stars this week, bringing its total to over 110,000 stars.
To use it, you input a topic or keyword, and it generates a script, searches for materials, adds subtitles and background music, and finally synthesizes a short video. The project provides a WebUI and an API, and it runs on Windows, macOS, and Linux.
One thing to note is that the final generated video is still affected by the script model, materials, and keyword extraction. If you plan to publish it officially, you need to double-check the content and copyright yourself.
Finally
That's all for this week's GitHub trending list. There are more and more projects related to Agent Infra and on-device AI. I look forward to seeing more interesting open-source projects. See you next week!