Top New Open-Source GitHub Projects This Week: AI Agents, Web Tools & Dev Kits #211

Hey creators, welcome to top trending and open source GitHub projects this week, part one, where you'll discover powerful new tools on GitHub that can instantly level up your workflow. Today

we explore tools like Far 7B, an AI that automates real computer tasks and Revo Grid, a lightning fast data grid for web apps. Let's dive in and see what you can

apps. Let's dive in and see what you can build with these gems. >> Welcome back to Manu AGI tutorials. Here

we explore the exciting world of AI, latest AI tools for you. So don't forget to hit that subscribe button and the notification bell so you don't miss out on the latest AI insights. So let's

start today's video.

>> Project number one, Farra 7B ondevice AI that uses your computer like a human.

Imagine having an assistant that doesn't just answer questions. It actually uses your computer, clicking, typing, and navigating just like you would to get things done. That's what FAR 7B

things done. That's what FAR 7B delivers. It's the first Agentic small

delivers. It's the first Agentic small language model SLM from Microsoft built specifically for computer use. With only

about 7 billion parameters, Farra 7B is compact enough to run on local devices, giving you speed, privacy, and direct control. Instead of generating just

control. Instead of generating just text, Farra 7B operates via pixelin, action out. It sees screenshots of a web

action out. It sees screenshots of a web page or UI, reasons about what to do, then outputs exact coordinates for clicks, scrolls, or keystrokes. No need

for accessibility trees or separate parsers. Behind the scenes, it's built

parsers. Behind the scenes, it's built on a base multimodal model and trained using a synthetic data pipeline called Farragen that generated over 145,000 verified trajectories across tens of

thousands of real websites. On

benchmarks, Farra 7B achieves strong performance, even beating some larger Agentic models on live web interaction tasks. This model is aimed at

tasks. This model is aimed at developers, power users, and teams who want to automate web-based tasks, booking tickets, filling forms, comparing prices, summarizing info while

keeping data local, and preserving privacy. It runs on device or via

privacy. It runs on device or via self-hosted infrastructure with much lower latency and cost compared to cloud-based agents. Give it a try and

cloud-based agents. Give it a try and see how your browser-based workflows transform into hands-free automated actions. Project number two, Ora

actions. Project number two, Ora browser. Clean, native Mac OS browser

browser. Clean, native Mac OS browser with minimalist design. Here's something

that feels like the future. Imagine a

web browser that's light, fast, and feels truly native on Mac OS with smooth tab management and a design focused on reducing noise. That's what Aura browser

reducing noise. That's what Aura browser brings. It's an open- source browser

brings. It's an open- source browser built with Swift and WebKit designed to deliver a clean, secure, and beautiful browsing experience specifically for Mac OS. Aura offers native feeling

OS. Aura offers native feeling performance, a vertical sidebar for tab/spaces rather than cluttered horizontal tab bars, an interface design inspired by

modern minimalist browsers. It signs and verifies its app updates using ED25519 cryptographic keys, a detail that underscores its emphasis on security and

user trust. This browser is ideal for

user trust. This browser is ideal for users who value privacy, simplicity, and speed over feature bloat. If you prefer a distraction-free browsing session on Mac OS, Aura gives you a sharp, native

looking alternative to mainstream browsers. Give it a try and see how

browsers. Give it a try and see how smooth and clean your browsing feels.

Project number three, Claude Code Open Router. Route Claude Code to any LLM via

Router. Route Claude Code to any LLM via Open Router. Imagine you're using Claude

Open Router. Imagine you're using Claude Code, the terminal-based AI coding assistant, but instead of being locked to a single provider, you can call almost any large language model by name

just by writing its name in your prompt.

That's what claude code open router does. It adapts claude code to route

does. It adapts claude code to route requests through open router giving you access to a wide variety of models like Grock, Gemini, GLM 4.6, GPT5 and more

via a unified interface. The project is MIT licensed and lightweight. You grab

your open router API key, drop it into a nenbin, then claude code plus this agent will intercept prompts like use grock to summarize this file or ask Gemini to

rewrite this function, forward them to open router and save output into a local file under llm-output/phere.

This matters now because it breaks you free from model lock, giving developers, hobbyists, and AI toolbuilders flexibility to pick a model by need.

Lowcost models for quick tasks, heavier models for complex work while staying inside the same workflow. It bridges

Claude Code's convenience with Open Router's model agnostic reach. Give it a try and see how your coding and prompt experiments multiply across model choices. Project number four, Folio AI

choices. Project number four, Folio AI powered GitHub portfolio generator.

Imagine you drop your GitHub username into a tool and instantly get a polished professional portfolio website without writing a single HTML line yourself.

That's exactly what Folio does. It's an

open-source GPL 3.0 web app built with TypeScript/Next.js that fetches your GitHub data via the GraphQL API, runs AI to generate summary

texts, highlights, and descriptions, and then renders a full responsive portfolio site. Folio uses a backend with Posgrex

site. Folio uses a backend with Posgrex Prisma for data storage and caches GitHub metadata to reduce rate limits and speed up generation. It also

optionally integrates LinkedIn data and can fetch live project screenshots for rich visuals. As soon as you supply a

rich visuals. As soon as you supply a username and optionally a GitHub token to include private repos, Folio builds your profile, project list, languages/contributions,

graph, etc. The result, a readyto-share, clean portfolio site ideal for developers, freelancers, and creators who want to show off their code and projects quickly and elegantly. If you

want to save time and get a professional web presence from existing GitHub work, Folio is worth trying. See how fast your portfolio goes live. Project number

five, RetJS visual programming engine and node editor toolkit. Imagine

dragging boxes and wires in a browser.

And those boxes actually do work. That's

what RetJS brings to life. It's an

open-source TypeScript first framework that helps developers build visual node-based editors and workflows directly in the browser. With RetJS

brings to lifeJS, you define nodes with inputs, outputs, or controls. Connect

them visually like wiring and run them through either a data flow engine or a control flow engine to execute logic or process data. The framework supports

process data. The framework supports major UI stacks, React, Vue, Angular, even LIT. So you can integrate a node

even LIT. So you can integrate a node editor into your web app without reinventing the wheel because it separates visualization from logic.

RetJSJS gives creators flexibility. You

pick exactly which plugins and UI frameworks you need. No unnecessary

bloat. This makes it ideal for building things like visual data pipelines, drag and drop automation tools, education apps, game logic editors, or custom workflow designers. If you want a

workflow designers. If you want a powerful way to turn codeheavy logic into an accessible graphical interface and give users a way to build, connect, and run workflows by drag and drop, give

RetJS Carl 1JS a try and see how your next project becomes visual and interactive. Project number six, Aloha

interactive. Project number six, Aloha Mini. Budgetfriendly dual arm robot you

Mini. Budgetfriendly dual arm robot you can build at home. Imagine opening a box in your room and assembling a fully functional two-ear armed robot, no factory needed, that can reach from the

floor up to table height. perceive with

multiple cameras and run realworld tasks. That's exactly what Aloha Mini

tasks. That's exactly what Aloha Mini delivers. An open-source 3D printable

delivers. An open-source 3D printable dual arm mobile robot with a motorized vertical lift. Aloha Mini gives you a

vertical lift. Aloha Mini gives you a robot platform that's affordable, US $600 if you self-print and assemble, yet surprisingly capable. It includes a 060

surprisingly capable. It includes a 060 cm vertical lift for floor-to-table reach, a five camera perception system, top, front, back, and dual arm cameras,

and in full hardware plus software open source under Apache 2.0, making it ideal for robotics lovers, educators, researchers, or hobbyists wanting

hands-on experience. The repository

hands-on experience. The repository provides 3D printable STL files, a bill of materials, and instructions for assembly and software setup. You can

swap the onboard compute either Raspberry Pi according to your needs and the design is modular enough for custom upgrades. It matters now because

upgrades. It matters now because robotics is often expensive and closed source. Aloha Mini lowers the bar,

source. Aloha Mini lowers the bar, giving anyone access to a capable mobile manipulation platform for experimentation and learning. Try

building it and see how easily a homegrown robot turns into a sandbox for embodied AI automation or creative robotics projects. Project number seven,

robotics projects. Project number seven, Splinter memory efficient tokenizer for LLMs. Imagine a tokenizer for large language models that is fast, memory

efficient, and works in plain Python.

That's what Splinter gives you. Splinter

is an open- source Python package on GitHub under Splinter-Ml/Splinter.

Designed to make tokenization and token handling easier for ML workflows built with PyTorch. Splinter handles text

with PyTorch. Splinter handles text tokenization under the hood, but in a way that focuses on efficiency. Ideal

when you build or train LLMs or other NLP models and want a lightweight tokenizer that doesn't hold up your pipeline. Because it's a Python library,

pipeline. Because it's a Python library, it can be dropped into existing projects with minimal friction. The simplicity

and light footprint make it useful for researchers, developers, and engineers working with custom models or data sets.

It helps especially when you care about resource usage, speed, and clean integration into Python-based ML stacks.

Give it a try and see how smoothly your token pre-processing becomes part of your ML workflow. Project number eight, Flux.2. NextG open image gen and editing

Flux.2. NextG open image gen and editing AI. Imagine describing a scene in words

AI. Imagine describing a scene in words or providing a few reference pictures and getting back a highquality photorealistic image or a refined edit that respects lighting, style,

composition, text, and even consistent characters or products. That's exactly

what Flux.2 delivers. It's an openweight image generation and editing model from Black Forest Labs, released under Apache 2.0 or a non-commercial license for some

variants. Flux.2 2 is a 32 billion

variants. Flux.2 2 is a 32 billion parameter rectified flow transformer that works both for textto image generation and imageto image editing. It

can accept up to 10 reference images to maintain style, character or object consistency so you can generate coherent multi-seene art, product mockups or redesigns all in one go. What makes it

matter now is its leap in realism and utility. Flux.2 2 produces outputs with

utility. Flux.2 2 produces outputs with 4 megapixel resolution, handles complex typography and layouts, and understands structured prompts, making it suitable

for marketing visuals, UI mock-ups, concept art, or professional-grade image work. This tool targets digital artists,

work. This tool targets digital artists, creators designers marketers anyone needing highquality visuals with full creative control without needing to craft every pixel by hand. Try it now

and see how smoothly your concept to image workflow transforms. Project number nine, Revo Grid. Excelstyle

higherformance data grid for web apps.

Imagine embedding an Excel-like spreadsheet inside your web app that can display millions of cells, scroll smoothly, let users edit, filter, sort, copy, paste, yet remain fast,

responsive, and lightweight. That's what

Revo Grid delivers. It's an open-source MIT licensed data grid component built with StencilJS compatible with major JS frameworks React, Vue, Angular, Spelt,

PlainJS. At its core, Revo Grid uses a

PlainJS. At its core, Revo Grid uses a virtual DOM plus smart row/column virtualization. So only visible data is

virtualization. So only visible data is rendered, meaning even huge data sets are handled with minimal performance hit. You define columns and data source,

hit. You define columns and data source, ei array of objects, and Revo grid handles rendering, editing, resizing, custom cell templates as editors, pinned, frozen rows or columns,

grouping filtering sorting exporting data, CSV, Excel, copy paste, much like a spreadsheet. Developers building

a spreadsheet. Developers building dashboards, analytics tools, inventory or finance web apps, data heavy admin panels. Anyone needing large-scale

panels. Anyone needing large-scale tabular data with interactive editing benefit from Revo Grid. It gives them spreadsheet class features without external dependencies or serverside rendering and works across frameworks

with the same API. If you handle large data in a web interface, try Revo Grid.

It might turn your data grid pain into smooth spreadsheet-like interactivity instantly. Project number 10, Infidex

instantly. Project number 10, Infidex high-performance.net search engine with fuzzy matching. Imagine building an app

fuzzy matching. Imagine building an app and being able to embed a search engine so fast and smart that it finds what you want, even if your users type wrong.

That's what Infidex brings to your net projects. This open-source MIT licensed

projects. This open-source MIT licensed search engine indexes documents in thousands per second and returns search results in mere milliseconds. Infidex

uses a unique pattern recognition approach. It extracts features like term

approach. It extracts features like term frequency and rarity from your data.

Embeds them into a multi-dimensional hypersphere and uses that for fuzzy matching enabling typo tolerant intelligent search without manual tuning or linguistic rules. It supports

advanced features like multi-field search with configurable weights faceted search and a SQL like query language called infascript for complex filtering and boosting. Integration is simple.

and boosting. Integration is simple.

Install via Nougat, create an engine instance, index your documents, and immediately start searching. Infidex is

fully thread safe, supports incremental indexing, add or update docs without reindexing everything, and exposes a clean API ideal for developers building content management systems, e-commerce

sites, knowledge bases, or any app needing fast, robust search without heavy dependencies. Give it a try and

heavy dependencies. Give it a try and watch how quickly your application search functionality becomes smarter and more userfriendly. Thanks for watching.

more userfriendly. Thanks for watching.

If you like this roundup, please hit like, subscribe to the channel, comment which repo you loved most, and enable notifications so you don't miss part two. Don't forget to start the repos and

two. Don't forget to start the repos and try out the demos linked in the description. And if you know more cool

description. And if you know more cool projects I should cover next time, drop them in the comments below.