Design
Preset
Background
Text
Font
Size
Width
Account Pricing

The Git Times

AI Models
Claude Opus 5.5 $20/M —GPT-6.1 Sol $10/M —Gemini 4 Argon $10/M —Grok 4.7 $6/M —DeepSeek V4 Pro 0813 $1.98/M —Qwen3.8 Max Prime $12/M —Kimi K3 $11.94/M —
Full Markets →

Open-Source Tool Lets Your Laptop Pool Every Network Connection It Has 🔗

net-merge splits traffic across Wi-Fi, Ethernet and phone tethers at once, with a free Go codebase that runs on Windows, macOS and Linux.

rohit267/net-merge · Go · ▲ 267 in 1d 5d old · Latest: v0.3.0

Most people sit on one internet connection at a time, even when several are within reach. net-merge, a Go project on GitHub, changes that by using all of a machine's network adapters together.

Wi-Fi, Ethernet, USB tethers and phone hotspots combine into a single pool, and the speed of each link adds up for multi-connection workloads.

The mechanism is per-connection load balancing rather than packet splitting. Each new TCP or UDP connection is assigned to one link, so an application that opens many connections spreads them across every available path. Steam downloads, browsers, cloud uploads and torrent clients benefit most. The project's own test, run with Wi-Fi, an Android USB tether and an iPhone hotspot, reported 232 to 303 Mbit/s across nine simultaneous downloads, against roughly 121 Mbit/s for the fastest link alone.

Several design choices stand out. Faster links receive more connections, and a link without internet access receives none until it recovers. The tool sits on TUN and Wintun interfaces, so applications need no proxy configuration. If net-merge crashes, the machine's normal routing keeps working. It claims about 0.18 CPU cores per 100 Mbit/s and carries no account, server or VPN dependency.

Release 0.3.0 adds native packages for Debian, Fedora, openSUSE, Alpine and Arch across nine CPU architectures, including Raspberry Pi targets, and pauses merging automatically when fewer than two links work. Installation is a single paste into a terminal or PowerShell.

The catch: A single large file over a single connection will not get faster, because each connection rides one link. The project has four open issues, the Windows installer is not yet code-signed and triggers SmartScreen warnings, and the project is young at version 0.3.0. Builders should test failover behavior on their own networks before relying on it for anything important.

Why this leads today Bandwidth aggregation is an old idea, but a single open-source TUN layer working across three desktop systems puts it within reach of everyday builders, a more practical change than a placeholder repo or niche benchmark.

Use Cases
  • Gamers pooling Wi-Fi and phone hotspots to speed up Steam downloads.
  • Remote workers combining Ethernet and tethered phones for faster cloud uploads.
  • Raspberry Pi builders aggregating multiple links on ARM-based gateway devices.

Source: rohit267/net-merge — based on the README and release notes.

More on the Front Page

Native PC Port Lets Gamers Run PS4 Title inFAMOUS Second Son Without Emulating the CPU 🔗

Community project reimplements PS4 system layer and translates GPU calls to Vulkan, with motion-sensor emulation and Linux and Windows builds.

HayatoG/iss_pc · C++ · 533 stars 4d old

A community project called iss_pc is porting inFAMOUS Second Son (PS4, CUSA00223, base 1.00 plus 1.

07 update) to native PC. The original game code runs directly on the host processor. The PS4 system environment is reimplemented in software, and the GPU workload is translated to Vulkan. Because the CPU is not emulated, the approach avoids the overhead that slows conventional console emulators.

The project ships no game files. Users must dump their own copy from a PS4. The core is shared with ports of The Last Guardian and Gravity Rush 2, and the project reports the game reaches gameplay and the first chapters.

Recent builds add a Qt Quick launcher, a save-management system with pre-launch backups and .zip export, and twelve experimental fixes, labelled BP01–BP12, that were backported from other PS4 ports and the shadPS4 project. A fix for flickering lights and light bleeding through characters was confirmed on both Windows and Linux. Motion-sensor and touchpad input are emulated for non-Sony controllers through L3+R3 toggling.

Builds are available for Linux x86-64 (glibc 2.39+) and Windows 10/11.

The catch: Performance remains unstable, at roughly 20–30 FPS on an RTX 2060, and known rendering faults persist. The tracker holds 158 open issues, and the project is still labelled in development, so it is not a finished replacement for the original hardware.

Use Cases
  • PC gamers who own a PS4 copy and want native playback of inFAMOUS Second Son.
  • Preservation researchers studying PS4 system reimplementation and native console-port techniques.
  • Linux developers interested in Vulkan-based GPU translation layers for console software.

Source: HayatoG/iss_pc — based on the README and release notes.

Blender Gets an AI Copilot That Edits Scenes Through Chat 🔗

Open-source add-on lets artists and developers route OpenAI, Claude, Gemini, or local Ollama models into Blender's scene, materials, and export tools.

LiCinCan/blender-ai-mcp-addon · Python · 723 stars 2d old

The mcp-for-blender-pro add-on puts an AI assistant inside Blender. It inspects the open scene, builds objects, edits materials, finds assets and runs multi-step tasks, with each action shown in a sidebar chat and checked in the viewport.

Larger jobs run through a Team chain. A manager splits the request into assignments, workers perform scene operations one at a time, and reviewers approve or send back findings. A blocked task gets one correction attempt before the workflow stops.

Connections cover OpenAI, Claude, Gemini, OpenRouter, Ollama and custom endpoints. The bundled Model Context Protocol server exposes 21 tools for creation, editing, inspection and export. Install it from a ZIP on Windows x64 with Blender 4.2 or later.

The catch: the add-on is a first release with four open issues, and it only ships a Windows x64 build, so macOS and Linux users must wait or build it themselves.

Use Cases
  • Product artists arranging display layouts and assigning materials through chat.
  • Technical artists automating repetitive scene cleanup and export preparation.
  • Developers connecting local Ollama models to Blender via MCP.

Source: LiCinCan/blender-ai-mcp-addon — based on the README and release notes.

Rust Headless Browser Gives AI Agents a Lean Web Runtime 🔗

Moli speaks CDP, WebDriver Classic and BiDi from one endpoint, and renders pages on demand to keep memory footprints small.

lexmount/moli · Rust · 19.4k stars 2mo old

Moli, a headless browser written in Rust, is aimed squarely at AI agents that need to fetch pages, extract content, and drive real web interactions. It runs on Linux, macOS and Windows, and exposes a single endpoint that speaks three automation protocols: CDP, WebDriver Classic and WebDriver BiDi.

Its design choice is on-demand layout and rendering. Pages are parsed and laid out only when a task requires it, which keeps the resource footprint light compared with a full desktop engine.

Agents can request output as Markdown, as a compact semantic tree, or as PNG and PDF exports. Explicit interactions are handled through --eval combined with --dump, so only the supplied expression changes page state.

Recent releases focus on navigation and renderer correctness, including fixes to Fetch configuration during navigation and detached-node resolution.

The catch: with 383 open issues and a young codebase, a builder should test Moli against their own target sites before trusting it in production.

Previously in The Times “covered” — Oct 11

Use Cases
  • Agent developers fetching and extracting pages as Markdown for language models
  • QA engineers running Playwright tests against a lightweight CDP endpoint
  • Data teams scraping dynamic sites into compact semantic trees for pipelines

Source: lexmount/moli — based on the README and release notes.

Kaku Lets You Reach Your Mac's Terminal From Any Browser 🔗

Version 0.24.0 adds remote access, one-time pairing and browser uploads for running AI coding sessions

tw93/Kaku · Rust · 6.1k stars 8mo old

The latest release of Kaku, a macOS terminal built on WezTerm, lets builders open their Mac's terminals in a browser on another machine. Running kaku remote exposes existing sessions, including ones already running Claude Code or Codex.

Access works through a few mechanisms:

  • Pairing: each link binds to one browser and works once. A bookmark reopens the page afterwards.
  • Relay: the relay forwards only content encrypted between the browser and the Mac.
  • Web terminal: tabs, splits and copy-paste match the native app. Files dropped on the page upload to the Mac, and Cmd+Click on a path downloads it.

Kaku ships with JetBrains Mono, automatic themes, shell completion and optional AI command suggestions, which you configure with your own service.

The catch: the project has one open issue and its last release arrived months ago, so the remote relay's security and reliability under real use remain unproven.

Use Cases
  • Developers checking long-running Claude Code sessions from a laptop away from their desk
  • Engineers uploading files to a Mac terminal from a second computer via browser
  • Teams sharing one Mac's Codex or shell session with a colleague through paired links

Source: tw93/Kaku — based on the README and release notes.

Hobbyist Tool Strips Ads and Bloat From Android TVs Over ADB 🔗

Reversible debloat script installs a clean, tvOS-style launcher built for 1 GB sets, scripted with Claude Code.

Gathin23/tv-vibehack · Java · 791 stars 1d old

Cheap Android TV and Google TV sets ship with preinstalled apps and promotional home screens. Gathin23/tv-vibehack removes much of that over ADB, then replaces the home screen with a minimal launcher.

The project works by:

  • Disabling packages rather than uninstalling them
  • Installing com.gathin.tvhome, a native launcher with a black background, a four-app dock, and D-pad navigation
  • Including a restore script that returns the stock launcher

Users enable developer options on the TV, then paste a prompt from PROMPT.md into Claude Code or another assistant. The assistant backs up the device, debloats it, builds the launcher, and checks each step with screenshots.

Its launcher is designed for sets with 1 GB of RAM and runs no background services or network calls.

The catch: the project was built in an afternoon, has 20 open issues, and its scope is limited to Android TV sets, so builders should confirm their model is supported before running it.

Use Cases
  • Owners of low-memory Xiaomi sets removing ads from their home screen
  • Developers scripting reversible debloat routines for fleets of Android TVs
  • Builders using AI coding agents to automate ADB device configuration

Source: Gathin23/tv-vibehack — based on the project README.

Single-File WebGPU Ocean Renders Sunset Sea Without Any Engine 🔗

Six FFT cascades simulate waves from kilometre swells down to millimetre ripples in one 61 KB page

Noniv/web-ocean · HTML · 602 stars 1d old

A new GitHub project, Noniv/web-ocean, renders a real-time sunset sea in the browser using nothing but WebGPU. It has no engine and no rendering library.

The whole scene opens from a single HTML file in recent Chrome.

The renderer runs a Tessendorf FFT simulation across six cascades of 256 by 256 samples. Wavelengths span roughly a kilometre down to a few millimetres, which is what produces the sun sparkle near the camera. The mesh is a 17-level CDLOD quadtree that reaches 160 km, with vertices 6.25 cm apart up close.

Builders get:

  • Whitecaps, sun glitter and a physically based sky with clouds
  • A free camera that stays above the water surface
  • URL parameters for resolution, camera pose and wind settings
  • A minified build at 61 KB, versus 102 KB unminified

Since the project was created on 16 October, it has drawn 97 forks and four open issues.

The catch: It requires a recent Chrome with WebGPU and a capable GPU, and laptops may silently fall back to integrated graphics unless forced to the dedicated one.

Use Cases
  • Game developers prototyping water shaders without adopting a full engine
  • Graphics students studying FFT ocean simulation in readable WebGPU code
  • Web designers needing a lightweight animated sea background for landing pages

Source: Noniv/web-ocean — based on the project README.

Open Source Web Projects Pivot Toward Browser-Native Compute and AI Tooling 🔗

Single-file WebGPU renderers, in-browser LLM routing, and reverse-engineered desktop UIs signal a shift away from heavy frameworks.

Trendweb-frameworks
Caho1/open-claude-uiNoniv/web-oceanlingdojo/kana-dojoTanStack/queryrzashakeri/beautify-github-profiletheopenco/llmgatewaycockpit-project/cockpitsrizzon/git-cityNaiboWang/EasySpiderAOrbitron/Eridanuscurl/curljstrieb/github-statsddalcu/mlx-serveheroiclabs/nakamaMewPurPur/GodSVG

A cluster of recent repositories suggests open source is collapsing the distance between "web app" and "native application." Noniv/web-ocean ships a real-time WebGPU ocean renderer with FFT waves and a physically based sky in a single 61 KB HTML file, no engine required.

It is a clear signal that the browser is now the compute target, not just the display layer.

Alongside it, AI tooling is moving into the same stack. theopenco/llmgateway routes requests across providers through one API, while ddalcu/mlx-serve runs an OpenAI- and Anthropic-compatible inference server natively on Apple Silicon with a Zig backend and no Python. Caho1/open-claude-ui reconstructs Claude's desktop interface as 30 React components and documents its interaction physics in a 3,000-line analysis manual.

Meanwhile, the data layer is consolidating around TanStack/query, which now spans React, Solid, Svelte, and Vue, and cockpit-project/cockpit provides a browser front end for server administration.

Technically, the pattern is: low-level languages (Zig, C, WebGPU) pushed into browsers and small binaries, and frameworks standardizing server-state handling across UI libraries.

The catch: Much of this is demo-grade. A single-file renderer proves capability, not production viability, and reverse-engineered UIs carry legal and maintenance risk. Cross-framework abstractions like TanStack's remain only as strong as their adapters, and browser GPU APIs still vary by platform. The trend is real; its maturity is not yet proven.

Use Cases
  • Graphics engineers prototyping GPU rendering without installing engines or toolchains.
  • Mac developers running local OpenAI-compatible LLM inference without Python environments.
  • Platform teams routing requests across multiple LLM providers through one gateway.

Open source data infrastructure is splitting into specialized, AI-ready layers 🔗

From lakehouse ingestion in Rust to a financial database built for safety, projects are narrowing focus while converging on AI workloads.

Trenddata-infra
PimpinPumpkin/VelaTanStack/queryPart-DB/Part-DB-servertheopenco/llmgatewaylakesoul-io/LakeSoulSnailclimb/JavaGuidemicrosoft/ML-For-Beginnersdeepspeedai/DeepSpeedNaiboWang/EasySpiderFFmpeg/FFmpegcurl/curltigerbeetle/tigerbeetleSanderMertens/flecs

The data-infra projects clustering on GitHub this quarter share a shape: each takes one hard layer of the stack and rebuilds it for a specific workload rather than offering a general platform.

LakeSoul, written in Rust, pitches itself as a cloud-native lakehouse combining realtime ingestion, concurrent upserts, incremental analytics and vector search in one system.

Its bet is that BI and AI pipelines should stop maintaining separate stores.

TigerBeetle, in Zig, targets the opposite end: a financial transactions database where correctness under failure matters more than feature breadth. Together with llmgateway, which routes requests across multiple LLM providers through one API, the cluster shows AI traffic becoming a first-class data path that needs its own routing, metering and analytics.

On the client side, TanStack Query handles server-state caching and fetching across React, Solid, Svelte and Vue, reflecting how frontends now treat remote data as a managed resource.

Meanwhile DeepSpeed continues to push distributed training efficiency, and Vela shows privacy-driven demand for maps and navigation that run without Google Play Services.

The catch: Much of this is convergence on buzzwords rather than proven architecture. "Lakehouse plus vector search" and "AI gateway" are early categories, several projects are single-maintainer efforts, and the pieces rarely interoperate yet. Specialization is real; a shared standard is not.

Use Cases
  • Data engineers consolidating batch and streaming pipelines into one open lakehouse store.
  • Fintech teams needing a correctness-first ledger database for transaction processing.
  • Platform teams routing requests across multiple LLM providers with unified cost analytics.

LLMs Are Leaving the Chat Window and Taking Over Your Tools 🔗

Open source projects now wire models into TV launchers, 3D editors, game engines, and native inference servers, using MCP and OpenAI-compatible APIs as the connective tissue.

Trendllm-tools
Caho1/open-claude-uiGathin23/tv-vibehackLiCinCan/blender-ai-mcp-addonNoniv/web-oceanrikkahub/rikkahubtheopenco/llmgatewayNirDiamant/GenAI_AgentsNirDiamant/agents-towards-productionAOrbitron/Eridanusddalcu/mlx-serveCoplayDev/unity-mcp

A new cluster of open source repos shows LLMs moving out of the chat box and into the software people already use. The common thread is the tool layer: models are given structured hooks into external applications rather than asked to produce text alone.

  • blender-ai-mcp-addon exposes Blender scene operations and bpy scripting to multiple model providers through MCP, turning a copilot into an actor inside the 3D pipeline.
  • unity-mcp bridges assistants to the Unity Editor so they can manage assets, edit scripts, and control scenes.
  • Eridanus centers its OneBot bot framework on LLM function calling, dispatching actions rather than replies.
  • tv-vibehack was built in an afternoon with Claude Code to debloat Android TV over ADB, a concrete example of agent-written system tooling.

Infrastructure is following the same logic. llmgateway offers one API across providers, while mlx-serve ships a native Zig inference server for Apple Silicon that speaks OpenAI and Anthropic protocols, removing Python from the stack entirely.

Technically, the pattern is convergence on compatible interfaces: MCP for tool exposure, OpenAI-style endpoints for model swapping, and code-first agent tutorials such as agents-towards-production for deployment.

The catch: Much of this is early and fragmented. Protocols differ in maturity, several projects are single-author experiments, and "built with Claude Code" is a weak signal of production reliability. Whether tool-wired models hold up outside demos remains unproven.

Use Cases
  • Blender artists using natural-language commands to drive scene edits and scripting.
  • Unity developers letting an assistant manage assets and scenes inside the editor.
  • Apple Silicon users running a private, Python-free local LLM server for their apps.

Deep Cuts

Eridanus Turns a QQ Chatbot Into an LLM Function-Calling Agent 🔗

Its Live2D desktop-pet mode runs without any QQ account attached at all.

AOrbitron/Eridanus · Python · 208 stars

Most OneBot bots route messages through hand-written command handlers. Eridanus replaces that routing with LLM function calling at its core, so the model decides which tool to invoke and with what arguments.

Built in Python on the OneBot 11 protocol, it connects to QQ through standard adapters and speaks to both openai and gemini APIs. That makes it a framework, not just a bot: you register functions, and the model chooses among them.

The more unusual feature is the pure Live2D desktop-pet mode. It runs the same LLM-driven character without requiring a QQ connection, which means the agent layer is separable from the chat transport. For builders experimenting with embodied or persistent assistants, that separation is the interesting part.

The catch: the project is small, Chinese-language documentation is the primary source, and its rough edges are unproven outside its original authors' setups.

Use Cases
  • Chinese-speaking developers building QQ bots with LLM tool routing
  • Indie builders prototyping Live2D desktop companions powered by Gemini or OpenAI
  • Framework experimenters testing function-calling patterns on OneBot 11 infrastructure

Source: AOrbitron/Eridanus — based on the project README.

Quick Hits

solana-classic An experimental SVM-based blockchain testnet aimed at art, NFTs, and meme communities, built in Rust for developers who want to build and experiment in the open. 529
MOSS-Image A 5B-parameter base diffusion transformer for image generation, offering strong text rendering, layout control, and efficient fine-tuning for custom visual models. 544
xscreensaver The listed repository is unofficial or unverified: no official XScreenSaver GitHub repo exists, so treat this entry with caution before cloning or installing anything. 792
kana-dojo A minimalist, Next.js-built platform for learning Japanese, inspired by Duolingo and Monkeytype, with plenty of beginner-friendly issues for new contributors. 3.7k
git-city Turns your GitHub profile into an interactive 3D pixel art building inside a shared city, letting you explore other developers' profiles as skylines. 5.8k
ethereum-flashloan-mev-bot A smart-contract arbitrage bot paired with an external automation script that controls its execution, illustrating flash-loan MEV strategies on Ethereum. 2.8k
Who shipped it

The Business Desks

Fresh on Hugging Face

Model Drops

The newest model releases builders are picking up right now.
Beyond GitHub

The AI Wire

What builders are reading today — the headlines, papers, and announcements that aren't trending repos.

From the labs & arXiv

OpenClaw Gateway Now Survives Sleep and Clock Jumps Without Restarting 🔗

Release 2026.10.4 also closes an owner-impersonation path in Tlon group DMs and unhangs shutdown on stalled workers.

openclaw/openclaw · TypeScript · ▲ 54 in 1d 10mo old · Latest: v2026.10.4

OpenClaw, the open-source assistant that runs on your own machine and answers through chat platforms such as Slack, Telegram, and WhatsApp, has fixed a reliability flaw that matters to anyone running it unattended. Its Gateway used to restart itself when its ownership heartbeat looked stale, which happened after laptops slept or system clocks jumped.

Release 2026.10.4 stops that behavior and recovers state ownership instead.

What changed

  • The Gateway keeps running through sleep and wall-clock changes.
  • openclaw gateway restart now works from verified external macOS terminals.
  • Shutdown and restart no longer hang on a stalled transcript worker or in-flight session events.
  • Tlon group DMs no longer treat other participants as the owner, closing an impersonation path.
  • Manual /compact settles the active run before compacting rather than aborting it.
  • Headless browser pages stay responsive while a site waits for a passkey.

The project's architecture separates a trusted Gateway from untrusted execution, with state, memory and credentials kept on the user's hardware. Models and agent harnesses are swappable plugins.

The catch: The project carries 9,399 open issues, a figure that signals a large backlog against a project of this age, and this release is a stability patch with no new features. Builders should check their own deployment against the open issue tracker before relying on it for long-running service.

Previously in The Times “covered” — Oct 17

Use Cases
  • Developers running a personal assistant on a laptop that sleeps daily.
  • Teams deploying a shared OpenClaw Gateway across Slack and Teams channels.
  • Privacy-conscious engineers keeping model credentials and chat state on local hardware.

Source: openclaw/openclaw — based on the README and release notes.

More Stories

Microsoft's Free Course Teaches Classic Machine Learning in 12 Weeks 🔗

Twenty-six lessons, 52 quizzes and a Scikit-learn focus let beginners skip deep learning entirely

microsoft/ML-For-Beginners · Jupyter Notebook · 91.6k stars Est. 2021

Microsoft's ML-For-Beginners repository offers a 12-week, 26-lesson curriculum on classic machine learning, built around Python and scikit-learn. It deliberately sets deep learning aside; that material lives in the company's separate AI for Beginners course.

Each lesson bundles a pre-lesson quiz, written instructions, a solution, an assignment and a post-lesson quiz. The lessons are project-based and use datasets drawn from many regions of the world.

  • Language: Jupyter Notebook, with R content alongside Python
  • Scope: Classic algorithms and workflows, not neural networks
  • Pairing: Designed to follow the companion Data Science for Beginners curriculum

The repository has been maintained since March 2021 and received a commit within the past day. Its sixteen open issues suggest active upkeep.

The catch: The course stops at classic methods, so learners who want modern deep-learning skills will need to move on to a separate curriculum.

Use Cases
  • Students learning scikit-learn fundamentals through structured, quiz-checked lessons
  • Instructors seeking a free, ready-made 12-week machine learning syllabus
  • Career switchers building portfolio projects on real-world regional datasets

Source: microsoft/ML-For-Beginners — based on the project README.

Open Tutorial Series Walks Builders From GenAI Prototype to Production 🔗

Repository now spans 28 company-contributed tutorials covering memory, guardrails, GPU scaling and multi-agent coordination

NirDiamant/agents-towards-production · Jupyter Notebook · ▲ 3 in 1d Est. 2025

The NirDiamant/agents-towards-production repository is a Jupyter-based collection of code-first tutorials for taking GenAI agents from notebook to live deployment. Its most recent push, on October 17, extends that catalog rather than changing its core.

The tutorials cover the unglamorous parts of agent work:

  • Stateful workflows built on LangGraph
  • Vector memory and RAG pipelines
  • Deployment via Docker and FastAPI endpoints
  • Operations: security guardrails, GPU scaling, observability and evaluation

Companies have contributed 28 step-by-step tutorials, and partners sponsor the project. The repository has accumulated 2,881 forks, a sign of active use.

Its paid course, Prompt to Production, is a separate product from the free notebooks.

The catch: the repository is a curated set of notebooks, not a framework, and its material is only as current as the libraries it demonstrates, so builders should verify each tutorial against today's versions before relying on it.

Use Cases
  • Backend engineers productionizing LangGraph agents behind FastAPI endpoints
  • Platform teams adding observability and guardrails to internal agent deployments
  • Startup developers evaluating GPU scaling options for multi-agent workloads

Source: NirDiamant/agents-towards-production — based on the project README.

YOLOv5 Adds Instance Segmentation Without Changing Its Simple Workflow 🔗

Version 7.0 segmentation models target real-time vision tasks, with training, validation and deployment handled through the same commands used for detection.

ultralytics/yolov5 · Python · ▲ 3 in 1d Est. 2020

The ultralytics/yolov5 repository, a PyTorch framework for object detection, classification and now instance segmentation, shipped v7.0 with segmentation models that Ultralytics says beat current real-time state-of-the-art benchmarks on MS COCO.

The release's main change is workflow. Builders can train, validate and deploy segmentation models using the same paths they already use for detection, rather than learning a separate pipeline.

Export support covers ONNX, TensorRT, CoreML and TFLite, so the same weights can move from a training box to edge hardware.

The project is about six years old and has been stable through several major versions. Its detect.py and PyTorch Hub entry points remain the quickest way to test a model on local images or video.

The catch: the repository is maintained, but the core code is one of several Ultralytics YOLO lines, and the company now steers new work toward the newer ultralytics package, so teams starting fresh should weigh whether YOLOv5 is the right base.

Use Cases
  • Robotics teams running real-time object detection on edge hardware.
  • Manufacturers training custom defect detectors on proprietary image datasets.
  • Researchers benchmarking instance segmentation against published COCO results.

Source: ultralytics/yolov5 — based on the README and release notes.

Quick Hits

GenAI_Agents Teaches Generative AI agent design through 50+ hands-on notebooks, progressing from simple conversational bots to coordinated multi-agent systems. 24.6k
DeepSpeed Speeds up distributed training and inference for large deep learning models, offering memory and parallelism optimizations that make scaling practical. 43.2k
ai-engineering-from-scratch A hands-on curriculum for building and shipping AI engineering projects, turning foundational concepts into working systems you can deploy for others. 70.8k
dify Combines agentic workflows, RAG pipelines, and broad model and tool support in one collaborative workspace, deployable from cloud to self-hosted. 158.8k
n8n Lets teams mix visual workflow building with custom code to automate processes across 400+ integrations, with native AI features and self-hosting options. 207.9k

ros2_control Gives Robot Builders One Framework for Hardware and Controllers 🔗

The C++ project, now on its ROS 2 Rolling, Lyrical, Kilted, Jazzy and Humble branches, decouples hardware drivers from control logic.

ros-controls/ros2_control · C++ · 1k stars Est. 2017

Robot builders have long wired motor drivers directly into control code, making every hardware swap a rewrite. ros2_control attacks that coupling.

It is a generic, deliberately simple framework for ROS 2 that separates the hardware interface from the controllers that command it.

The design splits the stack in two. Hardware plugins expose joints and sensors as state and command interfaces. Controllers, such as joint trajectory or velocity controllers, read and write those interfaces without knowing what sits underneath. Swapping a simulated arm for a physical one means changing a plugin, not the control logic.

The project is actively maintained across five ROS 2 distributions, with branches for Rolling, Lyrical, Kilted, Jazzy and Humble. Published Docker images under ghcr.io/ros-controls give teams a reproducible build environment. Documentation and API references are published per distribution, and the project credits major contributions from companies and institutions listed on control.ros.org.

Who should care: integrators bringing new actuators online, researchers moving controllers between simulation and hardware, and teams standardizing control code across robot models.

The catch: The project carries 149 open issues, so check the tracker for your specific hardware before committing. Its generality also means you own more integration work than with a vendor-specific driver stack.

Use Cases
  • Robotics integrators connecting new actuators to existing ROS 2 controller code
  • Researchers porting controllers between Gazebo simulation and physical hardware
  • Engineering teams standardizing joint control across multiple robot platforms

Source: ros-controls/ros2_control — based on the project README.

More Stories

BehaviorTree.CPP Adds Try-Catch Cleanup Nodes for Robot Control Logic 🔗

Release 4.9.0 also lets nodes exchange derived pointer types and wraps errors with the failing node's identity.

BehaviorTree/BehaviorTree.CPP · C++ · 4.3k stars Est. 2018

BehaviorTree.CPP, the C++17 library for building decision trees in robotics and game AI, shipped version 4.

9.0 with changes aimed at developers debugging live systems.

The headline addition is TryCatch, a sequence-like node with a cleanup branch that runs only when the sequence fails or is halted. Builders have requested it for years.

Other changes:

  • Polymorphic port support: nodes producing shared_ptr<Derived> can connect to ports expecting shared_ptr<Base>, including multi-level chains.
  • Exception tracking: errors thrown during tick() now carry the node's name, path, and registration name.
  • Parser swap: the lexy dependency is replaced by a hand-written Pratt parser, reducing compile time and binary size.

Trees remain XML files loadable at run time.

The catch: the project has 41 open issues, and the release notes do not say whether the new parser has been tested against the full range of XML files users already maintain.

Use Cases
  • Robotics engineers sequencing arm motions with guaranteed cleanup on failure
  • Game developers building NPC AI that loads behavior trees at run time
  • ROS 2 teams replacing finite state machines with reactive, concurrent behaviors

Source: BehaviorTree/BehaviorTree.CPP — based on the README and release notes.

TinyTorch v0.2.0 Teaches Students to Build a Deep Learning Library 🔗

Release adds transformer milestones, a TinyGPT trainer, and fixes for silent autograd and optimizer bugs

harvard-edge/cs249r_book · Python · ▲ 10 in 1d Est. 2023

The Harvard CS249r curriculum's TinyTorch framework now walks students through transformers, not just convolutional networks. Version `tinytorch-v0.

2.0` completes Milestones 05–07, the transformer-era stage of a 20-module course.

Students write a complete deep learning library in pure Python and NumPy, then use it to train a TinyGPT model on Shakespeare. Milestone 05 also adds TinyCopilot, a code-completion model scored against held-out prompts using Python AST parsing.

The release also fixes silent correctness bugs in autograd and optimizers, and removes test passes that were never verified. The lab guide now acts as an instructor rather than a reference.

The catch: the framework is pure NumPy by design, so students learn the internals but will not see GPU-scale training performance, and the project's seven open issues are a reminder that it is still maturing.

Previously in The Times “covered” — Oct 14

Use Cases
  • University instructors teaching a hands-on course on deep learning internals
  • Self-taught engineers building autograd and attention from scratch in NumPy
  • Students comparing their own optimizer and transformer code against a reference implementation

Source: harvard-edge/cs249r_book — based on the README and release notes.

Rerun Lets Robot Teams Stream Multimodal Data Straight Into Training 🔗

Version 0.39.0 adds raw Bayer image support, marimo notebook embedding and richer table controls for sensor datasets

rerun-io/rerun · Rust · 11.6k stars Est. 2022

Rerun's latest release, **0.39.

0**, targets teams whose robot and sensor data lives in many formats at once. The Rust-built toolkit ingests images, point clouds, transforms, joint states and video at different rates, then keeps them in sync in its viewer.

New in this version:

  • Raw Bayer images via a pixel-format encoding, so unprocessed camera output needn't be converted first.
  • PLY mesh and 2D point support when opening files.
  • Marimo notebook embedding of the Viewer.
  • Table column settings with search, display modes and bulk show/hide.

The same datasets can be queried with dataframes or SQL and streamed into training without export jobs. Python ships the Viewer in the rerun-sdk package; Rust and C++ need the separate rerun CLI.

The catch: the project carries 1,239 open issues, and the Rust crate for 0.39.0 is unpublished because it depends on an unreleased egui build, so Rust users must work from source or the CLI.

Previously in The Times “covered” — Oct 12

Use Cases
  • Robotics engineers scrubbing synchronized camera, lidar and joint-state logs.
  • ML teams streaming multimodal episode mixes directly into training runs.
  • Researchers running SLAM and hand-tracking pipelines as table edits.

Source: rerun-io/rerun — based on the README and release notes.

Quick Hits

image_pipeline Adds image processing components such as rectification, calibration, and stereo disparity to ROS, letting robots turn raw camera streams into usable perception data. 968
robotcode Gives Robot Framework testers a full development environment, with language server support, a debugger, and editor extensions for VS Code and PyCharm, plus CLI tools. 308
nicegui Lets Python developers build interactive web interfaces, dashboards, and tools without writing frontend JavaScript, using a clean, declarative API. 16.3k
newton Runs GPU-accelerated robot and physics simulations on NVIDIA Warp, giving roboticists and researchers a fast, open engine for contact-rich dynamics. 5.8k
OpenBot Turns a cheap smartphone and a roughly $50 electric chassis into a working robot, with Android software for person following and real-time autonomous navigation. 3.5k

Strix Gains Git Blame Context and Proxy-Verified Findings for Automated Pentests 🔗

Latest release ties vulnerability reports to HTTP evidence and local commit history, and adds CI/CD gating for pull requests.

usestrix/strix · Python · ▲ 167 in 1d Est. 2025 · Latest: v1.7.0

Strix, the open-source AI penetration testing tool, has shipped v1.7.

0 with changes that tighten how its findings are verified and reported. The release requires proxy-validated findings to link to the specific http_exchange_ids that prove them, so reports carry traceable request-and-response evidence rather than model assertions.

Reports now also pull in local Git blame for technical details. Agents are prompted to name the commit and author responsible for a vulnerable line, which helps teams route fixes to the right owner.

Other fixes matter for day-to-day use:

  • Model routing: Frontier-model checks now match on the model name alone, not the provider route.
  • Configuration: STRIX_API_TYPE forces either the Responses or Chat Completions API.
  • Sandboxing: Extra files are placed as agent-writable sandbox files on every backend, and read-only local sources mount as :ro bind mounts.
  • Dependencies: litellm>=1.101.0 is now required.

The project's GitHub Actions integration lets teams run scans on every pull request and block insecure code before merge. Strix runs inside Docker and requires an LLM API key from a supported provider.

The catch: The project has 477 open issues, and the release is a v1.x line whose changes are largely integration and reporting fixes. Teams running it in CI should test the STRIX_API_TYPE and sandbox changes against their own backends before relying on it as a merge gate.

Previously in The Times “covered” — Oct 15

Use Cases
  • Security teams gating pull requests with automated AI penetration testing in CI.
  • Bug bounty researchers generating reproducible proof-of-concept exploits quickly.
  • Developers auditing their own web applications before production deployment.

Source: usestrix/strix — based on the README and release notes.

More Stories

Self-Hosted IPsec VPN Installer Keeps Getting Easier to Deploy 🔗

Setup script now covers WireGuard, OpenVPN and Headscale alongside IPsec on one server

hwdsl2/setup-ipsec-vpn · Shell · 28.8k stars Est. 2016

The setup-ipsec-vpn project lets builders stand up their own VPN server with a single shell command. It runs Libreswan for IPsec and xl2tpd for L2TP, and supports IKEv2 with modern ciphers such as AES-GCM.

The script targets Ubuntu, Debian, CentOS/RHEL, Alpine Linux and Raspberry Pi OS. Credentials are generated at install time, and it produces profiles that auto-configure iOS, macOS and Android clients. Helper scripts manage users and certificates.

The project's wider ecosystem now includes a Docker image and optional installs of WireGuard, OpenVPN and Headscale on the same host.

Its ~1,000-plus-star-equivalent fork count and steady maintenance show a mature, widely reused tool.

The catch: The project is a single-purpose shell installer, so operators still own patching, firewalling and key handling, and it has a long track record but only one open issue on record to judge recent responsiveness.

Use Cases
  • Travelers securing laptops on hotel and airport networks with IKEv2
  • Small teams running a self-hosted VPN on a cheap cloud VPS
  • Hobbyists turning a Raspberry Pi into a home network tunnel

Source: hwdsl2/setup-ipsec-vpn — based on the project README.

Decepticon Lets Red Teams Start Autonomous Hacking Engagements Without Hand-Written Scope Files 🔗

Version 1.5.2 generates the rules of engagement from a name and target URL, then opens the Red agent directly.

BitterSecurity/Decepticon · Python · 6.3k stars Est. 2025

The latest release of Decepticon, an open-source autonomous penetration-testing agent, removes a setup step that blocked first runs. Previously, starting an engagement required a hand-written scope JSON and a manual approval command.

Now the CLI asks for an engagement name and a target web URL, creates the rules of engagement and matching approval automatically, and opens the Red agent ready for its first request.

The release fixes a first-run failure, Red scope is unavailable, that hit workspaces opened without scope.

  • Scope is enforced. Targets must be exact root HTTP or HTTPS URLs on default ports. Paths, custom ports, wildcards, credentials, queries and fragments are rejected.
  • Existing data survives. Engagements, approvals, conversations and database volumes carry over.
  • Planning is optional. Soundwave planning is available through /mode plan.

Specialist agents run in an isolated sandbox, and the optional Blue Cell sensor reports suspicious traffic without blocking it.

The catch: The project has one open issue and a 1.5.1 release that was withheld over packaging errors, so builders should verify engagement scope before pointing it at anything live.

Previously in The Times “covered” — Oct 11

Use Cases
  • Red teams launching authorized web-app penetration tests from a single CLI command
  • Security engineers evaluating autonomous agents against staging environments before production
  • Blue teams monitoring local web services for suspicious requests during controlled exercises

Source: BitterSecurity/Decepticon — based on the README and release notes.

Quick Hits

blackbird Searches for matching accounts across social networks using a username or email address, useful for mapping an identity's online footprint during investigations. 9.2k
maigret Compiles a comprehensive profile of a person from a single username by checking roughly 6,000 websites, with results organized for review. 39.2k
mastg A thorough open-source manual for mobile app security testing and reverse engineering, with step-by-step methods for verifying OWASP MASVS-aligned weaknesses. 13.3k
NetExec A multi-protocol network exploitation framework that automates authentication testing and command execution across large numbers of hosts in enterprise environments. 5.9k
caddy A web server written in Go that serves HTTP/1, HTTP/2, and HTTP/3 and provisions HTTPS certificates automatically, with minimal configuration. 78.1k
yakit A desktop platform bundling traffic interception, vulnerability scanning, and penetration testing workflows into one toolkit for security professionals. 7.8k

ClickHouse Ships Monthly Stable Releases as Real-Time Analytics Goes Mainstream 🔗

The open-source column store now targets AI observability and lakehouse workloads, with releases arriving roughly every month.

ClickHouse/ClickHouse · C++ · ▲ 13 in 1d Est. 2016 · Latest: v26.9.21.9-stable

ClickHouse, the C++ column-oriented database for real-time analytical reporting, continues to move on a fast cadence. Its latest stable tag, **v26.

9.21.9-stable**, lands on the heels of a monthly release train: the 26.9, 26.8, and 26.7 release calls all happened within the last quarter, each with recorded walkthroughs and slides.

The project's recent emphasis is visible in its event calendar. October brings a run of AI Builders and Databases meetups in Tel Aviv, London and Dubai, plus hands-on sessions on agent observability with Langfuse. The pairing signals where ClickHouse is pointing its users: storing and querying high-volume telemetry from LLM applications alongside conventional analytics.

The topic list also reflects the shift. Tags include lakehouse, ai, cloud-native and hacktoberfest, and the repository is mirrored into a browser-based code viewer on github.dev. For builders, the practical question is whether one engine can serve both dashboards and agent traces without a second store.

Architecturally, ClickHouse runs as a single binary that can be self-hosted, embedded, or consumed as ClickHouse Cloud. Its MPP design shards data across nodes and favors wide scans over row lookups.

The catch: Activity is high but so is the backlog. The repository carries 8,503 open issues, and its scale means that upgrading across monthly stable versions requires disciplined testing. Teams whose workloads are mostly point lookups or small transactional writes should evaluate whether a columnar OLAP engine fits before adopting it.

Previously in The Times “covered” — Oct 13

Use Cases
  • Platform teams storing LLM agent traces for real-time observability dashboards
  • Data engineers running lakehouse-style analytics over self-hosted petabyte-scale event logs
  • Application developers embedding a fast analytical engine for user-facing reporting features

Source: ClickHouse/ClickHouse — based on the README and release notes.

More Stories

CC Switch Gives Builders One Panel for Every Coding Agent's Config 🔗

Desktop app now routes Kimi K3, DeepSeek, and Claude Desktop traffic through one place, with fixes for settings resets.

farion1231/cc-switch · Rust · ▲ 210 in 1d Est. 2025

CC Switch, a Rust and Tauri desktop app, lets developers switch API providers in one click across Claude Code, Codex, OpenCode, OpenClaw, Grok Build, and Hermes Agent. It also manages MCP servers, skills, and prompts in one place, replacing hand-edited JSON, TOML, and YAML files.

The latest release, v4.0.15, targets reliability. It fixes a Windows bug where updates reset hidden apps and config directory overrides to defaults. Routed requests now recognize Claude Code sessions again, and a corrupted Codex tool call no longer breaks every later turn. DeepSeek accepts replayed tool history.

New features include a bar or area toggle for the usage trend chart, per-model image-input flags for Hermes, and model-list fetching for Claude Desktop providers that sign in through Copilot or Codex OAuth.

The catch: the project carries 1,492 open issues, and its scope spans six agent tools that each change their config formats, so breakage remains a live risk.

Previously in The Times “covered” — Oct 12

Use Cases
  • Developers switching between Claude Code and Codex providers without editing config files.
  • Teams managing shared MCP servers and skills across several coding agents.
  • Engineers routing agent traffic through relay services while tracking usage costs.

Source: farion1231/cc-switch — based on the README and release notes.

Syncthing 2.1 Adds Proxy Support and Device Grouping 🔗

Release lets users route sync traffic through HTTP and HTTPS proxies, and tune block indexing for smaller databases

syncthing/syncthing · Go · 89.5k stars Est. 2013

Syncthing, the open-source peer-to-peer file synchronizer, has shipped version 2.1.

6 with changes aimed at networks and large setups. Users can now route traffic through HTTP and HTTPS proxies that support CONNECT, not just SOCKS, by setting all_proxy.

The 2.1 line also adds a group attribute to organize devices and folders in the GUI. Folders can disable block indexing via blockIndexing, trading minimal transfer size for lower database overhead. GUI session length and cookie path are now configurable, including a non-expiring option.

Syncthing's design puts data safety and encryption ahead of convenience, and it runs on common desktop platforms, with Docker and APT options.

The catch: the project carries 388 open issues, and this is a point release on a 2.x line, so builders should check the issue tracker against their own platforms before rolling it out broadly.

Use Cases
  • Home users syncing files between laptops and phones without a cloud provider
  • Teams behind corporate HTTP proxies replicating folders across office machines
  • Sysadmins running containerized Syncthing nodes to back up shared directories

Source: syncthing/syncthing — based on the README and release notes.

RustDesk Offers Self-Hosted Alternative to TeamViewer and AnyDesk 🔗

Latest 1.5.0 release ships Windows, macOS, Linux, Android, iOS and web builds across x86-64 and ARM64

rustdesk/rustdesk · Rust · ▲ 82 in 1d Est. 2020

RustDesk's **1.5.

0** release packages the remote desktop client for Windows, macOS, Linux, Android, iOS and web, with x86-64 and AArch64 builds for most platforms. The release notes list installers, Flatpak and APK downloads, and a browser client at rustdesk.com/web.

The app is written in Rust with a Flutter GUI. Its selling point is control: users can run the project's rendezvous and relay servers, host their own, or build a replacement. Desktop builds also support the deprecated Sciter front end.

Building from source requires vcpkg with libvpx, libyuv, opus and aom, or the project's Docker container. The first build is slow.

The project has over 19,000 forks and was last pushed two days before this report.

  • The catch: The repository carries 190 open issues, and Sciter support is being phased out, so self-hosters should check which GUI path their platform supports before deploying.
Use Cases
  • IT teams hosting remote support without relying on third-party relay servers.
  • Privacy-conscious users needing peer-to-peer desktop access on Linux and Wayland.
  • Developers embedding a self-hosted remote desktop stack into internal tooling.

Source: rustdesk/rustdesk — based on the README and release notes.

Quick Hits

linux Provides the Linux kernel source code, the core operating system layer that manages hardware, processes, memory, and devices for systems worldwide. 252.3k
FFmpeg Offers a complete toolkit for recording, converting, streaming, and processing audio and video across nearly every format and codec. 65.4k
RuView Turns ordinary WiFi signals into real-time presence detection, vital sign monitoring, and spatial mapping without using any cameras or video. 98.6k
hugo Builds static websites at remarkable speed, letting developers ship large content sites in seconds with a flexible, templated workflow. 90.1k
codebase-memory-mcp Indexes codebases into a persistent knowledge graph across 162 languages, letting AI assistants answer structural queries in sub-milliseconds with far fewer tokens. 46.8k

Internet-in-a-Box Ships Calibre-Web Library Into Its Default Offline Stack 🔗

Release 8.2 adds ebook library management to every install size, including the smallest Raspberry Pi builds.

iiab/iiab · Jinja · ▲ 2 in 1d Est. 2017 · Latest: release-8.2

Internet-in-a-Box (IIAB) turns a Raspberry Pi or Linux machine into a local learning hotspot. Users connect phones, tablets, or laptops over a nearby Wi-Fi network and reach offline copies of Wikipedia, Khan Academy videos, OpenStreetMap, and other educational content.

Release 8.2 changes the default toolset. Calibre-Web is now installed and enabled by default, including on SMALL-sized installs. It adds an eBook library interface and, in experimental form, support for video, audio, and images.

The release also tightens operator tooling. iiab-diagnostics now reports the Calibre-Web version, systemd service status, and three log files. It includes the last 300 lines of /var/log/xklb.log, and pipes output through cat -v so control characters are visible when debugging.

Installation has two paths: a one-line installer for Raspberry Pi OS, Ubuntu 24.04+, Linux Mint 22+, and Debian 12+, or prebuilt microSD images for the Pi. Content is loaded and arranged through a drag-and-drop Manage Content interface.

The project's contributor guide was recently published, signaling a push to widen the volunteer pool beyond its core maintainers.

The catch: The project carries 365 open issues, and the experimental media support in Calibre-Web is not yet production-grade. Builders deploying to schools or clinics should test video and audio handling on target hardware before relying on it, and should plan storage carefully, since offline libraries grow quickly.

Previously in The Times “covered” — Oct 17

Use Cases
  • Rural schools deploying offline Wikipedia and Khan Academy on Raspberry Pi hardware.
  • Clinics and prisons offering curated medical and educational content without internet access.
  • Community libraries building local ebook collections accessible over a home Wi-Fi network.

Source: iiab/iiab — based on the README and release notes.

More Stories

Pocket-Sized Raspberry Pi Zero Camera Adds GIF Recording and Custom Filters 🔗

Open-source Optocam Zero now boots in five seconds and shoots 2592x2592 JPEGs from a 14500 battery.

dorukkumkumoglu/optocamzero · Python · 884 stars 6mo old

The open-source Optocam Zero, a pocket camera built on a Raspberry Pi Zero, has gained GIF recording and playback. Its newest additions are a custom filter maker, which lets users build up to five filters in the hotspot interface, and a faster boot.

The build is deliberately simple. Case parts are fully 3D printed, and the electronics use common off-the-shelf components. The 240x240 display previews at a steady 25-30 fps. Captures are 2592x2592 JPEGs saved in the background while preview continues.

  • Battery: 14500 Li-ion, roughly 70–80 minutes per charge
  • Charging: USB-C, usable while charging
  • Size: 51×71×18mm

Images transfer over the built-in hotspot interface. The repository includes the print files, parts list and build guide.

The catch: The project is about seven months old with a 1.2.1 release, and the latest release notes are cosmetic, so it is unclear how the design holds up across many independent builds.

Use Cases
  • Makers building a playful pocket camera from printed parts and common electronics.
  • Hobbyists wanting a Kodak Charmera-style toy camera they can modify.
  • Educators teaching embedded imaging with a Raspberry Pi Zero project.

Source: dorukkumkumoglu/optocamzero — based on the README and release notes.

Open-Source FPGA Wi-Fi Design Claims to Match Commercial Chips 🔗

Latest release adds deterministic IQ timing for MIMO and radar, and runs on a low-cost Zynq 7020 board

open-sdr/openwifi · C · 4.8k stars Est. 2019

The openwifi project has released v1.5.

0, a full IEEE 802.11a/g/n stack built on software-defined radio hardware. The Linux driver lives in openwifi, while the FPGA baseband design sits in the companion openwifi-hw repository.

The release's headline claim is that its digital design performs as well as, or better than, commercial Wi-Fi chips. The project says a comparison report backs this up, though readers should check it themselves.

Key changes include:

  • Improved PHY algorithms for indoor multipath, including better frequency-offset estimation and soft-decoding LLR calculation
  • Removed ADC/DAC FIFOs, giving more deterministic IQ timing for distributed MIMO and radar
  • Fixed CSI fuzzer bugs, so fuzzing and defuzzing work as intended

Support for the low-cost Zynq 7020 remains in place. The stack is dual-licensed under AGPLv3 and a commercial license, and it expects users to follow local spectrum rules.

The catch: the project carries 98 open issues, and its performance claims rest on its own test report rather than independent verification.

Use Cases
  • Researchers building custom 802.11 MAC and PHY experiments on SDR hardware
  • Engineers prototyping distributed MIMO systems needing coherent IQ timing
  • Security teams running Wi-Fi packet injection and CSI-based sensing tests

Source: open-sdr/openwifi — based on the README and release notes.

Kubernetes Feature Detector Cuts API Load in Topology Updates 🔗

Release v0.20.0 switches topology writes to server-side apply and adds a 2-minute kubelet timeout

kubernetes-sigs/node-feature-discovery · Go · 1.1k stars Est. 2016

Node Feature Discovery (NFD) detects hardware and kernel features on each Kubernetes node and publishes them as labels, annotations, and extended resources. Schedulers can then place workloads on nodes with the right CPU instructions, GPUs or network adapters.

The v0.20.0 release targets cluster scale. nfd-topology-updater now writes NodeResourceTopology objects with server-side apply in a single request, instead of a GET followed by an UPDATE. That lowers load on kube-apiserver and etcd.

Other changes:

  • Reads from the kubelet configz endpoint time out after two minutes.
  • Builds use Go 1.27 and Kubernetes v1.37 libraries.
  • nfd-gc no longer deletes data for a node recreated under the same name.

Operators must update RBAC for the topology updater.

The catch: the project has 41 open issues, and the topology updater change requires an RBAC update that existing deployments must apply before upgrading.

Use Cases
  • Platform teams scheduling GPU workloads onto nodes with matching hardware labels.
  • Cluster operators reducing apiserver and etcd write load in large topology-aware fleets.
  • HPC engineers placing pods that need specific CPU instruction sets or RDT features.

Source: kubernetes-sigs/node-feature-discovery — based on the README and release notes.

Quick Hits

gdsfactory Design chips, circuit boards, and 3D-printable parts entirely in Python, spanning photonics, analog, quantum, and MEMS hardware. 1.1k
OpenWeedLocator Detects weeds in crops and fallow fields from camera images using a low-cost, open-source device you can build yourself. 502
ESPHome-Treadmill-FTMS Turns an ESP32-S3 into a treadmill controller that supports Bluetooth FTMS, Zwift grade-matched incline, and Home Assistant integration. 40
streamdeck Provides the SDK for building custom plugins and actions that control Elgato Stream Deck buttons and displays. 266
litex Lets you assemble custom FPGA-based hardware systems with Python tooling instead of wrestling through vendor-specific design flows. 4.2k

Luanti Ships Security Fixes for Client and Server in 5.17.0 🔗

The open-source voxel engine, formerly Minetest, urges immediate upgrades; C++ core exposes Lua for game modding.

luanti-org/luanti · C++ · ▲ 1 in 1d Est. 2011 · Latest: 5.17.0

Luanti, the open-source voxel game engine formerly known as Minetest, has released version 5.17.

0 with fixes for critical security vulnerabilities affecting both the client and server. The project advises all users to upgrade immediately.

The engine is written in C++17 and exposes gameplay logic through Lua, letting developers build and modify voxel games without rebuilding the core. Its layout separates compiled bin, read-only share and writable user data, and worlds live as folders under user/worlds/. Server deployments can run from Docker, and builds are documented for GNU/Linux, Windows and macOS.

Version numbering follows major.minor.patch since 5.0. The 5.16.0 to 5.17.0 changelog is linked from the release notes, and the security advisory is the release's headline change. Operators running dedicated servers should treat this as a priority update rather than an optional one.

The project's tracker shows 1,505 open issues, a sign of a large contributor base and an active backlog. Builders evaluating the engine for a server or game should check the issue tracker for their specific platform before committing.

The catch: The release notes do not enumerate the vulnerabilities, and the changelog is the only detailed source. Teams running hosted servers should confirm the fix against their own deployment rather than assume the upgrade closes every exposure.

Previously in The Times “covered” — Oct 15

Use Cases
  • Game studios building custom voxel titles with Lua-scripted gameplay logic
  • Server operators hosting multiplayer Luanti worlds who need the security patch now
  • Modders extending existing voxel games through documented Lua APIs and configuration

Source: luanti-org/luanti — based on the README and release notes.

More Stories

OpenSpeedy Fixes Numpad Hotkeys and a Silent Speed-Slider Bug 🔗

Ring-3 hooks let Windows users change game speed on x86 and x64 processes without touching the kernel.

game1024/OpenSpeedy · TypeScript · 19.8k stars Est. 2025

The latest release of OpenSpeedy, the open-source game speed modifier, lets numpad keys work as hotkeys. Before, numpad digits were recorded as their main-keyboard counterparts, so a numpad 1 registered the wrong key and never fired.

With NumLock off, it registered an unrelated key entirely.

The fix reads physical key positions, so Num1 and 1 now appear as distinct bindings. A second patch repairs a settings-page bug where a speed hotkey changed game speed but left the displayed multiplier stale until restart.

OpenSpeedy works by hooking Windows system time functions from user mode, avoiding kernel changes.

The catch: the project has three open issues, and its own README warns that anti-cheat systems may ban accounts and that excessive speed can crash physics engines.

Use Cases
  • Single-player gamers slowing down fast-paced titles for precise timing practice.
  • Modders and researchers studying how games depend on system time functions.
  • Developers testing game logic under altered time scales before shipping builds.

Source: game1024/OpenSpeedy — based on the README and release notes.

Unity MCP Lets AI Assistants Drive the Editor Directly 🔗

Version 10.3.0 shipped October 4 as the project adds multi-instance routing and tool groups.

CoplayDev/unity-mcp · C# · 15.1k stars Est. 2025

MCP for Unity now lets an AI assistant operate the Unity Editor through natural language. The bridge uses the Model Context Protocol, so any compatible client can connect, including Claude Desktop, Cursor, VS Code and Gemini CLI.

Once connected, a model can:

  • Create scenes and GameObjects
  • Edit C# scripts
  • Manage assets
  • Run tests and profile builds

Recent work targets reliability. Recent release notes cover Console filter fixes, port discovery on reconnect, and Claude Code registration visibility. The project offers 50 tool entrypoints under the MIT license, and requires Unity 2021.3 LTS or newer plus Python 3.10+.

The catch: With 104 open issues and a fast-moving beta branch, builders should test it on non-critical projects before trusting an agent with their editor.

Use Cases
  • Indie developers automating scene setup and GameObject creation through chat prompts.
  • Unity teams using AI agents to run tests and profile builds.
  • Tool builders connecting local LLMs to Unity projects via Model Context Protocol.

Source: CoplayDev/unity-mcp — based on the README and release notes.

Fan-Built Super Mario Bros. Remake Adds Level Editor and Custom Characters 🔗

Version 1.1 fixes trampoline sounds, checkpoint bugs and localization, while a level-sharing hub and portable mode widen access.

JHDev2006/Super-Mario-Bros.-Remastered-Public · GDScript · ▲ 9 in 1d Est. 2025

The Super-Mario-Bros.-Remastered-Public project, a fully open-source rebuild of the original NES Mario titles, has reached its **1.

1-stable** release. The project recreates Super Mario Bros., The Lost Levels, Super Mario Bros. Special and All Night Nippon from the ground up in GDScript on the Godot engine.

The latest changes are mostly bug fixes:

  • Red Piranha Plants no longer linger in warp zones.
  • Checkpoints survive inside level packs.
  • A Goomba that spawned beside a checkpoint in 2-1 has been removed.
  • French and German localization has been updated.

Custom content is the larger draw. Resource packs reskin graphics and sound, custom characters can be added, and a level editor feeds a Level Share Square. A portable.txt file enables portable mode.

The catch: the project has 178 open issues and its physics still show jitter at corners and camera limits, quirks the maintainer attributes partly to Godot itself.

Use Cases
  • Game designers building and sharing custom Mario-style levels with a community.
  • Modders creating resource packs and custom characters for a classic platformer.
  • Godot developers studying a large, complete GDScript platformer as a reference project.

Source: JHDev2006/Super-Mario-Bros.-Remastered-Public — based on the README and release notes.

Quick Hits

GodSVG Edit SVG files through a structured, visual interface rather than raw markup, with a late-alpha build that runs on desktop platforms and in the browser. 2.8k
imgui Drop a immediate-mode GUI into your C++ project with minimal dependencies, ideal for debug tools, editors, and game interfaces without heavy framework overhead. 76.7k
nakama Stand up a scalable backend for games that handles multiplayer sessions, matchmaking, leaderboards, chat, and social features out of the box. 13.5k
flecs Build high-performance games and simulations on an entity component system that is fast enough to handle large numbers of entities in C or C++. 8.8k
shaders Customize the look of your Niri desktop with a ready-made collection of GLSL shader effects for windows and screen rendering. 253
The Git Times AI Desk
Ask about today's stories — or hit “Ask about this” on any article to focus on one.

Unlock the Git Times AI desk to ask about today's stories and the AI model market.

Upgrade to Premium
Answers by the Git Times AI desk · verify before you ship