Design
Preset
Background
Text
Font
Size
Width
Account Pricing Monday, July 27, 2026

The Git Times

“You never change things by fighting the existing reality. To change something, build a new model that makes the existing model obsolete.” — Buckminster Fuller

AI Models
Claude Opus 5 $25/M GPT-5.6 Luna $6/M Gemini 3.1 Pro Preview $12/M Grok 4.5 $6/M DeepSeek V4 Pro $0.87/M Qwen3.7 Max $4.42/M Kimi K3 $15/M
Full Markets →
Fresh on Hugging Face

Model Drops

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

Just Shipped

Significant new releases from the AI and dev-infra repos builders run on.

ML for Trading 3rd Edition Ships Precomputed Artifacts for Rapid Strategy Testing 🔗

Developers can now download case study outputs—model predictions, backtests, and fitted models—to skip retraining and iterate faster on quantitative workflow.

stefan-jansen/machine-learning-for-trading · Jupyter Notebook · 20.1k stars Est. 2018 · Latest: v3.0.0-artifacts

The stefan-jansen/machine-learning-for-trading repository released version v3.0.0-artifacts, bundling precomputed results for all nine case studies from the book’s third edition.

This release targets developers who want to inspect strategy performance, model behavior, or backtest mechanics without waiting hours or days for retraining on large datasets. Artifacts range from 33 MB for the etfs case study to 1.6 GB for the high-frequency us_equities_panel, with each asset isolated so users download only what they need. The download script supports selective retrieval—uv run python scripts/download_artifacts.py --cs etfs fetches just the ETF workflow, including 10 prediction sets, 7 training runs, and 55 backtests. This accelerates experimentation by letting developers focus on strategy refinement, feature engineering, or deployment logic instead of recomputing baseline results. The repo continues to integrate modern ML techniques like PatchTST, TabPFN, and retrieval-augmented generation for financial research, maintaining its end-to-end workflow from data sourcing to live execution. The catch: Artifacts are tied to specific book versions and dependencies; reproducing or extending results may require matching the exact Python environment, library versions, and data sources used in the original runs, which could pose a barrier for teams using divergent stacks or seeking to validate findings independently.

Why this leads today The repo offers a rare, production-grade framework for live trading systems, filling a critical gap in reliable, end-to-end tools for quantitative developers navigating real-world execution and risk.

Use Cases
  • Quant researchers validate ML trading strategies using precomputed backtests
  • Data scientists inspect model predictions from time-series architectures like PatchTST
  • Engineers accelerate workflow iteration by skipping expensive retraining steps

Source: stefan-jansen/machine-learning-for-trading — based on the README and release notes.

More on the Front Page

Crisp Unifies External Display Controls in a Free macOS Menu Bar App 🔗

It replaces paid tools like BetterDisplay with native SwiftUI control over brightness, HiDPI, and virtual displays.

didriksg/Crisp · Swift · 340 stars 4d old

Crisp consolidates macOS’s scattered external display settings into a single menu bar panel, offering hardware DDC brightness control, HiDPI scaling on any monitor, and virtual display creation—all in SwiftUI. Unlike BetterDisplay or Lunar, it’s free, open-source, and requires no subscription.

The app routes brightness keys to the display under the cursor, enables true darkness below hardware floors, and lets users save named presets for resolution, brightness, and arrangement. Recent updates added disconnect/reconnect for physical displays on Apple Silicon, drag-to-arrange canvas with main display switching, and Simplified Chinese support. The panel now mirrors macOS Control Center’s Displays UI with refined typography and effect tiles for Dark Mode, Night Shift, and True Tone. Install via brew install --cask didriksg/tap/crisp or download the unsigned .dmg from GitHub.

The catch: Crisp requires macOS 15 (Sequoia) or later, leaving older Intel Macs on Monterey or Ventura without support despite its Apple Silicon display disconnect feature depending on newer hardware.

Use Cases
  • Developers scaling external 4K monitors to HiDPI for sharper code and UI
  • Designers switching ICC profiles and presets for color-accurate workflows
  • MacBook users managing brightness and arrangement across multiple displays via menu bar

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

Vercel Labs' Scriptc Compiles TypeScript to Native Binaries Without a JavaScript Engine 🔗

The tool uses build-time execution and explicit FFI to produce small, fast executables with zero runtime overhead for static code.

vercel-labs/scriptc · TypeScript · 1.1k stars 4d old

Scriptc compiles TypeScript directly to native code by running type-checked programs at build time in an isolated VM and baking results as literals. It avoids embedding Node, V8, or any JavaScript engine in the output binary unless --dynamic is used for any types or npm dependencies.

Checked casts like JSON.parse(...) as Config insert runtime validation that throws descriptive errors on type mismatches. The project uses a typed IR as its core interface, with LLVM as the default code generator and a readable C backend for debugging. Recent v0.0.17 fixed Windows path handling and executable naming, enabling reliable cross-platform builds. The catch: As a pre-1.0 tool with only three open issues but limited real-world adoption, its long-term stability and performance at scale remain unproven for production systems.

Use Cases
  • Systems programmers building CLI tools in TypeScript
  • Embedded developers needing small, fast binaries from TS
  • Teams replacing Node scripts with native executables for latency-sensitive tasks

Source: vercel-labs/scriptc — based on the README and release notes.

OpenWork Lets AI Agents Share Workflows Across Tools and Teams 🔗

Desktop app enables cross-agent reuse of skills and services without lock-in

different-ai/openwork · TypeScript · 17.3k stars 6mo old

OpenWork v0.18.

3 released as an open-source alternative to Claude Cowork, letting users install the desktop app or integrate via MCP with Claude Code, Cursor, Codex, or other agents. It shares skills, plugins, and connections to Google Workspace and Microsoft 365 across tools and machines. Admins can publish capabilities and manage access for organizations. The catch: Windows installer remains unsigned pending production code signing, requiring manual override for installation.

Use Cases
  • Developers reuse AI workflows across Cursor and Claude Code
  • Teams share MCP-connected services via centralized admin interface
  • Individuals sync personal skills between desktop app and AI agents

Source: different-ai/openwork — based on the README and release notes.

Vorssaint Utils Unifies macOS Menu Bar Tools in One Open-Source App 🔗

Latest release fixes startup crashes and window-drag interference while adding per-app volume control

vorssaint/vorssaint-utils · Swift · 3.3k stars 1mo old

Vorssaint Utils consolidates over a dozen macOS utilities—volume mixer, app uninstaller, window switcher, clipboard history, and system monitor—into a single menu bar icon. Built in Swift and SwiftUI, it lets users install only the features they need, with each disabled component consuming zero resources.

The v3.1.15 update resolves startup failures, restores modifier-click behavior during window dragging, and fixes smooth scrolling and brightness keys. All tools run locally with no telemetry, subscriptions, or accounts required.
The catch: With 61 open issues and recent fixes to core stability, long-term reliability remains uncertain for daily professional use.

Use Cases
  • Developers adjusting per-app audio during video calls
  • Designers snapping windows and monitoring system resources
  • Users uninstalling apps without leaving residual files or processes

Source: vorssaint/vorssaint-utils — based on the README and release notes.

Unsloth Enables Local LLM Training on AMD GPUs with 2x Speedup 🔗

Latest release adds ROCm optimization and Triton kernels for Radeon and Instinct hardware.

unslothai/unsloth · Python · 69k stars Est. 2023

Unsloth’s v0.1.

501-beta release brings local LLM training and inference to AMD GPUs across Windows, WSL, and Linux. The update uses custom Triton kernels and math algorithms to train 500+ models up to 2× faster with 70% less VRAM, including support for RDNA2, RDNA4, and Strix Halo architectures. Users can install via curl -fsSL https://unsloth.ai/install.sh | sh on Linux/macOS/WSL or irm https://unsloth.ai/install.ps1 | iex on Windows. Features include GGUF export, tool calling, code execution, and OpenAI-compatible API endpoints for local model serving. The project supports model comparison, web/PDF search, and connecting local models to agents like Claude Code and Codex.
The catch: Despite rapid development, 986 open issues suggest ongoing stability and usability challenges, particularly around complex multi-GPU setups and Windows-specific driver compatibility.

Use Cases
  • Fine-tune LLMs locally on AMD Radeon GPUs
  • Deploy local models via OpenAI-compatible API endpoints
  • Train reinforcement learning models with reduced VRAM usage

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

Stealth-Patched Firefox Makes Playwright Undetectable to Anti-Bot Systems 🔗

Fingerprint set in engine, not page, with humanized input for full bot evasion

feder-cr/invisible_playwright · Python · 1.8k stars 2mo old

The invisible_playwright project patches Firefox at the C++ level to embed realistic browser fingerprints directly into the engine, avoiding JavaScript shims that anti-bot systems can detect. It passes standard detection tests by answering "yes" to both "Is this a real browser?

" and "Is a real person using it?" through authentic GPU/WebGL, Canvas, font, and WebRTC signals, plus humanized mouse paths with pressure, timing, and trusted events. Driven by standard Playwright API, it requires no code changes to existing scripts. The catch: While the browser layer is stealth, the project notes that 90% of proxies are public and IP reputation remains the primary blocker — undetected automation fails if the IP is on deny lists.

Use Cases
  • Scrape e-commerce sites without triggering CAPTCHAs
  • Automate login flows on bot-protected SaaS platforms
  • Test anti-fraud systems with realistic user behavior simulation

Source: feder-cr/invisible_playwright — based on the README and release notes.

Open Source AI Agents Shift From Assistants to Autonomous Workflows 🔗

Projects now orchestrate multi-agent systems for coding, trading, and media without constant human prompting

The open source AI agent landscape is rapidly evolving beyond simple chatbots into persistent, self-directed systems capable of executing complex, multi-step workflows. Repositories like makecindy/cindy and gptme/gptme offer terminal-based agents that write code, browse the web, and use local tools autonomously.

Others, such as hahhforest/pi-textbook and can1357/oh-my-pi, provide structured pathways to build Pi-style agents from scratch, emphasizing reproducibility and skill composition. Multi-agent orchestration is gaining traction: stablyai/orca manages fleets of parallel coding agents, while agent-of-empires/agent-of-empires unifies access to Claude Code, Codex, and Gemini agents via TUI or web interfaces. Specialized agents are emerging for niche domains—HKUDS/Vibe-Trading for algorithmic trading, calesthio/OpenMontage for agentic video production with 500+ skills, and yanhua1010/self-media-content-workflow for automated content creation. Skill marketplaces and agent harnesses like zhinjs/zhin and 1jehuang/jcode enable hot-reloading, multi-channel integration, and secure tool orchestration. This shift reflects a move from AI as a reactive assistant to AI as a proactive, modular workforce—where agents don’t just respond but plan, delegate, and persist across sessions.
The catch: Despite rapid innovation, the ecosystem remains fragmented across languages and frameworks, with limited interoperability between agent systems; many projects prioritize demo-friendly autonomy over robust error handling, long-term state management, or security hardening, raising concerns about reliability in production use beyond controlled experiments.

Use Cases
  • Developers delegate coding tasks to terminal-based AI agents
  • Traders run multi-agent adversarial analysis frameworks
  • Creators automate video production using agent skill pipelines

AI Agent Skills Marketplace Emerges as Open Source Standard 🔗

Modular, composable tools turn LLMs into programmable workers across coding, research, and automation

A defining pattern in open source is the rise of LLM agent skill marketplaces—reusable, plug-and-play modules that extend AI agents with domain-specific capabilities. Projects like gptme/gptme and nanmicoder/cc-haha position the terminal or desktop as an agent workspace, where skills are dynamically loaded.

finna/Finn-loop exemplifies this with its 3-skill factory (spec, build, review) for Claude Code, while different-ai/openwork offers an open alternative to Claude Cowork. Specialized skills are proliferating: AgriciDaniel/claude-seo delivers 25 sub-skills for search optimization, Imbad0202/academic-research-skills automates scholarly workflows, and coreyhaines31/marketingskills equips agents with CRO and SEO tooling. Even niche domains are covered—xbtlin/ai-berkshire implements value investing frameworks, and ZhuLinsen/daily_stock_analysis enables automated market monitoring. The technical shift is clear: agents are no longer monolithic prompts but composable systems where skills like bradautomates/claude-video (video understanding) or virgiliojr94/book-to-skill (PDF-to-skill conversion) act as interchangeable components. This mirrors app stores but for AI behavior—skills are versioned, shared, and combined via manifest files or skill registries, enabling agents to adapt to new tasks without retraining.
The catch: Despite rapid growth, skill interoperability remains fragmented—many tools target specific agents (Claude Code, Codex) with incompatible formats, quality varies widely, and long-term maintenance of skill suites is unproven at scale, risking a tower of Babel in agent extensibility.

Use Cases
  • Developers automate code review using deterministic pipelines + LLM agents
  • Researchers build persistent agents that browse, analyze, and cite sources
  • Marketers deploy AI agents for SEO audits and campaign optimization

AI Agents Merge with Dev Tools in Open Source Surge 🔗

Autonomous agents now embed in terminals, browsers, and workflows, blurring lines between coding, automation, and AI

Open source is rapidly integrating AI agents directly into developer toolchains, moving beyond chatbots to embedded, action-oriented systems. Projects like gptme put persistent agents in the terminal that write code, run commands, and browse the web autonomously.

Similarly, agent-of-empires manages multiple AI coding agents (Claude, OpenCode, Gemini) via TUI or web for mobile access, while oomol-lab/open-connector acts as an auth gateway linking 1,000+ SaaS platforms to AI agents through SDKs, CLI, and MCP. Beyond terminals, AI is penetrating UI layers: nexu-io/codex-slides enables AI-generated slide decks inside Codex, turning research into presentations in minutes, and agegr/pi-web offers a web UI for the Pi coding agent. Even automation tools are gaining stealth — feder-cr/invisible_playwright patches Firefox at the engine level for undetected web automation. This pattern reflects a shift from AI as a copilot to AI as an embedded operator within dev environments, handling end-to-end tasks with minimal human intervention. The catch: Many of these agents remain brittle, prone to hallucinations in complex workflows, and lack standardized interfaces, making interoperability and reliability significant hurdles for production use despite rapid prototyping appeal.

Use Cases
  • Developers automate code generation and terminal tasks
  • Teams build autonomous AI slide decks from research
  • Engineers orchestrate multi-agent systems across SaaS tools

Quick Hits

remotion Remotion lets developers build dynamic, data-driven videos using React components, enabling programmatic video creation with full code control. 54.4k
OptMem OptMem equips AI agents with persistent memory via a minimal 426-token prompt and plug-and-play script for long-term context retention. 490
leaked-llm-prompts Leaked-LLM-Prompts aggregates publicly exposed AI system prompts, offering insight into real-world model behaviors and guardrails for security analysis. 274
netmon Netmon provides self-hosted network monitoring with automated speed tests, ARP-based device discovery, and humor-infused AI reports sent to Telegram or Discord. 229
adelan-support Adelan-Support delivers a lightweight, extensible Python framework for building AI-powered customer support agents with tool use and conversation memory. 410
undo Undo reverses the filesystem impact of the last shell command by tracking and inverting file mutations for safe, targeted rollback. 201
Beyond GitHub

The AI Wire

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

From the labs & arXiv

Julia v1.12.6 Refines HPC Workflows with AI Disclosure Rules 🔗

Latest patch enforces transparency for AI-assisted contributions in technical computing language

JuliaLang/julia · Julia · 49k stars Est. 2011 · Latest: v1.12.6

The JuliaLang/julia repository released v1.12.6 on July 27, 2026, a point update focused on contributor guidelines rather than core language changes.

This release introduces a mandatory disclosure rule: any pull request containing substantial contributions from generative AI tools must now include explicit details, with contributors required to review all changes before submission. The policy applies to issues, discussions, and comments as well, aiming to maintain code integrity in a language widely used for high-performance computing, machine learning, and scientific computing. Julia remains a dynamic, high-performance language designed for technical computing, offering Python-like ease with C-like speed through just-in-time compilation and multiple dispatch. Installation is streamlined via juliaup, which manages multiple versions and provides the latest stable release. The project continues to emphasize community contributions, welcoming bug fixes, documentation, tests, and performance improvements from developers of all levels. Despite its maturity—over 15 years of active development—the repository shows sustained engagement, with the last commit just zero days ago and 4,717 open issues indicating ongoing evolution.
The catch: Julia’s rapid release cadence and evolving AI disclosure requirements may increase cognitive load for contributors needing to track policy shifts alongside technical updates, potentially slowing adoption in regulated or compliance-sensitive environments.

Use Cases
  • Scientists simulate climate models
  • Engineers optimize fluid dynamics
  • Researchers train ML prototypes

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

More Stories

Caveman Skill Trims AI Agent Output by 65% Without Losing Accuracy 🔗

v1.9.1 release standardizes savings metric and improves installer reliability across agents

JuliusBrussee/caveman · JavaScript · 93.4k stars 3mo old

JuliusBrussee/caveman is a Claude Code skill that reduces AI agent output tokens by an average 65% through caveman-style speech while preserving technical accuracy. The v1.

9.1 release, tagged 1 day ago, hardens installer integrity and removes optimistic savings ranges, now reporting a flat 65% reduction based on a 10-prompt benchmark. It works by wrapping responses in tight, grammatically minimal phrasing — e.g., “New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo.” — without altering code, commands, or error output. Supported across 30+ agents including Cursor, Windsurf, and Copilot via a single install script.
The catch: Open issues exceed 400, suggesting ongoing integration challenges despite recent hardening efforts.

Use Cases
  • Developers cutting API costs in high-volume AI coding sessions
  • Teams standardizing terse agent output for log analysis
  • Users hitting token limits on constrained agent subscriptions

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

rasbt/LLMs-from-scratch Adds GPT-3.5 Weight Loading for Finetuning 🔗

New feature lets builders adapt pretrained models using PyTorch notebooks without full retraining.

rasbt/LLMs-from-scratch · Jupyter Notebook · 99.9k stars Est. 2023

The rasbt/LLMs-from-scratch repository now includes code to load weights from larger pretrained models like GPT-3.5 for instruction tuning, expanding beyond its original from-scratch training focus.

Built in Jupyter Notebooks with PyTorch, it walks through tokenization, attention mechanisms, and training loops step by step. Recent commits show active maintenance, with the last push days ago and only two open issues. Builders can clone the repo and finetune LLMs on custom datasets using familiar Python tools.
The catch: Finetuning relies on external model weights, which may violate licensing terms of some pretrained checkpoints.

Use Cases
  • ML engineers learn LLM architecture by coding transformers
  • Researchers adapt public models for domain-specific tasks
  • Students implement pretraining pipelines from scratch in PyTorch

Source: rasbt/LLMs-from-scratch — based on the project README.

TensorFlow 2.21 Drops Python 3.9 and TensorBoard Support 🔗

Release adds int2/4 quantization and JPEG XL decoding while removing legacy dependencies

tensorflow/tensorflow · C++ · 196.6k stars Est. 2015

TensorFlow 2.21.

0 removes support for Python 3.9 and decouples TensorBoard as a standalone dependency, streamlining the core framework. The update enhances tf.lite with int2, int4, and int16x8 quantization operators for efficient edge deployment and adds JPEG XL decoding in tf.image. Developers can now identify unspecified tensor shapes via NoneTensorSpec in tf.data. Despite these advances, the project maintains 2, **maintains a steep learning curve for newcomers due to its vast API surface and inconsistent documentation across versions.

Previously in The Times “covered” — Jul 21

Use Cases
  • Researchers training vision models with JPEG XL inputs
  • Developers deploying int4-quantized models on microcontrollers
  • Data engineers validating pipelines with optional tensor specs

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

Quick Hits

FinGPT FinGPT provides open-source financial LLMs trained for real-world trading and analysis, enabling developers to build AI-driven investment tools with HuggingFace accessibility. 21k
AutoGPT AutoGPT empowers users to create autonomous AI agents that execute complex tasks independently, focusing on goal-driven automation without constant human oversight. 185.7k
hermes-agent Hermes-Agent evolves with user interaction, adapting its behavior and knowledge over time to deliver increasingly personalized and context-aware AI assistance. 221.2k
openai-cookbook OpenAI Cookbook offers practical, ready-to-run Jupyter examples that demystify API usage, helping builders integrate GPT models efficiently into real applications. 74.9k
pytorch PyTorch delivers flexible,charged GPU computing and research and GPU-accelerated tensor computations and dynamic neural networks, forming the backbone of modern AI research and deployment. 102k

iDynTree v15.1.0 Fixes Python CI and URDF Handling for Floating-Base Robot Control 🔗

The multibody dynamics library updates bindings and model parsing to support humanoid and free-floating robot development workflows.

gbionics/idyntree · C++ · 234 stars Est. 2014 · Latest: v15.1.0

iDynTree released v15.1.0 with a focus on stabilizing its Python bindings and improving URDF file handling, critical for robot model integration.

The update resolves continuous integration failures in the Python interface, ensuring reliable access to dynamics algorithms for control and simulation tasks. By fixing format handling in URDF and SDFormat parsers, the library maintains its core strength: enabling dynamic base link changes without model reloads—a key requirement for humanoids and free-floating robots where the computational base shifts during motion. Built on an undirected graph structure (iDynTree::Model), the C++ library supports mixed, body, and inertial representations of 6D quantities, accommodating various whole-body control formulations. It also retains torque estimation capabilities inspired by the iCub humanoid, allowing joint torque inference without direct sensors. These features serve researchers and engineers developing locomotion, manipulation, and estimation pipelines for legged and floating-base systems.

The catch: Despite recent activity, the project carries 197 open issues, suggesting ongoing maintenance challenges that could affect stability in production environments relying on cutting-edge features or rapid bug resolution.

Use Cases
  • Control humanoid robots with dynamic base link switching
  • Estimate joint torques without mounted sensors
  • Parse and modify URDF models for simulation tools

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

More Stories

py-xiaozhi Adds WebRTC AEC and Parallel Mixing in v2.1.1 🔗

Update enhances real-time voice AI latency and audio processing for edge devices.

huangjunsen0406/py-xiaozhi · Python · 3.4k stars Est. 2025

The py-xiaozhi project released v2.1.

1, integrating built-in WebRTC AEC and parallel audio mixing to reduce voice interaction latency under 20ms. Built on Python’s async architecture, it supports multimodal workflows across Windows, macOS, Linux, and ARM platforms like Raspberry Pi and Jetson Nano. The framework enables real-time voice streaming, vision-language tasks, and IoT control via MCP integrations. Sponsored by GitDo.net and Token能量站, it evolved from xiaozhi-esp32 and is now an upstream dependency for D-Robotics’ RDK deployments. The catch: Despite active development, the project has only one open issue and limited public documentation on scaling voice pipelines in multi-device IoT fleets.

Use Cases
  • Developers building voice-controlled robots on ESP32-S3
  • Engineers deploying multimodal AI agents on Raspberry Pi
  • Teams creating cross-platform IoT assistants with local LLMs

Source: huangjunsen0406/py-xiaozhi — based on the README and release notes.

GTSAM Shifts to C++17, Drops Boost for Factor Graphs 🔗

Pre-4.3 develop branch signals API breaks; stable 4.2.2 remains for legacy support

borglab/gtsam · Jupyter Notebook · 3.6k stars Est. 2017

The borglab/gtsam project advances its develop branch toward C++17, removing Boost dependencies and introducing API-breaking changes ahead of version 4.3.

This shift affects smoothing and mapping workflows in robotics and vision, where factor graphs replace sparse matrices for state estimation. While the latest release, 4.2.2, maintains stability, users relying on deprecated features must disable GTSAM_ALLOW_DEPRECATED_SINCE_V42 or stick to the maintained 4.2 line. The project continues to offer C++, Python, and MATLAB wrappers, with recent fixes improving ISAM2 robustness and Python packaging for versions 3.11–3.14.
The catch: Active development prioritizes future compatibility over backward stability, forcing early adopters to refactor code or risk instability in the pre-release branch.

Use Cases
  • Robotics engineers fusing lidar and IMU data for SLAM
  • Vision researchers optimizing bundle adjustment in SfM pipelines
  • Autonomous vehicle teams estimating pose from multimodal sensor streams

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

Zenoh 1.9.0 Tightens QUIC Reliability for Geo-Distributed Edge Systems 🔗

Mixed-reliability UDP/QUIC links now prioritize streams while fixing 500ms scouting delay regression.

eclipse-zenoh/zenoh · Rust · 3k stars Est. 2020

Eclipse Zenoh’s 1.9.

0 release refines its geo-distributed pub/sub/storage stack with QUIC enhancements: mixed-reliability support and priority-based stream handling improve UDP link robustness. The release patches a 500ms scouting delay regression from 1.7.2 and addresses RuntimeState leaks, boosting stability for embedded and ROS2 deployments. Built in Rust, Zenoh unifies data-in-motion, -at-rest, and computations but demands expertise in distributed systems tuning.
The catch: Open issues (261) and complex configuration raise adoption barriers for teams lacking low-latency networking experience.

Use Cases
  • Industrial IoT sensors syncing data across factory edge nodes
  • ROS2 robotics fleets sharing real-time lidar and telemetry
  • Geo-distributed storage clusters querying replicated datasets with low latency

Source: eclipse-zenoh/zenoh — based on the README and release notes.

Quick Hits

yarp YARP provides a flexible, cross-platform middleware for robot development, enabling seamless communication and integration across diverse hardware and software components. 598
horus HORUS delivers a high-performance, real-time robotics runtime in Rust, offering low-latency control and safety-critical reliability for next-gen robotic systems. 393
tesseract Tesseract offers a powerful motion planning environment for complex robotic manipulation, supporting collision checking, trajectory optimization, and multi-robot coordination in C++. 376
openarm OpenArm is a fully open-source humanoid arm designed for physical AI research, enabling robust interaction in contact-rich, real-world tasks like assembly and tool use. 2.8k
ardupilot ArduPilot is a mature, feature-rich open-source autopilot suite supporting planes, copters, rovers, and subs with advanced navigation, stability, and mission planning capabilities. 15.6k

Nuclei v3.11.0 Enforces Digital Signatures for JavaScript Templates 🔗

Breaking change blocks unsigned JS templates to harden scanner against runtime exploits

projectdiscovery/nuclei · Go · 30k stars Est. 2020 · Latest: v3.11.0

Nuclei’s latest release, v3.11.0, introduces a mandatory digital signature requirement for any custom template using the javascript: protocol.

Unsigned JavaScript templates are now skipped during loading, a move designed to close a growing attack surface. The change extends sandboxing and code-template protections first seen in v3.10.0, addressing risks from Go-backed modules exposed through Nuclei’s JS runtime. Official templates from the nuclei-templates repository remain unaffected—they are pre-signed with ProjectDiscovery’s public key. Users maintaining private or community JavaScript templates must now sign them before use; workflows referencing unsigned versions will fail silently. The scanner retains its ultra-fast parallel processing, YAML-based DSL, and multi-protocol support across HTTP, DNS, TCP, and cloud configs, with CI/CD integrations for Jira, Splunk, and GitLab. Nuclei continues to serve as a standalone CLI tool, with the project warning that running it as a service may introduce security risks.
The catch: The signing requirement adds complexity for teams managing custom JavaScript-based detection logic, potentially delaying adoption of new template features until signing pipelines are established.

Previously in The Times “covered” — Jul 26

Use Cases
  • Security teams scanning internal APIs for misconfigurations
  • DevOps engineers integrating vulnerability checks into CI pipelines
  • Bug bounty hunters validating subdomain takeover risks at scale

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

More Stories

AI-guided hackingtool adds local model support for offline security testing 🔗

Users can now run the toolkit entirely offline using self-hosted LLMs

Z4nzu/hackingtool · Python · 78.5k stars Est. 2020

The Z4nzu/hackingtool project updated its AI layer to support local model inference, allowing penetration testers and red teamers to operate without external API keys or internet connectivity. This enhancement strengthens its air-gapped usability for sensitive engagements while maintaining the same guided workflow: describe a goal in plain English, receive the exact command, and execute step-by-step with tool recommendations.

The toolkit still bundles 215 curated security tools across 21 categories, including recon, web, wireless, and post-exploitation, all accessible via /goal and /find commands.
The catch: Despite recent updates, the project has 127 open issues and relies on a complex dependency tree that may delay fixes for critical bugs in niche tool integrations.

Use Cases
  • Penetration testers running offline web app assessments
  • Red teamers simulating wireless attacks in isolated labs
  • DFIR analysts automating log analysis and timeline reconstruction

Source: Z4nzu/hackingtool — based on the project README.

Web-Check adds QRATOR and DDoS-Guard WAF detection for deeper OSINT 🔗

Latest update enhances firewall analysis in the TypeScript-based website scanner

lissy93/web-check · TypeScript · 34.3k stars Est. 2023

The lissy93/web-check project recently updated its firewall.js module to detect QRATOR and DDoS-Guard web application firewalls, expanding its OSINT capabilities.

Built in TypeScript, the tool scans domains for IP info, SSL chains, DNS records, headers, cookies, trackers, and server architecture to uncover potential attack vectors. It supports deployment via Netlify, Vercel, Docker, or source, with a live demo at web-check.as93.net. The catch: With 84 open issues and a release cycle focused on incremental feature adds, questions remain about long-term maintenance and scalability for enterprise use.

Use Cases
  • Security analysts audit client websites for exposed misconfigurations
  • Developers verify third-party service integrations and tracking scripts
  • Sysadmins benchmark server headers and DNS security extensions post-deploy

Source: lissy93/web-check — based on the README and release notes.

CarterPerez Dev Updates Cybersecurity Project Repository with New DDoS Tool 🔗

Repository now includes 70 hands-on projects across four skill tiers and certification roadmaps

CarterPerez-dev/Cybersecurity-Projects · Go · 4.6k stars 8mo old

The CarterPerez-dev/Cybersecurity-Projects repository added a DDoS Mitigation Tool in its latest commit, expanding its collection of 70 cybersecurity projects. Built in Go and Python, the resource organizes projects into Foundations, Beginner, Intermediate, and Advanced tiers, each with guided learning paths.

It also features 10 certification roadmaps for roles like SOC Analyst and Pentester, alongside learning resources covering tools, courses, and frameworks. The Foundations tier uses single-file Python scripts with heavy inline comments and Numpy-style docstrings to teach concepts from zero. Last updated 26 days ago, the project shows active maintenance with two open issues and 674 forks.
The catch: The repository relies heavily on external guided courses from CertGames for full walkthroughs, which may limit accessibility for users seeking entirely self-contained, free-to-use project instructions without third-party dependencies.

Use Cases
  • Learn cybersecurity fundamentals through commented Python scripts
  • Build a DDoS mitigation tool using Go and Python source code
  • Follow certification roadmaps for SOC Analyst or Pentester roles
  • Study security concepts via structured learning resources and courses

Source: CarterPerez-dev/Cybersecurity-Projects — based on the project README.

Quick Hits

Azure-Sentinel Azure/Azure-Sentinel delivers cloud-native SIEM with AI-driven threat detection and automated response across your entire enterprise infrastructure. 6k
Decepticon PurpleAILAB/Decepticon is an autonomous hacking agent that simulates advanced adversary tactics for realistic red team operations. 4.9k
NetExec Pennyw0rth/NetExec is a versatile network execution tool enabling post-exploitation, credential spraying, and lateral movement across protocols. 5.7k
ProxmoxVE community-scripts/ProxmoxVE provides shell-based helper scripts to automate deployment, backup, and management tasks in Proxmox VE environments. 29.1k
sherlock sherlock-project/sherlock hunts down social media accounts by username across hundreds of platforms to identify digital footprints and impersonation risks. 87.1k

Bun v1.3.14 Sharpens JavaScript Toolchain with Faster Startup and Rust Core 🔗

The all-in-one runtime now ships TypeScript and JSX support out of the box, reducing dev setup friction.

oven-sh/bun · Rust · 95k stars Est. 2021 · Latest: bun-v1.3.14

Bun v1.3.14 sharpens the JavaScript toolchain by combining a fast runtime, bundler, test runner, and package manager into a single executable.

Built in Rust and powered by JavaScriptCore, it delivers significantly reduced startup times and memory usage compared to traditional Node.js setups. Developers can run TypeScript and JSX files directly with bun run index.tsx without additional transpilation steps. The release includes improved install scripts for Linux, macOS, and Windows, with specific guidance for older kernels and CPU compatibility.

Bun’s integrated toolchain replaces multiple dev dependencies—testing via bun test, script execution with bun run start, and package management through bun install and bunx. Its Node.js-compatible package manager aims to simplify workflows while avoiding the bloat of large node_modules directories. The project maintains a rapid release cadence, with canary builds available on every commit to main.

The catch: Bun’s reliance on JavaScriptCore instead of V8 may cause subtle behavioral differences in edge-case JavaScript execution, requiring testing when migrating complex Node.js applications.

Previously in The Times “covered” — Jul 26

Use Cases
  • Backend devs replacing Node.js with faster startup
  • Full-stack teams streamlining TS/JSX tooling
  • CLI tool authors reducing dependency footprint

Source: oven-sh/bun — based on the README and release notes.

More Stories

Ollama v0.32.4 boosts Apple GPU performance with MLX engine support 🔗

Quantization fixes and speculative decoding upgrades improve Qwen3 and draft model handling

ollama/ollama · Go · 177k stars Est. 2023

Ollama’s latest release adds Laguna support on Apple GPUs via the MLX engine, accelerating inference on recent Mac hardware. It now quantizes draft-model output heads during speculative decoding, improving efficiency.

A fix for Qwen3 MoE decoding handles differently quantized experts, with gate/up projection speedups of 4–9% on M5 Max chips. The update streamlines local LLM serving for developers using llama.cpp backends.
The catch: Open issues remain high at 3,533, signaling ongoing stability and compatibility challenges across model variants and platforms.

Previously in The Times “covered” — Jul 21

Use Cases
  • Run Gemma 4 locally via `ollama run gemma4`
  • Integrate with Claude Code using `ollama launch claude`
  • Serve models through REST API at localhost:11434/api/chat

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

Rust Compiler Fixes LLVM Optimization Bug in Latest Release 🔗

Patch 1.97.1 resolves miscompilation risk from recent LLVM update

rust-lang/rust · Rust · 114.9k stars Est. 2010

The Rust project released version 1.97.

1 to address a miscompilation bug triggered by an LLVM optimization. The fix reverts a rustc change and updates the LLVM submodule to prevent incorrect code generation. This patch ensures reliability for systems relying on Rust’s safety guarantees. Despite rapid fixes, the compiler’s deep LLVM integration means updates can introduce subtle regressions requiring careful backporting. The catch: Rust’s release process depends on frequent LLVM rebases, which can delay stability for embedded or long-term support targets.

Use Cases
  • Building performance-critical backend services
  • Developing memory-safe embedded firmware
  • Creating cross-platform CLI tools with Cargo

Source: rust-lang/rust — based on the README and release notes.

Bitcoin Core v31.1 release improves network security and validation 🔗

Latest update strengthens peer-to-peer block verification and wallet resilience

bitcoin/bitcoin · C++ · 89.7k stars Est. 2010

Bitcoin Core version 31.1, released July 2026, enhances block validation logic and peer-to-peer message handling to reduce eclipse attack risks.

The update refines transaction relay policies and strengthens mempool protection against low-fee spam. Developers note improved test coverage in wallet encryption modules, though full GUI testing remains dependent on the separate bitcoin-core/gui repo. The catch: Over 675 open issues persist, highlighting ongoing strain on review bandwidth despite frequent commits.

Use Cases
  • Developers building full-node blockchain infrastructure
  • Enterprises validating Bitcoin transactions with cryptographic guarantees
  • Researchers testing peer-to-peer network resilience under adversarial conditions

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

Quick Hits

native A Zig toolkit enabling developers to build native desktop applications with low-level control and high performance. 7k
FFmpeg A comprehensive multimedia framework for recording, converting, and streaming audio and video across platforms. 62.4k
zed A blazing-fast, collaborative code editor built in Rust, designed for real-time multiplayer editing and deep IDE-like responsiveness. 87.6k
RuView Transforms everyday WiFi signals into real-time spatial awareness, vital signs, and presence detection — no cameras needed. 86.8k
cc-switch A unified cross-platform desktop assistant integrating Claude Code, Codex, OpenCode, and more into a single AI workflow hub. 121.6k

TuyaOpen v1.9.0 cuts battery drain with new DTIM WiFi low-power mode 🔗

Adds support for ZECTRIX_T5AI_NOTE_4 and RK3576 boards, boosting edge AI hardware options

tuya/TuyaOpen · C · 1.8k stars Est. 2023 · Latest: v1.9.0

TuyaOpen’s latest release targets battery-powered IoT devices with WiFi ultra-low-power DTIM mode, reducing standby consumption via the new ultra_lowpower_demo example. The update expands hardware support to include the ZECTRIX_T5AI_NOTE_4 and SPARKLEIOT_T5_DEV T5AI boards, plus the LCKFB T5AI board with GC0308 camera and SC7A20 IMU integration. On Linux, RK3576 board support now enables direct deployment of the your_chat_bot application, resolving prior memory overflow bugs in your_robot_dog.

The framework continues to unify C/C++ SDK access across Tuya T-Series, ESP32, and Raspberry Pi hardware, pairing with Tuya Cloud’s multimodal AI for ASR, KWS, TTS, and LLM integration (ChatGPT, Gemini, Deepseek). Developers retain OTA updates, remote monitoring, and compatibility with Google Home and Alexa.
The catch: Despite active commits, 57 open issues persist, and the C-only SDK may deter teams preferring higher-level abstractions or Rust-based toolchains for AIoT prototyping.

Previously in The Times “covered” — Jul 25

Use Cases
  • Build battery-operated voice sensors with extended lifespan
  • Deploy vision-enabled AI gateways on RK3576 Linux boards
  • Create Tuya-certified smart devices with local LLM inference

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

More Stories

Open-source 6-DOF CAD controller hits v1.0 with 3D-printable parts 🔗

Built around Arduino Pro Micro and Hall-effect joysticks for native HID navigation

HackMan3D/HackMan3D-Orbit-Controller · C++ · 170 stars 3w old

HackMan3D Orbit Controller v1.0.

0 delivers a fully 3D-printable six-degree-of-freedom input device for CAD and slicer workflows. Using four Hall-effect joysticks and an Arduino Pro Micro, it emulates 3Dconnexion-style HID controllers without requiring intermediate software. Firmware supports configurable speed profiles, dead zones, and axis inversion, with optional mouse-emulation mode for slicers like Cura and PrusaSlicer. Assembly guides, wiring diagrams, and bill of materials are provided for builds on Creality Cloud or MakerWorld.
The catch: Open issues remain unresolved, and long-term durability of printed parts under frequent use is unproven.

Use Cases
  • Engineers navigate Fusion 360 with pan, zoom, and rotate
  • Makers adjust slicer views in Bambu Studio via mouse emulation
  • Designers configure shortcut buttons for frequent CAD commands

Source: HackMan3D/HackMan3D-Orbit-Controller — based on the README and release notes.

Fanpico adds new board support and WiFi stability fixes 🔗

Latest release v1.7.8 enables FanPico-0200 hardware and improves wireless reconnection logic

tjko/fanpico · C · 228 stars Est. 2022

The tjko/fanpico project released v1.7.

8, adding support for the FanPico-0200 board and refining WiFi rejoin logic for Pico W variants. This open-source firmware lets builders program fan curves via temperature sensors or PWM inputs using an RP2040-based controller. It supports up to 8 fans, multiple temperature probes, and SCPI-like commands for real-time tuning without OS dependencies. Configuration persists in flash, enabling standalone operation in PCs, 3D printers, or network racks. The catch: WiFi functionality requires the Pico W variant, limiting wireless use to specific hardware builds and excluding standard Pico models.

Use Cases
  • Fine-tune PC fan curves using temperature feedback
  • Control 3D printer cooling via custom PWM profiles
  • Monitor rack equipment fans with SNMP and SSH access

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

Quick Hits

PMSG Control-C/PMSG (C++): PMSG is a wearable face computer platform using Seeed Studio XIAO, enabling developers to flash and customize real-time biometric interaction devices. 41
vdbrink.github.io vdbrink/vdbrink.github.io (CSS): This documentation hub offers practical tips and tricks for home automation with Node-RED and Home Assistant, streamlining smart home setup and integration. 48
SonosESP OpenSurface/SonosESP (C): A modern touchscreen controller for Sonos speakers built on ESP32-P4, delivering responsive, local control without reliance on cloud or apps. 43
FlightTracker ColinWaddell/FlightTracker (Python): Turns a Raspberry Pi into a real-time flight tracker using dot matrix display, showing live aircraft, satellite, weather, and time data from ADS-B or FlightRadar24. 182
maker.js microsoft/maker.js (TypeScript): A powerful 2D vector modeling library for CNC and laser cutters, enabling parametric design, path generation, and precise shape manipulation in code. 2k
espectre ESPectre - Motion detection system based on Wi-Fi spectre analysis (CSI), with Home Assistant integration. 8.9k

PlayCanvas Engine Adds WebGPU Shadow Cascades and Gaussian Splatting Support 🔗

Latest release fixes normal mapping bugs and texture limits while expanding 3D rendering capabilities for browser-based games and visualizations.

playcanvas/engine · JavaScript · 16.3k stars Est. 2014 · Latest: v2.21.1

The PlayCanvas engine released v2.21.1 with targeted fixes for WebGPU rendering, including proper shadow cascade masking and corrected normal mapping via derivative TBN.

Developers gain first-class support for loading and rendering 3D Gaussian Splats (.ply) directly in the engine, a technique gaining traction for real-time photorealistic scenes. The update also resolves an MSDF font rendering bug affecting small text sizes was patched, and WebGPU sampled texture limits were raised to prevent runtime errors in complex scenes. Installation remains straightforward via npm install playcanvas or project scaffolding with create-playcanvas. The engine continues to abstract WebGL2 and WebGPU differences, letting developers write once and deploy across desktop and mobile browsers.
The catch: While WebGPU support is advancing, fallback to WebGL2 remains necessary for broad browser compatibility, and Gaussian Splatting performance on mobile devices is not yet optimized for complex scenes.

Previously in The Times “covered” — Jul 23

Use Cases
  • Build browser-based 3D games with real-time lighting
  • Render photogrammetry captures using Gaussian Splats
  • Create cross-platform WebXR visualizations for VR headsets

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

More Stories

JoltPhysics v5.6.0 Adds GPU Hair Simulation and Compute Shader Support 🔗

New release enables strand-based hair physics and DX12/Vulkan/Metal compute offloading for games and VR.

jrouwe/JoltPhysics · C++ · 11.1k stars Est. 2021

JoltPhysics v5.6.

0 introduces a GPU-accelerated strand-based hair simulation using Cosserat rod physics, supporting environment collisions and skinning to scalp meshes. The release also adds interfaces to run compute shaders on DX12, Vulkan, and Metal, enabling parallel collision queries and background physics batching without stalling the main thread. These features build on Jolt’s core strength: deterministic, multicore-friendly rigid body simulation used in titles like Horizon Forbidden West. The library continues to prioritize lock-free background loading and immediate visibility of cross-thread changes.
The catch: GPU hair simulation remains marked as work in progress, with known limitations in collision shape support and unresolved items listed in Hair.h.

Use Cases
  • Game developers simulating realistic character hair in VR
  • Engine teams offloading collision queries to GPU compute
  • Studios loading/unloading physics bodies without hitching the main thread

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

Pumpkin Minecraft Server Gains Rust-Based Bedrock Support 🔗

Developers report progress on Java and Bedrock edition compatibility with performance focus

Pumpkin-MC/Pumpkin · Rust · 10.2k stars Est. 2024

Pumpkin, the Rust-built Minecraft server, has advanced chunk saving systems, moving parts of Minecraft Bedrock Edition parity. The project leverages Rust's concurrency for low-latency world ticking and entity handling, targeting server admins seeking alternatives to Paper or Velocity.

With 680 forks and active commits, it shows growing community interest despite 283 open issues. The server supports TOML configuration, RCON, and plugin extensibility via inventories and chat systems.
The catch: Bedrock Edition remains a work-in-progress with mobs, animals, and villager AI still unimplemented, limiting production use for cross-platform networks.

Previously in The Times “covered” — Jul 20

Use Cases
  • Host Java Edition servers with Rust performance
  • Test Bedrock Edition features in development
  • Run lightweight servers with TOML config

Source: Pumpkin-MC/Pumpkin — based on the project README.

LÖVE Framework Updates Android Launcher and iOS Game Sorting 🔗

Release 11.5 fixes LuaJIT bugs and improves macOS Apple Silicon performance

love2d/love · C++ · 8.6k stars Est. 2019

The LÖVE 2D game framework released version 11.5, adding an Android launcher for easier .

love file loading and alphabetically sorting iOS game selectors. It updated LuaJIT to fix inconsistent 'pairs' behavior and disabled JIT by default on macOS arm64 due to unreliable memory. The release also resolved Android audio pausing issues and iOS recording crashes. Developers use LÖVE to build cross-platform 2D games in Lua with love2d.org/wiki guidance.
The catch: Open issues remain at 120, indicating ongoing maintenance challenges despite recent fixes.

Use Cases
  • Indie developers creating 2D mobile games
  • Hobbyists prototyping Lua-based game mechanics
  • Educators teaching game programming with Lua

Source: love2d/love — based on the README and release notes.

Quick Hits

Fyrox FyroxEngine/Fyrox (Rust): A high-performance, Rust-based 3D and 2D game engine offering real-time rendering, physics, and editor tools for building cross-platform games with memory safety and zero-cost abstractions. 9.5k
internal MustardOS/internal (GLSL): The core rendering and graphics subsystem of MustardOS, featuring low-level GLSL shaders for custom visual effects and hardware-accelerated display pipelines in a minimalist OS environment. 106
photon sixthsurge/photon (GLSL): A visually striking, gameplay-enhancing shader pack for Minecraft that improves lighting, shadows, and atmospheric effects without sacrificing performance or breaking vanilla mechanics. 1.8k
godot godotengine/godot (C++): A free, open-source, multi-platform 2D and 3D game engine with a user-friendly editor, scene system, and GDScript scripting — ideal for rapid prototyping and full-scale game development. 114.7k
SpacetimeDB clockworklabs/SpacetimeDB (Rust): A real-time, distributed database built in Rust that enables sub-millisecond reads/writes and automatic synchronization across nodes for low-latency, scalable applications. 24.8k
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