Design
Preset
Background
Text
Font
Size
Width
Account Pricing

The Git Times

AI Models
Claude Opus 5 $25/M GPT-5.6 Luna $1.20/M Gemini 3.1 Pro Preview $12/M Grok 4.6 $6/M DeepSeek V4 Pro 0813 $3.37/M Qwen3.8 2.4T A95B $6/M Kimi K3 $15/M
Full Markets →

GamePhanes Benchmarks Coding Agents Through Full Game Build Loops 🔗

Evaluates agents on runtime behavior, debugging, and regression resistance in Godot environments

GamePhanes/GamePhanes · JavaScript · ▲ 105 in 1d 5d old

GamePhanes introduces Terminal-Bench, a benchmark that tests coding agents not by file edits or command outputs, but by their ability to build, debug, and repair playable Godot games through terminal workspaces. Agents receive evaluator-controlled probes and must produce executable runtime feedback—importing projects, accepting input, modifying state, and delivering intended behavior.

The system judges engineering work: code changes, debugging efficacy, and resistance to regressions, not just task completion. Built on the Harbor task contract, GamePhanes provides six playable reference projects and plans 20 tasks across gameplay, engine/runtime, UI/interaction, content/systems, and delivery/quality. Each task is a self-contained Harbor directory with versioned calibration runs completed successfully using Oracle 1.0 and NOP 0.0 baselines. Early trials like Kimi K3 show agents can make substantive changes over 30 turns but still fail hidden boundary assertions, revealing gaps in robust, long-horizon reasoning. The benchmark shifts focus from syntactic correctness to functional, interactive outcomes—aligning evaluation with real-world software engineering demands.
The catch: The benchmark remains narrow in scope, targeting only Godot and JavaScript-based agents, with production tasks still unverified and no public leaderboard yet established.

Why this leads today GamePhanes provides a practical, open-source environment for testing and benchmarking AI agents in Godot, filling a gap in tooling for developers integrating intelligent behaviors into games.

Use Cases
  • Evaluate coding agents on full-game debugging workflows
  • Test agent ability to modify runtime game states
  • Measure regression resistance in interactive software repair

Source: GamePhanes/GamePhanes — based on the project README.

More on the Front Page

Codex Router Unifies External Models in Codex Without Replacing Native Picker 🔗

Routes Kimi, DeepSeek, and other LLMs through a local service that preserves Codex’s original model catalog and default settings.

duolahypercho/codex-router · JavaScript · 2.9k stars 1mo old

Codex Router lets builders use Anthropic, Kimi, DeepSeek, xAI, GitHub Copilot, and future models inside the Codex App and CLI via a single local router. It speaks the Responses API and merges routed models into Codex’s native model picker without overwriting the original GPT catalog.

Credentials are entered once through a hidden terminal prompt and shared across all installed clients—Codex, DeepSeek Harness, and Gemini CLI—eliminating repeated key entry. Model visibility and selection are stored locally in model-picker.json as an explicit allowlist, then republished to every client. The router does not alter Codex’s active task configuration or default model unless explicitly opted in; that choice survives rebuilds and can be restored. Recent updates include surfacing LM Studio models in the panel, adding official Xiaomi MiMo API support, fixing Qwen and GLM context window limits using provider-reported values, and adding UI languages for Arabic, Hindi, Japanese, and Korean. A new doctor feature flags context windows providers have disproved, and subagents can now specify reasoning effort for child models.
The catch: Desktop packages for Windows and Linux are unsigned tester artifacts, and macOS support remains excluded until Developer ID signing and notarization are completed, limiting broad desktop deployment.

Use Cases
  • Developers testing multiple LLMs in Codex without reconfiguring keys
  • Teams standardizing external model access across Codex, Harness, and Gemini CLI
  • Builders preserving native Codex defaults while experimenting with routed models

Source: duolahypercho/codex-router — based on the README and release notes.

Prime Agent adds recursive subagents to persistent coding workflows 🔗

Latest release fixes syntax highlighting and adjusts recursion depth defaults

PrimeIntellect-ai/prime-agent · TypeScript · ▲ 18.5k in 1d 3mo old

PrimeIntellect-ai/prime-agent is a TypeScript-based coding agent that uses a Recursive Language Model to treat context as variables and subagents as function calls inside a persistent REPL. It combines a Python control environment with durable harness state, enabling long-running autonomous tasks to retain context across sessions.

Subagents spawn via rlm(...), skills are executable Python packages, and agents communicate directly without user mediation. The v0.8.1 release fixed multi-line string highlighting in python tool views, raised default RLM recursion depth from 1 to 2, and adjusted ACP prompt resolution to wait for settled subagent work.
The catch: With 92 open issues and a narrow TypeScript/Python focus, production stability and broad language support remain unproven at scale.

Use Cases
  • Developers automate multi-step refactoring across large codebases
  • Researchers run unattended literature review and synthesis pipelines
  • Teams deploy background agents for continuous integration monitoring

Source: PrimeIntellect-ai/prime-agent — based on the README and release notes.

Pi agent toolkit adds PowerShell support and safer updates on Windows 🔗

Release v0.84.3 introduces managed update workflows and model control via `/thinking`

earendil-works/pi · TypeScript · 97.5k stars Est. 2025

The Pi agent toolkit now includes an optional PowerShell tool for Windows command execution, configurable through defaultTools. Users can stage, verify, and atomically activate updates for installer-managed installations, reducing risks during upgrades.

Model selection is enhanced with a /thinking selector, searchable defaults, session-scoped persistence, and Ctrl+S to save global defaults. Built on TypeScript, Pi provides a unified LLM API, agent loop, TUI, and coding agent CLI across OpenAI, Anthropic, and Google providers.
The catch: Pi runs with the user's full permissions by default, requiring containerization or sandboxing for stronger security boundaries.

Use Cases
  • Developers automate code edits via CLI on Windows
  • Teams manage LLM agent updates without downtime
  • Engineers control model reasoning levels interactively

Source: earendil-works/pi — based on the README and release notes.

Try Omarchy Brings Hardware-Accelerated Linux to Apple Silicon Macs 🔗

Packages Arch Linux with Omarchy Quattro QEMU runtime into a signed macOS app for native performance.

themartiano/try-omarchy · Shell · 384 stars 2d old

Try Omarchy lets Apple Silicon Mac users run the upstream Omarchy desktop as a hardware-accelerated app without manual setup. It bundles a pinned ARM64 Arch Linux image, Omarchy Quattro using Apple’s Hypervisor Framework, and a Swift launcher into one notarized .

dmg. After dragging to Applications, the first launch provisions the account and starts a persistent VM stored under ~/Library/Application Support/Try Omarchy/. Features include HiDPI resizing, audio routing, and Command-to-Super shortcuts via accessibility permissions. The project relies on Homebrew, Docker, and specific library versions like libslirp 4.9.2 and SDL2 2.32.10. The catch: Video decoding remains CPU-only, slowing playback at high resolutions until an improved video path is completed.

Use Cases
  • Developers testing Omarchy on Mac without Linux setup
  • Designers needing HiDPI Linux environments on Apple Silicon
  • Users exploring Arch Linux with QEMU/HVF acceleration on macOS

Source: themartiano/try-omarchy — based on the README and release notes.

Nanobot v0.3.0 upgrades WebUI into a true agent front door 🔗

Chat now shows tool results, model fallbacks, and subagent work inline

HKUDS/nanobot · Python · 47.4k stars 6mo old

The latest release merges 260 PRs and adds 38 contributors, focusing on making the WebUI feel like a native agent workspace. Chats display quoted follow-ups, file diffs, code previews, and multimodal tool results directly in conversation.

Users can switch model presets from the composer and trigger long-running goals with /goal. Settings now adapt to narrow screens and dark mode, with added Brazilian Portuguese and improved Traditional Chinese support. Underneath, nanobot routes turn state through clearer context and lets subagents handle authorized work. The framework remains Python-based, self-hosted, and MCP-integrated, with OpenAI-compatible API access. The catch: Rapid growth has left 743 open issues, raising questions about maintenance pace and stability for production-critical workflows.

Use Cases
  • Developers automate local file and shell tasks via chat
  • Teams deploy private AI agents in Discord or Slack
  • Researchers test multi-agent workflows with local LLMs

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

AI Agents Reshape Open Source Through Modular Skill Sharing 🔗

Projects unify agent memory, skills, and observability to build composable, self-improving workflows across domains

Trendai-agents
LB623/no-negative-echoGamePhanes/GamePhanesinkboard/system-atlasbam-bam-2/solo-skillsalchaincyf/huashu-excelhalofyai/halofysodiumsun/agenttraildeepset-ai/haystackcomet-ml/opikPrimeIntellect-ai/prime-agentHKUDS/nanobotearendil-works/piKilo-Org/kilocodesantifer/career-opsvolcengine/OpenVikingakitaonrails/ai-memorytt-a1i/archifyblader/humanizerayghri/i-have-adhdcalesthio/OpenMontagechaitanyagiri/munder-difflinHiThink-Tech/Financial-APIbojieli/ai-agent-bookcan1357/oh-my-pivectorize-io/hindsighttinyhumansai/openhumantjboudreaux/cc-thinking-skillsHKUDS/CLI-AnythingTencentCloud/TencentDB-Agent-Memoryplannotator/effective-htmlchuspeeism/dashi-ppt-skillifixai-ai/iFixAillm-as-a-verifier/llm-as-a-verifierlangchain-ai/openwikimaka-agent/maka-agentMemTensor/memmy-agentinternet-court/internet-court-skillSnailclimb/JavaGuidestefan-jansen/machine-learning-for-tradingGlaube-TY/siyuan-homepagecurl/curl

Open source is converging on AI agents as modular, interoperable units — not monolithic tools but pluggable skills that agents compose dynamically. Repos like solo-skills package 26 ready-to-use agent skills for solo entrepreneurs, while system-atlas turns architecture talks into explorable isometric maps with auto-generated SYSTEM.

md. Memory layers are emerging as shared infrastructure: halofy provides organizational governance for agent identity and policy, memmy-agent offers a persistent, shared context hub across Claude Code and Codex, and vectorize-io/hindsight enables agents to learn from past interactions. Observability is becoming native: agenttrail maps real-time tool calls and file changes from Claude Code, Codex, and Cursor, and llm-as-a-verifier gives agents fine-grained, training-free feedback on output correctness. Skill ecosystems are domain-specific yet reusable: huashu-excel delivers end-to-end Excel data workflows trusted for audit, calesthio/OpenMontage lets agents run full video production pipelines, and dashi-ppt-skill generates browser-editable presentations from themes. Even niche needs are covered: ayghri/i-have-adhd prevents agents from burying answers in verbose output, and blader/humanizer strips AI telltales from text. The pattern shows open source shifting from standalone apps to agent-native primitives — skills, memory, and observability — that agents discover, combine, and improve autonomously.
The catch: Much of this remains experimental; skills often lack standardized interfaces, memory systems struggle with context scaling, and verification tools like llm-as-a-verifier still lag in real-world agent loops — promising but not yet battle-tested at scale.

Use Cases
  • Solo entrepreneurs automate business tasks with pre-built agent skills
  • Teams audit AI agent outputs using verifier frameworks without retraining
  • Developers debug agent workflows with real-time observability maps

Open Source Shifts to Modular LLM Agent Tooling 🔗

Projects now focus on composable skills, memory, and observability for AI coding agents

Trendllm-tools
LB623/no-negative-echobam-bam-2/solo-skillssodiumsun/agenttraildeepset-ai/haystackcomet-ml/opiklobehub/lobe-uiHKUDS/nanobotearendil-works/piKilo-Org/kilocodesantifer/career-opsfreestylefly/awesome-gpt-image-2cathrynlavery/diagram-designMadsLorentzen/ai-job-searchvirgiliojr94/book-to-skillAlishahryar1/free-claude-codezhaoxuya520/reverse-skillWei-Shaw/sub2apicitrolabs/ego-liteAgriciDaniel/claude-obsidiantjboudreaux/cc-thinking-skillsFailproofAI/failproofaiTencentCloud/TencentDB-Agent-MemoryTiger3807861189/J-Space-Cognition-Suite-V3.7llm-as-a-verifier/llm-as-a-verifierlightningpixel/modlyMDX-Tom/gpt-5.6-instructsnflkd/fluent-koreananthropics/claude-plugins-communityjundot/omlxMemTensor/memmy-agentshy3130/tickflow-stock-panelinternet-court/internet-court-skilldatawhalechina/self-llmgoogle-gemini/cookbookpbatard/rufus

The open source landscape is rapidly evolving around LLM-as-a-tool patterns, where developers build and share discrete, reusable components that extend AI coding agents like Claude Code, Codex, and Cursor. Rather than monolithic frameworks, the trend favors modular skills—self-contained units that add specific capabilities such as job searching (santifer/career-ops), Korean language fluency (snflkd/fluent-korean), or reverse engineering workflows (zhaoxuya520/reverse-skill).

These skills often include executable scripts and integrate directly into agent workflows via plugin systems, as seen in the community-driven anthropics/claude-plugins-community marketplace.

Complementing skill ecosystems are tools for agent observability and control. sodiumsun/agenttrail provides real-time visualization of agent plans, tool calls, and file changes, while FailproofAI/failproofai hooks into agent loops to catch dangerous actions or infinite loops before they escalate. For persistent context, MemTensor/memmy-agent offers a shared memory hub that synchronizes state across multiple agents, enabling cohesive multi-agent behavior.

Orchestration layers are also maturing: deepset-ai/haystack enables production-grade LLM pipelines with explicit control over retrieval and routing, and comet-ml/opik delivers tracing and evaluation for agentic workflows. Meanwhile, lightweight frameworks like HKUDS/nanobot and earendil-works/pi lower the barrier to building custom agents with built-in loops, memory, and TUI interfaces.

This shift reflects a move from prompting LLMs ad hoc to engineering reliable, observable, and composable agent systems—treating AI not as a black box, but as a programmable runtime.

The catch: Much of this tooling remains fragmented across incompatible agent runtimes (Claude Code vs. Codex vs. OpenCode), with limited standardization in skill interfaces or memory formats. Many projects are early-stage prototypes, lacking long-term maintenance, security audits, or clear upgrade paths—raising concerns about reliability in production use despite rapid innovation.

Use Cases
  • Developers automate job applications using AI agents with tailored CVs
  • Teams monitor and debug AI coding agent behavior in real time
  • Engineers build custom agents with shared memory and modular skills

AI-Powered UI Libraries Redefine Web Framework Boundaries 🔗

Component ecosystems now embed LLMs, translation, and data agents directly into frontend workflows

Trendweb-frameworks
ChenShuo2004/cs-boardlxf746/outlook-auto-registerxnx3/translateant-design/ant-designlobehub/lobe-uisourcegraph/zoektmeasure-sh/measureduolahypercho/codex-routerearendil-works/pisanity-io/sanityHiThink-Tech/Financial-APIdnshe/DNSHE-FreeDomainszhu1090093659/dsh-webDeveloper-Y/cs-video-coursesgoogle-gemini/cookbooklissy93/web-checkchaitin/SafeLineprojectdiscovery/httpxcurl/curljstrieb/github-statsdmtrKovalenko/odiffkarlseguin/http.zig

Open source web frameworks are evolving beyond static UI into intelligent, self-augmenting layers where AI capabilities are natively woven into component ecosystems. Projects like lobehub/lobe-ui provide React-based components specifically engineered for AIGC apps, enabling seamless integration of generative AI outputs into interactive interfaces.

Similarly, sanity-io/sanity extends its structured content studio with AI-assisted field suggestions and auto-tagging, turning content modeling into a collaborative human-AI process. Ant-design/ant-design, while established, is seeing new AI-driven extensions emerge in its ecosystem — such as auto-generated form validators and contextual help overlays powered by LLMs — signaling a shift from passive component libraries to active development partners.

This pattern extends to tooling that bridges frontend and AI infrastructure: xnx3/translate delivers zero-config, in-browser HTML translation using client-side AI models, eliminating reliance on external APIs or language files; duolahypercho/codex-router acts as a dynamic proxy for multiple LLM backends (Kimi, DeepSeek) with guided OAuth and safe rollback, letting frontend apps switch models without redeployment; and earendil-works/pi offers a unified agent toolkit with TUI and CLI interfaces, enabling developers to orchestrate LLM workflows directly from dev environments.

Together, these repos reveal a convergence: web frameworks are no longer just about rendering views — they’re becoming orchestration layers for AI agents, data pipelines, and multimodal interfaces. The boundary between framework, tool, and intelligent service is dissolving, with component libraries now serving as the front-end nervous system for AI-augmented applications.

The catch: Much of this integration remains experimental — AI-powered components often lack robust error handling, consistent state management, or accessibility guarantees; model routing layers like codex-router introduce latency and failure points that are poorly documented; and the tight coupling of UI with volatile AI APIs risks creating fragile frontends that break when models update or prompts drift, suggesting a need for clearer abstraction layers before this pattern can scale reliably in production.

Use Cases
  • Developers build multilingual SaaS apps without i18n config files
  • Teams switch LLM backends in production without redeploying frontend
  • Content editors use AI-assisted field suggestions in real time
  • AI agent workflows are triggered directly from component interactions
  • Enterprise UIs auto-generate form validators from natural language descriptions

Quick Hits

kilocode Kilo enables full-stack development with an agentic AI platform that builds, ships, and iterates code faster than traditional workflows. 27k
ant-design Ant Design provides a comprehensive, enterprise-grade React UI library with consistent design language and production-ready components. 99.2k
learn Amosblomqvist/learn offers an AI-powered personal learning system that adapts to user progress and optimizes knowledge retention. 315
Gradient Gradient delivers a flexible Python framework for building and training machine learning models with minimal boilerplate. 423
career-ops Open-source AI job search: scan job portals, evaluate listings into a structured A-H report with a global 1-5 score, tailor your CV, track applications — runs locally in your AI coding CLI (Claude Code, Codex, OpenCode, Antigravity…) 68.5k
personal-edge-proxy A practical multi-inbound, multi-outbound personal proxy setup with Xray, Hysteria2, REALITY Vision, WARP and optional static SOCKS5 routing. 274
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 Assistant Enables Local AI Control Across Messaging Platforms 🔗

Latest patch resolves npm plugin update issues for singleton-array metadata from modern clients

openclaw/openclaw · TypeScript · ▲ 165 in 1d 9mo old · Latest: v2026.7.1-2

OpenClaw delivers a self-hosted personal AI assistant that runs locally on macOS, Linux, or Windows, connecting language models to messaging channels like WhatsApp, Slack, and Signal through a unified Gateway. Built in TypeScript, it provisions its own Node.

js runtime or integrates with existing installations (Node 22.22.3+, 24.15+, or 25.9+). The system treats inbound messages as untrusted input and requires explicit approval for pairing with unknown senders via openclaw pairing approve. Tools execute on the host by default unless sandboxing is configured, with security guidance available in dedicated guides.

The v2026.7.1-2 release fixes a critical npm plugin update issue where newer npm clients send singleton-array metadata, preventing official plugins from installing or updating to correction releases. This resolves #108336, ensuring plugin lifecycle reliability for users on npm 12 or npm 11.16+.

The installer automates onboarding, verifying model access, creating a workspace, and launching the Control UI for validation. Companion apps extend functionality to voice, canvas, camera, and device-local actions.

The catch: Despite broad channel support, the assistant remains designed for a single operator, with no built-in multi-user collaboration or role-based access controls, limiting its utility in team or shared environments.

Previously in The Times “covered” — Aug 23

Use Cases
  • Developers testing local AI workflows
  • Privacy-focused users avoiding cloud assistants
  • Power users integrating AI with messaging apps

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

More Stories

Gemini Cookbook Updates With Nano-Banana 2 Image Model Guide 🔗

Adds 512px generation, search grounding, and thinking mode examples for developers

google-gemini/cookbook · Jupyter Notebook · 17.7k stars Est. 2024

The google-gemini/cookbook repository now includes step-by-step guides for Nano-Banana 2, Google’s latest image generation model. Developers can access examples showing 512px output, image grounding with search, and integration with thinking mode for improved reasoning.

The cookbook also covers Gemini 3.7 Flash for agentic tasks and the Agents API using Antigravity for custom workflows. Organized into Quick Starts and practical Examples, it remains a hands-on resource for multimodal AI development.
The catch: Open issues number 42, suggesting unresolved bugs or feature gaps that may affect reliability in production use.

Use Cases
  • Developers building image generation apps with grounding
  • Teams creating video editors using natural language prompts
  • Engineers prototyping multimodal agents with Gemini API

Source: google-gemini/cookbook — based on the project README.

DeepSpeed Adds Muon Optimizer and SDMA Support for Faster LLM Training 🔗

Latest release v0.19.5 includes ZeRO-3 offload improvements and PyTorch-style backward compatibility updates.

deepspeedai/DeepSpeed · Python · ▲ 8 in 1d Est. 2020

DeepSpeed’s v0.19.

5 release integrates the Muon Optimizer and System DMA (SDMA) for ZeRO-3, enabling better overlap of collectives on AMD GPUs. The update also introduces PyTorch-style backward passes and low-precision master states in the Core API, improving usability for researchers. Recent work like SuperOffload and ZenFlow continues to push efficient LLM training on large-scale systems, with DeepSpeed powering models such as MT-530B and BLOOM. The library remains active with frequent commits and community engagement through monthly office hours. The catch: Despite rapid feature growth, 1,324 open issues suggest ongoing challenges in stability and documentation for new integrations.

Use Cases
  • Train trillion-parameter models on GPU clusters
  • Accelerate LLM inference with ZeRO offload
  • Optimize mixture-of-experts routing in PyTorch workflows

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

Developer-Y/cs-video-courses curates 11444 forks of academic CS video lectures 🔗

Last commit 1 day ago shows active maintenance despite 2016 origin and 9.9-year age

Developer-Y/cs-video-courses · Unknown · 83.2k stars Est. 2016

The repository aggregates verified university-level computer science video playlists across 20+ topics including algorithms, quantum computing, and systems. Contributors submit pull requests for actual college courses, excluding MOOCs and tutorials per strict guidelines.

With 11,444 forks and only 2 open issues, it serves as a widely adopted reference for self-directed learners seeking structured academic content. The catch: reliance on manual curation means emerging courses may lag behind rapid university syllabus updates.

Use Cases
  • Students supplementing university courses with lecture videos
  • Self-taught engineers building CS foundations
  • Educators sourcing teaching materials for advanced topics

Source: Developer-Y/cs-video-courses — based on the project README.

Quick Hits

google-research Advances AI and ML research through open notebooks showcasing cutting-edge models, algorithms, and reproducible experiments for developers and scientists. 38.6k
opencv Provides a comprehensive, optimized C++ library for real-time computer vision tasks including image processing, object detection, and motion analysis. 90.6k
dify Enables teams to build, deploy, and scale agentic AI workflows and RAG pipelines with unified tooling across cloud, VPC, or self-hosted environments. 153.6k
open-webui Offers an intuitive, extensible web interface for interacting with LLMs via Ollama, OpenAI, and other backends, simplifying AI experimentation and deployment. 150k
langchain Delivers a modular framework for developing LLM-powered agents with tools for reasoning, memory, and integration across diverse data sources and APIs. 145k

Newton v1.5.0 Adds Vectorized Joint Control for Scalable Robot Learning 🔗

Isolated multi-world workflows and improved contact reliability enhance batched simulation fidelity.

newton-physics/newton · Python · ▲ 5 in 1d Est. 2025 · Latest: v1.5.0

Newton v1.5.

0 expands scalable robotics simulation with experimental vectorized joint control via the new newton.controllers module, enabling model-based and caller-supplied impedance control for heterogeneous GPU robot batches (#3598). The release strengthens batched simulation through isolated implicit MPM, masked solver resets, and dedicated global-world gravity, preventing cross-world state loss in multi-environment workflows (#3411, #3412, #3415, #3601, #3649, #3656, #3724). Contact handling sees deterministic hydroelastic generation, persistent geometry, and stable box manifolds to improve reproducibility in contact-rich scenes (#3661, #3649, #3775, #3641, #3732, #3776, #3756). Opt-in MuJoCo Warp sleeping now supports tree sleeping and contact-based wake-up for large-scale articulated scenes (#3731), while cable mechanics gain shear and twist helpers. Built on NVIDIA Warp and extending its deprecated warp.sim, Newton integrates MuJoCo Warp as its primary backend, emphasizing GPU acceleration, OpenUSD support, and differentiability. Licensed under Apache-2.0, it requires Python 3.10+ and an NVIDIA GPU (Maxwell or newer) with driver 545+, though macOS runs CPU-only. Examples span basic pendulums to G1 humanoid robots via uv run --extra examples -m newton.examples.
The catch: Despite rapid development, 398 open issues suggest ongoing stability challenges, and macOS users lack GPU acceleration, limiting cross-platform fidelity for Apple-based workflows.

Previously in The Times “covered” — Aug 23

Use Cases
  • Robotics researchers training policies in batched GPU simulations
  • Simulation engineers validating contact-rich mechanical designs
  • Developers building extensible physics pipelines with OpenUSD assets

Source: newton-physics/newton — based on the README and release notes.

More Stories

Text-to-cad adds CAD viewer timeouts and worker streams for large asset handling 🔗

Latest release improves reliability when inspecting complex STEP and STL files in agent workflows

earthtojake/text-to-cad · Python · ▲ 38 in 1d 4mo old

The text-to-cad library now includes worker and cold process timeouts in its CAD viewer, preventing hangs when processing large assemblies. Skills for GCODE export have been fixed, and dependency issues resolved.

A new test suite validates the viewer’s ability to stream sizable assets without blocking agent pipelines. These updates target mechanical engineers and robotics teams using AI agents to generate, inspect, or handoff CAD data locally. The project supports STEP, STL, and URDF formats, integrating with MoveIt2 and fabrication tools.
The catch: Open issues remain at 10, and the last commit was a day ago, raising questions about long-term maintenance velocity despite recent activity.

Previously in The Times “covered” — Aug 22

Use Cases
  • Mechanical engineers generating robot URDF files from text prompts
  • CAD agents inspecting STEP assemblies for collision checks
  • Robotics teams slicing STL models for 3D printing workflows

Source: earthtojake/text-to-cad — based on the README and release notes.

Gazebo Sim releases Jetty with OGRE v2 rendering and ROS 2 integration 🔗

Latest release adds high-fidelity graphics and async messaging for robotics developers

gazebosim/gz-sim · C++ · ▲ 2 in 1d Est. 2020

Gazebo Sim’s Jetty release (gz-sim10_10.0.

  1. upgrades rendering to OGRE v2 for realistic lighting, shadows, and textures. It maintains support for multiple physics engines, sensor models with noise, and SDF-based model building. Users can run simulations remotely via Gazebo Transport or control them through CLI tools and a plugin-based GUI. The simulator integrates with ROS 2 and offers pre-built models like PR2 and TurtleBot via Gazebo Fuel. Despite active development, the project carries 620 open issues, suggesting ongoing stability challenges in complex multi-robot or high-fidelity scenarios.
    The catch: High realism demands significant computational resources, limiting accessibility for edge or low-power deployments.
Use Cases
  • Developers testing autonomous navigation algorithms
  • Researchers validating sensor fusion in noisy environments
  • Educators teaching robotics with interactive 3D simulations

Source: gazebosim/gz-sim — based on the README and release notes.

Webots R2025a Boosts ROS 2 Support and Adds New Robot Models 🔗

Release expands simulation fidelity for autonomous vehicles and fluid dynamics workflows.

cyberbotics/webots · C++ · ▲ 1 in 1d Est. 2018

Webots R2025a introduces new robot models, enhanced demos, and improved ROS 2 integration, refining its role as a multi-platform robotics simulator. Built in C++, it supports ROS 2 workflows for simulating autonomous vehicles, computer vision pipelines, and fluid dynamics interactions.

Pre-compiled binaries are available for Windows, Linux (including Snap and Docker), and macOS, lowering barriers for education and industry use. The simulator remains open source, with ongoing development funded by Cyberbotics’ consulting and training services.
The catch: 228 open issues indicate persistent bugs, some long-standing, which may affect reliability in complex, multi-robot simulations.

Previously in The Times “covered” — Aug 20

Use Cases
  • Engineering students simulating ROS 2-based autonomous navigation
  • Researchers testing computer vision algorithms on virtual robots
  • Educators demonstrating fluid dynamics in robotic submarine models

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

Quick Hits

openpilot openpilot enhances vehicle safety and autonomy by upgrading driver assistance systems across 300+ supported car models with real-time robotic control. 63.5k
CADAM CADAM transforms natural language descriptions into precise 3D CAD models instantly via a browser-based text-to-CAD interface. 5k
client-sdk-cpp The LiveKit C++ SDK enables developers to build low-latency, real-time audio, video, and data applications using the LiveKit protocol. 67
SSG-48-adaptive-electric-gripper The SSG-48 adaptive electric gripper provides force-feedback control for delicate manipulation tasks in robotics applications. 178
carla CARLA offers a highly configurable, open-source simulator for testing and validating autonomous driving algorithms in diverse urban environments. 14.3k

OpenAI’s Codex Security CLI Scans Code for Vulnerabilities Using AI Models 🔗

New release improves reliability by skipping duplicate findings and checking access before publication

openai/codex-security · TypeScript · ▲ 33 in 1d 1mo old · Latest: npm-v0.1.20

OpenAI’s @openai/codex-security CLI and TypeScript SDK help developers find, validate, and fix security vulnerabilities in code using AI-powered scanning. Built for Node.

js 22.13+ and Python 3.10+, it scans repositories locally or in CI, using the OPENAI_API_KEY for authentication. The tool integrates with Docker Compose for multi-repo scans and supports alternative inference providers via custom API keys and model selection.

The latest release, v0.1.20, focuses on reliability: it now checks Linear access before publishing findings, skips issues already recorded, and recovers publication identities after failures. Credential reuse and scan selection during publishing were also fixed. Documentation was updated to reflect these changes.

While the tool shows promise in automating security checks with AI, its reliance on external APIs and recent activity — 195 open issues and a last commit just hours ago — suggest it remains in early adoption.

The catch: The tool requires Trusted Access for Cyber approval for certain protected findings, limiting immediate use for some teams without program enrollment.

Previously in The Times “covered” — Aug 22

Use Cases
  • DevSecOps teams scanning Node.js apps for flaws
  • Security engineers validating fixes in CI pipelines
  • Developers checking Python projects for vulnerabilities pre-merge

Source: openai/codex-security — based on the README and release notes.

More Stories

Web-check adds subdomain lookup and WAF detection in latest update 🔗

Release 2.2.0 improves vulnerability scanning and configuration controls for security audits

lissy93/web-check · TypeScript · 34.6k stars Est. 2023

The open-source OSINT tool web-check now includes subdomain lookup and enhanced WAF detection in version 2.2.

0, released August 2026. Built in TypeScript, it aggregates IP, SSL, DNS, headers, cookies, trackers, and carbon footprint data into a single dashboard for site analysis. Users can deploy via Docker, Vercel, Netlify, or Render to audit any website’s attack surface and server configuration. Recent commits also fixed IPv6 DNS truncation and added config options to disable specific checks.
The catch: Active development relies heavily on a single maintainer, with 33 open issues and no clear roadmap for enterprise-scale features like role-based access or API rate limiting.

Use Cases
  • Security auditors scanning client sites for misconfigurations
  • Sysadmins verifying SSL chains and DNS security extensions
  • Developers auditing third-party services for trackers and vulnerabilities

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

SafeLine WAF v9.4.0 Adds CVE Detection for Kafka Connect and .NET Deserialization 🔗

Optimized SQLi, XSS, and bot defenses improve protection against evolving web threats

chaitin/SafeLine · Go · 22.4k stars Est. 2023

SafeLine v9.4.

0 enhances its self-hosted WAF with detection for CVE-2025-27817 (Apache Kafka Connect file read) and expanded .NET deserialization attack vectors. Updates optimize SQL injection, XSS, CSRF, and Java/PHP/ASP code injection logic, plus MongoDB NoSQL and bot detection. Master-slave deployment now supports forwarding rule sync, and auth session validity is configurable for Simple Auth and SSO. The project remains actively maintained with recent commits and 1.5k forks.
The catch: Open issues at 64 suggest ongoing tuning is needed for complex rule false positives in high-traffic apps.

Use Cases
  • DevOps teams blocking SQLi and XSS on self-hosted apps
  • Security engineers mitigating bot traffic and credential stuffing
  • Enterprises deploying layered defense with rate limiting and WAF rules

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

WerWolv/ImHex Adds Signed Windows Installers to Boost Security 🔗

Latest release fixes updater bugs and eliminates hex selection gaps on scaled displays

WerWolv/ImHex · C++ · ▲ 8 in 1d Est. 2020

ImHex v1.38.

1 now ships with signed Windows installers via SignPath, addressing long-standing supply-chain concerns for binary-analysis tools. The update also resolves updater failures on Windows and macOS, ensures remote SSH file opens switch tabs instead of erroring, and closes off-by-one gaps in process memory views. Pattern editor crashes and data inspector edit bugs are patched. Despite active maintenance, The catch: its C++ core and Dear ImGui UI may deter contributors preferring Rust or web-based stacks for extensibility.

Previously in The Times “covered” — Aug 20

Use Cases
  • Reverse engineers analyzing firmware binaries
  • Programmers debugging binary file formats
  • Security researchers patching executables with IPS support

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

Quick Hits

httpx httpx provides a fast, multi-purpose HTTP toolkit for running concurrent probes and security assessments using retryablehttp for reliable network interactions. 10.3k
HackBrowserData HackBrowserData extracts and decrypts saved browser credentials, cookies, history, and more across macOS, Windows, and Linux systems. 14.5k
ProxmoxVE Proxmox VE Helper-Scripts automate common administrative tasks like VM management, backups, and cluster configuration in Proxmox VE environments. 29.4k
reconftw reconFTW automates comprehensive domain reconnaissance by orchestrating top-tier tools for subdomain enumeration, port scanning, and vulnerability detection. 8k
nginx NGINX Open Source delivers high-performance HTTP and reverse proxy server capabilities for load balancing, caching, and serving web traffic efficiently. 31.5k

Lightpanda’s headless browser bypasses Chromium forks with native Zig build 🔗

New nightly release adds ARM64 Docker images and WSL2 forwarding for cross-platform automation

lightpanda-io/browser · Zig · ▲ 28 in 1d Est. 2023 · Latest: nightly

Lightpanda’s headless browser, written entirely in Zig, now runs natively on ARM64 via official Docker images, expanding its reach beyond x86_64 Linux and macOS. The project avoids Chromium or WebKit forks, instead implementing a custom rendering engine designed for AI agent interaction through the Chrome DevTools Protocol.

Recent commits show improved WSL2 integration, allowing Windows users to run Lightpanda in a Linux subsystem while forwarding localhost:9222 to host-based Puppeteer or Playwright scripts. Benchmarks cite fetching 933 real web pages on an AWS EC2 m5.large instance, demonstrating efficiency in network-heavy automation tasks. Installation remains straightforward via Homebrew, Arch AUR, or direct nightly binaries for Linux and macOS, though Windows requires WSL2 due to the absence of a native binary. The browser supports standard automation flags like --wait-until and --dump png for flexible output, and includes an experimental lightpanda agent for natural-language-driven control. Despite active development — last commit just 0 days ago — the project carries 96 open issues, indicating ongoing challenges in stability and feature completeness.
The catch: Lightpanda’s custom browser engine lacks full CSS and JavaScript compliance compared to Chromium, risking breakage on complex modern web apps.

Previously in The Times “covered” — Aug 18

Use Cases
  • AI agents scraping dynamic content via Playwright
  • Cross-platform UI testing in WSL2 environments
  • Low-overhead headless browsing on ARM64 servers

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

More Stories

llama.cpp 0.3.0 Adds Multimodal Support for Dots3-Note Models 🔗

Release integrates vision/audio processing via ggml backend with tensor-split optimization

ggml-org/llama.cpp · C++ · ▲ 177 in 1d Est. 2023

llama.cpp 0.

3.0 introduces native support for the dots3-note multimodal model, enabling vision and audio inference through a new DSA-ISWA KV cache. The update leverages ggml v0.22.0’s meta-backend tensor split and parallel Metal kernel compilation for improved performance on Apple silicon. Tensor-split (-sm tensor) now allows distributing model layers across CPU and GPU memory, addressing VRAM limits for large models. Multi-sequence rollback fixes enhance stability for DeepSeek 4 variants. The llama-server gains a LLAMA_SERVER_SLOTS_N_DIFF debug knob, and the web UI features tabbed chat navigation.
The catch: Quantization below 4-bit often degrades accuracy significantly, limiting usefulness for precision-sensitive tasks despite memory gains.

Previously in The Times “covered” — Aug 20

Use Cases
  • Developers running local LLMs on MacBooks with limited RAM
  • Teams deploying vision-language models via REST API in Docker
  • Researchers testing 3-bit quantized models on RISC-V hardware

Source: ggml-org/llama.cpp — based on the README and release notes.

Go's GitHub mirror sees steady commits as language stabilizes 🔗

Recent activity focuses on tooling improvements and minor bug fixes

golang/go · Go · ▲ 11 in 1d Est. 2014

The Go programming language's GitHub mirror at github.com/golang/go shows consistent maintenance with commits as recent as today.

Contributors address open issues—currently over 10,000—refining the compiler, standard library, and build tools. Recent work includes updates to the Go toolchain and runtime performance tweaks. Binary releases remain available at go.dev/dl for Linux, macOS, and Windows. The project continues to prioritize backward compatibility while adding incremental features. The catch: Despite active development, the high volume of open issues suggests ongoing challenges in balancing new feature requests with stability and long-term maintenance.

Use Cases
  • Building cloud-native microservices
  • Developing command-line interfaces
  • Creating high-performance network servers

Source: golang/go — based on the project README.

Quick Hits

prometheus Collects and stores time-series metrics for real-time monitoring and alerting of system performance and application health. 65.8k
fzf Enables fast, interactive filtering of lists and files in the terminal using fuzzy matching for rapid navigation. 82.7k
bun Provides a unified, high-performance toolchain for JavaScript/TypeScript development with native-speed execution and bundling. 95.7k
RuView Transforms ambient WiFi signals into actionable spatial and biometric insights without cameras or wearable sensors. 91.7k
cc-switch Integrates multiple AI coding assistants into a single cross-platform desktop interface for streamlined agentic development workflows. 129.5k
lazygit simple terminal UI for git commands 81.6k

RealSense SDK adds person detection and close-range depth for robotics vision 🔗

New filters and ROS2 playback improve real-time 3D sensing in v2.58.3 release

realsenseai/librealsense · C++ · ▲ 9 in 1d Est. 2015 · Latest: v2.58.3

The librealsense SDK now supports person detection over USB with improved accuracy via center-of-mass clustering and non-maximum suppression, expanding beyond its prior DDS-only limitation. Close-range depth sensing is enhanced on D555 and GMSL devices through a new embedded filter API (RS2_EXTENSION_CLOSE_RANGE_EMBEDDED_FILTER), loaded dynamically at runtime.

Developers gain native ROS2 playback for .db3 recordings, eliminating conversion steps in robotics pipelines. The React-based Viewer adds stream-window swapping, firmware updates from user-supplied bin files, and metadata panels for debugging. Intel branding is removed from camera visible names, aligning with the project’s move to the RealSenseAI organization.

Builders targeting drones, skeletal tracking, or facial authentication benefit from cross-platform support (Windows, Linux, macOS, Android, Docker) and wrappers for Python, ROS, C#, and Unity. The SDK remains the defacto standard for stereo depth in robotics, with active community maintenance and frequent commits.

The catch: Despite rapid feature additions, 477 open issues persist, including unresolved GPU memory leaks in CUDA pointcloud optimizations and inconsistent multi-FPS handling in the Viewer — raising concerns about long-term stability in production vision systems.

Previously in The Times “covered” — Aug 24

Use Cases
  • Robotics teams implementing real-time obstacle avoidance
  • Developers building 3D scanning pipelines for cultural heritage
  • Engineers integrating depth sensing into autonomous drone navigation

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

More Stories

OpenWifi FPGA Design Matches Commercial Wi-Fi Performance After NLNET Tests 🔗

Release v1.5.0 confirms open-source 802.11n stack works on low-end Zynq 7020 hardware

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

The open-sdr/openwifi project released v1.5.

0 with verified FPGA improvements matching or exceeding commercial Wi-Fi chip performance in NLNET-funded tests. Key updates include deterministic IQ timing via ADC/DAC FIFO removal, enhanced PHY algorithms for multipath environments, and CSI fuzzer bug fixes. The design maintains compatibility with Linux mac80211 and supports 802.11a/g/n across 70 MHz to 6 GHz. Builders can deploy it on affordable hardware like the Xilinx Zynq 7020 for AP, station, or radar-communication fusion use cases. The catch: 93 open issues persist, and advanced features require a paid subscription despite AGPLv3 core licensing.

Use Cases
  • Researchers testing wireless protocols with FPGA-based SDR
  • Engineers building low-cost Wi-Fi access points
  • Developers experimenting with joint radar and communication systems

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

LiteX Enables FPGA SoC Design with Mixed-Language Support 🔗

Latest release fixes SD card, CPU cores, and clock handling for Xilinx platforms

enjoy-digital/litex · Python · 4.1k stars Est. 2015

The enjoy-digital/litex project provides a Python-based framework for building FPGA cores and system-on-chip designs. It integrates buses, simple and complex cores, multiple CPUs including RISC-V, and supports VHDL/Verilog/(n)Migen/Spinal-HDL.

The 2026.04 release addressed SD card initialization, NEORV32 and Gowin EMCPU core issues, and Xilinx clock management fixes. With 748 forks and 110 open issues, the project remains active but shows signs of maintenance strain.
The catch: Ongoing reliance on Migen for core logic limits adoption among teams preferring native SpinalHDL or modern HDL-first workflows.

Use Cases
  • FPGA engineers building Linux-capable SoCs
  • Hardware designers integrating LiteDRAM or LitePCIe
  • Academics exploring multi-core RISC-V architectures

Source: enjoy-digital/litex — based on the README and release notes.

Quick Hits

silhouette-card-maker Enables DIY creation of custom card games and proxies using Silhouette cutting machines for personalized tabletop gaming. 201
nvim-highlite Generates minimalist, developer-friendly color schemes in Lua with lightweight logic for quick Neovim theming. 252
gdsfactory Provides a Python library to design chips, PCBs, and 3D-printable hardware with intuitive tools for photonics, quantum, and MEMS applications. 1k
axi Offers synthesizable SystemVerilog AXI IP modules and verification tools for building high-performance on-chip communication systems. 1.7k
ghdl Delivers a full-featured VHDL simulator supporting 2008/93/87 standards for accurate hardware design verification and testing. 2.9k
maker.js Enables 2D vector line drawing and shape modeling in TypeScript for precise CNC and laser cutter workflows. 2k

Assimp 6.0.5 refines GLTF skinning export and memory handling 🔗

Latest release fixes vertex processing and FBX output precision for game asset pipelines

assimp/assimp · C++ · ▲ 3 in 1d Est. 2010 · Latest: v6.0.5

The Assimp team released v6.0.

5, focusing on stability and precision in 3D asset import. Key changes include improved skinning data export to GLB/GLTF formats, ensuring skeletal animations retain correct joint influences when transferred to engines like Unity or Unreal. Memory usage in the JoinVerticesProcess step was significantly reduced, benefiting large-scale model processing in automated pipelines. Floating-point value output to FBX files was corrected to prevent precision loss during round-trip conversions. Build compatibility was extended to Haiku OS, and numerous edge cases were addressed, including implicit conversion errors, missing normal transformations, and index validation in mesh processing. The library continues to support over 40 3D formats via a unified C++ API, with bindings for C#, Java, Python, and more, alongside post-processing tools for triangulation, tangent space generation, and material merging. Assimp remains integrated via vcpkg and CMake, with experimental viewers and ports to Rust, Haxe, and JVM ecosystems.
The catch: Despite active maintenance, 542 open issues indicate ongoing challenges in handling niche or malformed file formats, and export capabilities lag behind import breadth for newer standards like USDZ.

Previously in The Times “covered” — Aug 22

Use Cases
  • Game developers importing FBX/GLTF assets into Unity or Unreal
  • CAD engineers converting STL/IFC models for visualization pipelines
  • Mobile developers processing 3D models on Android using Assimp’s native bindings

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

More Stories

Godot Demo Projects Update with New 4.7 Compatibility Samples 🔗

Release adds eight new demos covering 2D drawing, 3D physics, and rendering techniques

godotengine/godot-demo-projects · GDScript · 9.4k stars Est. 2016

The godotengine/godot-demo-projects repository released version 4.7-6ad6167, adding eight new demo projects for Godot 4.

  1. Four 2D demos showcase custom drawing, texture rendering, navigation mesh chunks, and polygon/line handling. Four 3D demos include sprite systems, navigation mesh chunks, physics interpolation, ragdoll physics, sky shaders, tonemapping, color correction, and visibility ranges (HLOD). These samples help developers learn engine features through runnable code. The project maintains branches for Godot 2.x, 3.x, and 4.x development tracks.
    The catch: Browser-based demos via GitHub Pages run with reduced performance compared to native desktop or mobile builds, limiting their usefulness for performance-critical evaluation.
Use Cases
  • Learn Godot 4.7 2D custom drawing techniques
  • Test 3D physics interpolation and ragdoll systems
  • Evaluate sky shaders and tonemapping in Godot 4.7
  • Study navigation mesh chunk implementation in 2D and 3D
  • Explore polygon lines and drawable textures in 2D games
  • Assess visibility ranges (HLOD) for large 3D scenes

Source: godotengine/godot-demo-projects — based on the README and release notes.

Magictools: MagicTools curates essential game dev assets for indie creators 🔗

Updated daily with free and paid resources spanning 2D art, 3D models, and audio tools

ellisonleao/magictools · Markdown · 17.2k stars Est. 2014

The MagicTools repository aggregates game development resources into a single Markdown list, helping builders find graphics, audio, engines, and asset marketplaces. Recent commits show steady maintenance, with the last update just one day ago adding new free sprite packs and texture tools.

It serves as a living index for prototyping, game jams, and production pipelines across Unity, Godot, and custom engines.
The catch: With 13 open issues and no formal categorization system, navigating niche assets requires manual scanning.

Use Cases
  • Indie devs sourcing free RPG icons for commercial projects
  • Artists finding Blender-compatible 3D models and matcap textures
  • Teams comparing paid asset stores like GameDev Market and Kenney Assets

Source: ellisonleao/magictools — based on the project README.

Luanti patches critical security flaws in voxel engine release 5.17.0 🔗

Urgent update addresses client and server vulnerabilities affecting all users

luanti-org/luanti · C++ · 13.5k stars Est. 2011

Luanti 5.17.

0 patches critical security flaws in its open-source voxel game engine, requiring immediate upgrade for both client and server operators. The release fixes exploitable weaknesses disclosed in the changelog, though specific CVEs aren't detailed in public notes. Built with C++17 and Lua modding, Luanti powers user-generated worlds across Windows, Linux, and macOS. Despite 1,510 open issues and a decade-plus codebase, the project maintains active development with a commit just one day ago.
The catch: Persistent high issue volume suggests ongoing stability challenges despite rapid patching.

Use Cases
  • Developers creating moddable voxel games
  • Educators teaching game design with Lua
  • Hosters running custom multiplayer servers

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

Quick Hits

Godot-Game-Template Provides a ready-to-use Godot project template with menus, scene loading, tools, and an example game to accelerate 2D/3D game development. 1.6k
Babylon.js Offers a high-performance, open-source 3D rendering engine for building immersive games and interactive experiences in the browser or native apps. 26k
SpacetimeDB Enables real-time, low-latency multiplayer game development with deterministic simulation and instant state synchronization via Rust-powered serverless architecture. 25.1k
mpv-config Delivers optimized mpv playback settings for Windows, enhancing video quality and usability through pre-configured shaders and playback options. 1.9k
mpv-anime-build Automatically detects and applies anime-optimized video enhancements (Anime4K, ArtCNN, etc.) and live-action pipelines in mpv for superior visual fidelity with minimal user setup. 183
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