Design
Preset
Background
Text
Font
Size
Width
Account Pricing

The Git Times

AI Models
Claude Opus 5.5 $20/M —GPT-6 Luna $0.50/M —Gemini 3.1 Pro Preview $12/M —Grok 4.7 $4.80/M —DeepSeek V4 Pro 0813 $3.50/M ▲342%Qwen3.8 Max Prime $12/M —Kimi K3 $15/M —
Full Markets →

Single-File WebGL2 Demo Delivers Photoreal Water Without Build Tools 🔗

Real-time caustics and FFT waves run in browsers using only HTML, shaders, and base64 textures

Aureliengmz/clearwater · HTML · ▲ 31 in 1d 3d old

Aureliengmz/clearwater renders photorealistic shallow water simulation in a single HTML file, requiring no external libraries, build steps, or assets. Using WebGL2 with float render targets, it implements Jerry Tessendorf’s FFT-based ocean wave model, Evan Wallace’s refracted-grid caustics, and Inigo Quilez’s seabed tiling techniques—all within client-side JavaScript and GLSL shaders.

The seabed texture is embedded as a base64-encoded JPEG at the file’s end, regenerable via a Python tool using NumPy and SciPy. Resolution dynamically adapts to maintain frame rate, and users can interact by dragging to look around or tapping to disturb the water surface.

This approach eliminates toolchain friction for developers experimenting with real-time water rendering, offering an immediately runnable reference for caustics, FFT waves, and LEAN mapping-based highlights. It serves as both a learning resource and a deployment-ready snippet for prototypes or educational demos where zero-install access matters.

The catch: The simulation is limited to shallow water effects and lacks deep-water physics, foam, or spray, making it unsuitable for full ocean scenes without extension.

Why this leads today It enables instant, no-installation photorealistic water simulation in any browser, removing barriers to real-time graphics experimentation and education through a self-contained, library-free WebGL2 implementation.

Use Cases
  • Learn real-time water rendering with FFT and caustics
  • Embed interactive water in static educational sites
  • Prototype fluid effects without build configuration

Source: Aureliengmz/clearwater — based on the project README.

More on the Front Page

Chat On Steroids gives ChatGPT durable local workflow control across platforms 🔗

It enables multi-agent Goal/Loop workflows with file access and terminal use while respecting provider limits

totec448-spec/chat-on-steroids · TypeScript · 4.1k stars 1mo old

Chat On Steroids (CoS) is a TypeScript-based local MCP workspace that lets ChatGPT interact with your desktop — reading and editing files, running tests, and managing terminals — without invoking Codex directly. It works as a cross-platform Electron app with Chrome integration, supporting Windows, macOS, and Linux.

Instead of bypassing limits, CoS uses your existing ChatGPT conversation, sharing its usage and context boundaries.

The tool structures work through Goal, Loop, Compact & Resume, and durable multi-agent workflows. Goal tracks unfinished tasks; Loop keeps workers operating within a defined brief. Workers maintain context so they can resume where they left off. Compact & Resume preserves session and worker history when starting a fresh chat, enabling long-running projects to continue seamlessly.

Release v2.1.16 focuses on reliability: it adapts to ChatGPT’s latest UI changes, prevents chats from getting stuck, and ensures workers receive tools promptly after their first message. Message queues now time out after 15 minutes of silence to avoid permanent blocking, and chats stay linked across app restarts.

The catch: CoS depends on ChatGPT’s frontend behavior, making it fragile to UI changes — though v2.1.16 shows improved resilience — and it cannot override provider-enforced safety or usage limits, so it won’t grant extra quota or bypass restrictions.

Use Cases
  • Developers automate file edits and test runs via ChatGPT
  • Teams run parallel agents on independent code modules
  • Users resume complex workflows after restarting the app

Source: totec448-spec/chat-on-steroids — based on the README and release notes.

OmO Agent Now Runs on Senpi Fork with Bun-First Installation 🔗

Local installs default to Bun, requiring OMO_RUNTIME=node to stay on Node.js

code-yeongyu/oh-my-openagent · TypeScript · 69.6k stars 9mo old

OmO’s native agent now runs on senpi 2026.9.

27, shifting install defaults to Bun when version 1.4.0+ is detected. The omo update command now suggests bun add -g omo-ai for local checkouts, while npm/pnpm/Yarn installs follow Bun’s model unless overridden. Writing tasks default to Claude Opus 5.5 at low tier, dropping Claude Fable 5.1 fallback. Users must pin models or lose writing capability without compatible Claude access.
The catch: Heavy reliance on Bun and Claude models creates friction for teams locked into Node.js or non-Anthropic providers.

Use Cases
  • Backend engineers generating full-stack features from prompts
  • Frontend teams scaffolding UI with agent-driven code
  • Researchers automating cross-source analysis into reports

Source: code-yeongyu/oh-my-openagent — based on the README and release notes.

Rust Tool Runs Decision Models Locally with TypeSafe API 🔗

Adds CUDA 12 GPU support and faster CPU inference for Laya model

ollaya-dev/ollaya · Rust · 559 stars 3d old

Ollaya serves open decision models like Laya and winnow via a local daemon, mimicking Ollama but for calibrated decisions instead of text generation. It pulls models by name, serves them with TypeSafe-compatible APIs, and runs in a single Rust binary.

GPU users on older NVIDIA drivers now get CUDA 12 packs, enabling RTX 4090-class performance without upgrading. The Laya model runs faster on CPU via fused attention operators. Models load as 3MB ONNX graphs, verifying weights from Hugging Face without re-hosting.
The catch: GPU support depends on detecting driver CUDA version, which may fail in custom or headless Linux setups without standard tooling.

Use Cases
  • Agents route decisions locally using claude mcp
  • Servers classify tickets with zero-shot accuracy
  • Developers test Jev-compatible workflows offline

Source: ollaya-dev/ollaya — based on the README and release notes.

cmux integrates AI coding agent alerts into macOS terminal workflows 🔗

Built on Ghostty, it offers vertical tabs, programmable splits, and remote SSH workspaces for parallel development

manaflow-ai/cmux · Swift · 27.4k stars 8mo old

manaflow-ai/cmux is a Swift-based macOS terminal that embeds notifications for AI coding agents like Claude Code and Gemini directly into the interface. Tabs light up and panes show blue rings when agents need input, with a centralized notification panel for quick navigation.

Users can split terminals horizontally and vertically, run browser panes authenticated via imported cookies, and orchestrate remote workspaces using cmux ssh with automatic SCP file drag-and-drop. The app reads Ghostty configs for themes and fonts, supports custom commands via cmux.json, and exposes a CLI and socket API for automation. Despite active development, The catch: its narrow focus on macOS and AI agent integration limits adoption in cross-platform or traditional terminal multiplexing use cases.

Use Cases
  • Developers monitoring AI agent progress during coding tasks
  • Teams running synchronized Claude Code teammate sessions
  • Engineers managing remote servers with integrated file transfer

Source: manaflow-ai/cmux — based on the README and release notes.

Cracked CapCut Pro Suite Unlocks Premium Features for Free on Windows and macOS 🔗

The tool patches the application to remove watermarks and enable 4K export without subscription

SpecterLouse/CapCut-Pro-macOS-Windows · Unknown · 438 stars 2d old

SpecterLouse/CapCut-Pro-macOS-Windows provides a cracked version of CapCut Pro that unlocks all premium features for free. It removes watermarks, enables 4K export, and grants access to the full library of effects, templates, and transitions.

The suite works offline after setup and includes patches for AI feature stability and hardware acceleration. Users report it works with the latest CapCut version and does not delete personal projects.
The catch: The tool violates CapCut’s terms of service and carries legal and security risks from using unverified, modified software.

Use Cases
  • Video editors accessing premium CapCut features without paying
  • Creators exporting 4K watermark-free videos for client work
  • Users testing CapCut Pro tools before committing to a subscription

Source: SpecterLouse/CapCut-Pro-macOS-Windows — based on the project README.

Open LLM Pipeline Rewrites AI Text to Evade Detection While Preserving Meaning 🔗

Uses DeepSeek and translation layers to humanize content across eight languages for free

asokurasu/text-humanizer · Python · ▲ 16 in 1d 3d old

The asokurasu/text-humanizer project offers a free, open-source Python tool that rewrites AI-generated text to appear human-written. It employs a four-step pipeline: DeepSeek LLM rewrites the text, translates it to Chinese, then to Turkish via Google Translate, optionally to Japanese via DeepL, and finally back to the original language using DeepSeek to restore fluency.

This process alters sentence structure and linguistic patterns to bypass AI detectors like ZeroGPT while keeping semantic meaning intact. The tool supports English, Japanese, Chinese, Korean, German, French, and Spanish. The catch: Reliance on external translation APIs (Google Translate, DeepL) introduces dependency and potential cost or availability issues for scaled use.

Use Cases
  • Developers avoiding AI content flags in automated reports
  • Writers refining AI drafts for natural tone and flow
  • Researchers testing LLM robustness against detection systems

Source: asokurasu/text-humanizer — based on the project README.

OneTake Replaces Cuts With Continuous Motion for AI-Generated Product Films 🔗

Uses an oracle to verify visual continuity between beats in generated launch videos

feitangyuan/onetake · Python · 446 stars 1d old

OneTake is a Claude Agent Skill that creates product launch films, teasers, and demos as a single continuous take, avoiding the disjointed feel of slide-style edits. Built in Python, it designs transitions where on-screen elements evolve — like a prompt bar growing into an app window or a line becoming a grid — so each beat flows from the last.

An oracle in probe.py validates continuity by checking what persists across frame boundaries, rejecting films that feel like slideshows.

Use Cases
  • Developers generating smooth product demo videos
  • Designers creating uninterrupted launch teasers
  • Teams producing AI-assisted feature walkthroughs

Source: feitangyuan/onetake — based on the project README.

AI Agents Shift from Chatbots to Autonomous Workflow Engines 🔗

Open source projects now chain agents with browsers, RAG, and skills for end-to-end task automation

Trendai-agents
unreallabsai/unreal-agentfreestylefly/WeChatBridgealexgreensh/anidoodlefeitangyuan/onetakecoder/xumTencent/BrowserSkilltotec448-spec/chat-on-steroidsxerrors/YuxiGoogleCloudPlatform/generative-aiAI4Finance-Foundation/FinRLGlaube-TY/siyuan-homepagelinorobot/linorobot2ddalcu/mlx-servespicylobstergames/shotcaller-godot

The emerging pattern in open source AI agents moves beyond conversational interfaces toward autonomous, multi-step workflow execution. Projects like Tencent/BrowserSkill let agents control real browsers without disrupting user sessions, enabling reliable web automation from any agent framework.

freestylefly/WeChatBridge integrates agents into native macOS workflows, forwarding chat logs to AI and Obsidian with one click. Meanwhile, feitangyuan/onetake demonstrates agents as cinematic directors—generating continuous product films where each frame evolves from the last, guided by an oracle-like continuity checker.

This cluster reveals a technical shift: agents are no longer isolated LLMs but orchestrated systems combining MCP (Model Context Protocol), skills, RAG, and sandboxed execution. xerrors/Yuxi provides a self-hosted platform unifying knowledge graphs, multi-agent coordination, and skill management for private deployment. tec448-spec/chat-on-steroids adds durable, goal-driven workflows to ChatGPT with Chrome integration and resume capabilities. Even lower-layer tools like ddalcu/mlx-serve (Zig) enable Python-free LLM inference on Apple Silicon, stripping agent backends of traditional dependencies.

The catch: Despite promising integration, the ecosystem remains fragmented—competing agent protocols, inconsistent skill definitions, and unclear interoperability between frameworks like MCP, LangChain, or custom orchestrator. Most demos work in narrow, controlled scenarios; production-grade reliability, state persistence, and security boundaries in multi-agent systems are still largely unproven at scale.

Use Cases
  • Developers automating cross-app tasks via agent-controlled browsers
  • Knowledge workers linking chat logs to AI agents and local notes
  • Teams generating continuous product demos using agent-directed cinematography

Type-safe APIs and local AI reshape open source web frameworks 🔗

Frameworks now prioritize deterministic tooling and self-hosted model integration over monolithic stacks

Trendweb-frameworks
ollaya-dev/ollayaalexgreensh/anidoodlergem227/knoweldge-basecrocodilestick/Calibre-Web-Automatedxerrors/Yuximiddleapi/orpcGoogleCloudPlatform/generative-aiAOrbitron/Eridanusmitmproxy/mitmproxyberty/bertykelvinBen/AppInfoScannerjstrieb/github-statsddalcu/mlx-serveheroiclabs/nakama

A clear pattern is emerging in open source web frameworks: a shift toward type-safe, locally executable backends that decouple API contracts from runtime infrastructure. Projects like middleapi/orpc exemplify this with its TypeScript-first approach to generating end-to-end typesafe APIs without codegen, reducing boilerplate while catching contract errors at compile time.

Similarly, ollaya-dev/ollaya extends the Ollama paradigm to decision models—Laya, decider, NLI, and GLiClass—offering a TypeSafe-compatible API to run open, local reasoning engines, effectively treating AI inference as a first-class API service. This mirrors ddalcu/mlx-serve’s Zig-native LLM server for Apple Silicon, which provides OpenAI/Anthropic-compatible endpoints without Python dependencies, emphasizing performance and deployment simplicity. Beyond AI, alexgreensh/anidoodle demonstrates how declarative, code-driven art and animation—rendered identically every time—can be treated as version-controlled API outputs, blurring the line between creative tooling and backend services. Even non-AI projects like rgem227/knoweldge-base and xerrors/Yuxi lean on MCP API calls to standardize knowledge agent interactions, suggesting a broader move toward interoperable, self-hosted cognitive layers. The trend reflects a maturing desire for deterministic, auditable, and locally controllable systems where APIs are not just interfaces but enforceable contracts—powered by strong typing, local execution, and standardized agent protocols.
The catch: While promising, this cluster remains fragmented across languages (TypeScript, Python, Zig, Go) and maturity levels, with many projects lacking robust observability, long-term support, or clear migration paths from established frameworks, raising concerns about ecosystem cohesion and production readiness beyond early adopter use cases.

Use Cases
  • Developers building internal tools with zero-runtime API contract violations
  • Teams deploying private AI agents that comply with data sovereignty policies
  • Creators generating reproducible, code-based visual assets for design systems

Deep Cuts

One Prompt Generates Looping 3D Icons with Real Transparency 🔗

Uses Python and Claude Code to turn text into shareable animated assets

samyost1/3dicon · Python · ▲ 26 in 1d

samyost1/3dicon transforms a single text prompt into a looping animated 3D icon with true alpha transparency — no green screens or post-processing. Built as a Claude Code skill in Python, it outputs ready-to-use GLTF or web-optimized formats that render smoothly in browsers, apps, or AR/VR interfaces.

Designers skip modeling pipelines; developers get production-grade micro-interactions from natural language. Ideal for dynamic UI feedback, onboarding flows, or brand-responsive icons that adapt to themes or user actions.
The catch: It’s early-stage, requires specific Claude tooling access, and output consistency varies with prompt complexity.

Use Cases
  • Frontend dev creates hover-state icons from descriptive prompts
  • UX designer prototypes animated feedback without 3D tools
  • AR app generates context-aware brand markers on the fly

Source: samyost1/3dicon — based on the project README.

Quick Hits

scriptc Compiles TypeScript directly to native binaries, enabling high-performance apps without runtime overhead. 5.2k
BrowserSkill Lets AI agents control your logged-in browser seamlessly via CLI and extension for uninterrupted automation. 7.4k
anidoodle Generates consistent, code-driven art and animations — from stickers to scored films — in dozens of styles. 524
pdoom-video Renders a dynamic music video for "I'm Upping My P(doom)" entirely through code, frame by frame. 466
company-brain Deploys a Slack-based AI teammate that retains team knowledge and autonomously executes tasks. 616
ai-system-design Teaches step-by-step AI system design using LLMs, RAG, and agents with practical, production-ready patterns. 440
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 v2026.9.6 fixes macOS app crash with notarized rebuild 🔗

The update resolves launch failures from September 24 while keeping core architecture unchanged for self-hosted AI assistants

openclaw/openclaw · TypeScript · ▲ 85 in 1d 10mo old · Latest: v2026.9.6

OpenClaw: OpenClaw released v2026.9.

6 on September 27, 2026, to address a critical crash in its macOS application. The original build failed at launch due to a notarization issue (#156861), prompting a same-day replacement with a properly signed DMG (#156881). Users who installed the faulty version must download the new DMG to restore functionality; the npm package remains unaffected.

This release does not alter OpenClaw’s core design: a TypeScript-based gateway that runs AI assistants locally across Discord, Slack, WhatsApp, and 20+ platforms, with data and credentials staying on-user hardware. Models like Claude or Codex are swappable plugins, and the system avoids phoning home beyond optional version checks. The installer still provisions Node.js 24.16+ or 26.1+ when needed, and onboarding verifies model access and workspace setup.

With 8,713 open issues and 82,164 forks, the project shows active adoption but sustained maintenance pressure. Contributors span 351 individuals, reflecting broad community involvement in the OpenClaw Foundation’s 501(c)(3)-stewarded effort.

The catch: Despite its local-first promise, OpenClaw relies on external model providers (Claude, Codex, etc.) for intelligence, meaning users remain dependent on third-party APIs and associated costs, rate limits, or policy changes—even as their data stays private.

Previously in The Times “covered” — Sep 26

Use Cases
  • Developers testing AI workflows without cloud lock-in
  • Teams deploying private assistants across Slack and Teams
  • Individuals running multimodal agents on personal laptops

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

More Stories

DeepSpeed Enables Trillion-Parameter Model Training on Standard GPU Clusters 🔗

Latest release adds configurable ZeRO checkpoint dtype and ARM SVE CPU Adam optimizer support

deepspeedai/DeepSpeed · Python · 43.2k stars Est. 2020

DeepSpeed’s v0.19.

7 update improves ZeRO-3 checkpoint flexibility with user-selectable export data types and adds ARM SVE-accelerated kernels for CPU-based Adam optimization. The library continues to support efficient training of massive models like MT-530B and BLOOM through ZeRO-offload, pipeline parallelism, and Mixture-of-Experts. Recent work includes SuperOffload for AMD GPU collective offloading and ZenFlow for stall-free CPU-GPU data transfer. The catch: Complex configuration and debugging remain barriers for teams without dedicated infrastructure expertise, despite ongoing usability improvements.

Use Cases
  • Train 1T+ parameter LLMs on 8-GPU servers
  • Accelerate inference for Mixture-of-Experts models
  • Run large-scale model distillation on CPU-GPU hybrid systems

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

Hermes Agent v0.21.5 delivers cross-platform persistence with serverless hibernation 🔗

Agents now wake on demand from Modal or Daytona backends, cutting idle cloud costs to near zero

NousResearch/hermes-agent · Python · ▲ 254 in 1d Est. 2025

The latest Hermes Agent release enables true serverless operation by hibernating environments when idle on Modal and Daytona, preserving state without ongoing compute charges. Users can switch between local, Docker, SSH, or cloud backends without reconfiguration, maintaining persistent memory and skill sets across sessions.

This update reinforces the agent’s ability to run anywhere — from a $5 VPS to GPU clusters — while supporting multimodal input via Telegram, Discord, or CLI with full TUI and voice memo transcription. Skills continue to self-improve through use, with autonomous creation and FTS5-powered cross-session recall.
The catch: Despite its portability, the agent’s learning loop and memory persistence rely on opaque internal mechanisms that offer limited visibility into how skills are validated or retained over time.

Previously in The Times “covered” — Sep 16

Use Cases
  • Developers testing AI workflows on low-cost serverless infrastructure
  • Teams deploying persistent agents across hybrid cloud and on-prem systems
  • Individuals maintaining long-context AI assistants accessible via messaging apps

Source: NousResearch/hermes-agent — based on the README and release notes.

Google Cloud Updates Generative AI Samples with Gemini 3.8 Flash 🔗

Repository now includes notebooks for Vertex AI Agent Platform and multimodal workflows

GoogleCloudPlatform/generative-ai · Jupyter Notebook · ▲ 7 in 1d Est. 2023

The GoogleCloudPlatform/generative-ai repository has been updated with notebooks showcasing Gemini 3.8 Flash and the Vertex AI Agent Platform.

Developers can explore function calling, retrieval-augmented generation, and multimodal apps using Imagen, Veo, and Chirp models. The repo includes setup guides for Colab and Workbench, plus ADK-based agent samples for conversational and multi-agent workflows. It also links to the Agent Starter Pack for production-ready templates. The catch: Despite recent updates, 90 open issues suggest ongoing maintenance gaps that may affect reliability for enterprise use.

Use Cases
  • Build multimodal generative AI apps on Google Cloud
  • Create search engines using Agent Search for enterprise data
  • Develop RAG-grounded agents with Vertex AI Gemini API

Source: GoogleCloudPlatform/generative-ai — based on the project README.

Quick Hits

AutoGPT Enables autonomous AI agents to plan, execute, and iterate tasks using LLMs, letting builders focus on goals rather than low-level workflow mechanics. 187.6k
ultralytics Provides a unified, state-of-the-art computer vision framework for detection, segmentation, classification, pose, and tracking across multiple YOLO versions. 62k
gepa Uses AI-driven reflective optimization to automatically refine prompts, code, and configurations through iterative self-improvement loops. 6.8k
julia Offers a high-performance, dynamic programming language designed for scientific computing with speed approaching C and usability of Python. 49.2k
500-AI-Machine-learning-Deep-learning-Computer-vision-NLP-Projects-with-code Delivers a curated collection of 500 ready-to-run AI/ML projects spanning computer vision, NLP, and deep learning with executable code for rapid prototyping. 37k

PX4 v1.17 adds Altitude Cruise mode for smoother multicopter flight 🔗

Release improves fixed-wing takeoff behavior and ROS 2 control interfaces for drones and rovers

PX4/PX4-Autopilot · C++ · ▲ 2 in 1d Est. 2012 · Latest: v1.17.0

PX4 v1.17 introduces Altitude Cruise mode, a new multicopter flight mode that maintains tilt and heading on stick release, enabling steady velocity cruise instead of abrupt stops.

The update also refines fixed-wing takeoff logic, allowing vehicles to keep climbing with level wings during navigation loss and use takeoff waypoints to define loiter positions. ROS 2 integration gains cleaner high-level interfaces for fixed-wing and rover control, with the in-tree Zenoh middleware now rmw_zenoh-compatible. Simulation expands with Gazebo Jetty support and Ackermann SIH, while three new INS drivers (MicroStrain, sbgECom, EULER-NAV) join the ecosystem alongside Septentrio GNSS resilience reporting and barometer auto-calibration. PX4 continues to run on NuttX, Linux, and macOS, supporting multirotors, fixed-wing, VTOL, rovers, and experimental platforms via the Pixhawk ecosystem. Built around uORB, a DDS-compatible middleware, it offers modular, thread-safe architecture with first-class MAVLink and DDS/ROS 2 support. The Dronecode Foundation maintains vendor-neutral governance under a BSD-3 license.
The catch: Despite broad hardware support, PX4’s complexity and C++ codebase present a steep learning curve for newcomers, and real-time performance tuning remains challenging on resource-constrained flight controllers.

Previously in The Times “covered” — Sep 25

Use Cases
  • Develop autonomous delivery drones with precise altitude hold
  • Build fixed-wing mapping UAVs resilient to GPS dropouts
  • Create ROS 2-powered rovers with PX4-based motion control

Source: PX4/PX4-Autopilot — based on the README and release notes.

More Stories

Text-to-CAD Adds Sheet Metal DFM and Engineering Drawing Skills 🔗

Release v0.6.6 introduces manufacturability checks and PDF part documentation for agent workflows

earthtojake/text-to-cad · Python · ▲ 31 in 1d 5mo old

The earthtojake/text-to-cad library now includes a DFM skill for sheet metal, CNC, and injection molding with integrated measurement validation, plus an engineering drawings skill that generates PDF projections from 3D parts. These additions, contributed by @amywork777, expand the agent skill set for automated manufacturability review and documentation handoff.

Skills are installed via the Skills CLI, which ensures each agent receives pinned dependencies matching the cadgen release at publication. Provider-native plugins support Codex, Claude Code, and Grok Build, though Codex requires v0.142.0+ for plugin recognition.
The catch: The project’s rapid skill expansion raises questions about long-term consistency and documentation depth across its growing agent workflows.

Previously in The Times “covered” — Sep 16

Use Cases
  • Mechanical engineers validate sheet metal designs for manufacturability
  • Robotics teams generate URDF/SRDF from CAD for simulation
  • Fabrication agents source and slice STEP/STL files for CNC workflows

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

NiceGUI 3.17.1 fixes routing regression breaking relative navigation links 🔗

Patch restores FastAPI-compatible route behavior after v3.17.0 introduced trailing-slash issue

zauberzeug/nicegui · Python · ▲ 2 in 1d Est. 2021

The latest NiceGUI release reverts a routing change that broke `ui.navigate.

to()` calls on prefixed APIRouter pages, restoring behavior consistent with FastAPI and v3.16. Developers building micro web apps, dashboards, or robotics interfaces in Python regain reliable relative navigation without manual path adjustments. The fix addresses #6281, reopened after v3.17.0 inadvertently stripped trailing slashes from mounted routes, causing link failures in nested applications. NiceGUI remains installable via PyPI, Docker, and conda-forge, with live reload and browser-based UI rendering.
The catch: Active issue count remains high at 73 open tickets, suggesting ongoing maintenance pressure despite recent fixes.

Previously in The Times “covered” — Sep 14

Use Cases
  • Engineers tuning motor controllers via browser-based dashboards
  • Data scientists adjusting ML hyperparameters with live plots
  • Robotics teams visualizing sensor data in real-time interfaces

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

Open-source adaptive electric gripper enables precise force control for assembly 🔗

Force adjustable from 5 N to 80 N with 48 mm stroke and 400 g mass, supports ROS2

Source-Robotics/SSG-48-adaptive-electric-gripper · CMake · 182 stars Est. 2024

The SSG-48 adaptive electric gripper uses Spectral micro BLDC drivers to deliver tunable gripping force between 5 N and 80 N, ideal for delicate and rigid object handling in assembly and human-robot collaboration. With a 48 mm stroke and 400 g mass, it integrates via ROS2 packages, URDF/MJCF files, and a Python API.

Mechanical designs, firmware, and assembly docs are open source, letting users build or customize the gripper for any robotic arm. Latest release V1.1 was tested with the PAROL6 arm and includes GUI software and BOM guidance. The catch: The project acknowledges experimental status, with no guarantees on safety or reliability, placing full responsibility on the builder for assembly, calibration, and operation.

Use Cases
  • Assembly line workers handling fragile electronics components
  • Researchers prototyping adaptive grippers for collaborative robots
  • Engineers integrating force-feedback grippers with ROS2-based arms

Source: Source-Robotics/SSG-48-adaptive-electric-gripper — based on the README and release notes.

Quick Hits

linorobot2 Enables autonomous navigation for mobile robots with support for 2WD, 4WD, and Mecanum drive platforms using ROS 2 and sensor fusion. 1k
openpilot Provides an open-source driving agent that enhances vehicle safety and automation through real-time perception and control on 300+ car models. 63.7k
PAR6-Collaborative-Robot-Arm Offers a precision open-source robotic arm designed for education, research, and AI integration using G-code control and modular hardware. 36
drake Delivers a comprehensive C++ toolkit for modeling, simulating, and verifying complex robotic systems with optimization and contact dynamics. 4.2k
rtabmap Supplies a real-time visual SLAM and mapping library with loop closure and 3D reconstruction for autonomous navigation in unknown environments. 4k

OpenAI's Codex Security CLI scans code for vulnerabilities using AI-driven policy and fixes 🔗

Latest release adds Windows compatibility, CI workflows, and improved error reporting for devsecops teams

openai/codex-security · TypeScript · ▲ 11 in 1d 2mo old · Latest: npm-v0.1.30

The @openai/codex-security package provides a TypeScript SDK and CLI tool for defining security policies, scanning codebases, and automatically suggesting fixes for vulnerabilities. It leverages OpenAI's Codex model to interpret security rules and detect issues across dependencies and custom code.

Developers initiate scans via npx @openai/codex-security scan, which analyzes files against drafted SECURITY.md policies stored outside the repository to avoid accidental commits. Findings are validated and prioritized, with optional automated remediation pull requests.

Recent updates in version npm-v0.1.30 focus on platform robustness and CI integration. The CLI now runs on Windows without requiring symlink privileges, addressing a key barrier for enterprise environments. New GitHub Actions and Azure Pipelines examples enable Bedrock-based scans in CI/CD pipelines. Error handling was improved to report full-output scan failures as explicit errors, reducing silent failures in automation. The tool also includes Docker Compose configurations for scanning multiple repositories and a separate findings service that stores results in SQLite with a paginated read-only dashboard at /dashboard.

The catch: The tool requires Node.js 22.13.0+ and Python 3.10+, and some advanced features like protected findings and Trusted Access for Cyber approvals are gated behind OpenAI's cybersecurity program, limiting immediate accessibility for teams outside the approved workflow.

Previously in The Times “covered” — Sep 17

Use Cases
  • DevSecOps engineer scanning Node.js apps for dependency flaws
  • Security team enforcing custom policy across monorepos
  • CI/CD pipeline blocking merges on critical vulnerabilities

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

More Stories

Community Proxmox scripts streamline self-hosted installs with one-command setup 🔗

Latest release fixes bugs in Manyfold, SteamCMD, and Nginx Proxy Manager containers

community-scripts/ProxmoxVE · Shell · ▲ 8 in 1d Est. 2024

The community-scripts/ProxmoxVE project offers Shell-based one-command installers for hundreds of self-hosted services on Proxmox VE, including Home Assistant, Jellyfin, and databases. Users paste a script into the Proxmox shell, choose Default or Advanced mode, and get a running container or VM in under five minutes.

Each install includes a post-install helper for updates and configuration without manual file edits. The September 26 release addressed writable data paths for Manyfold, SteamCMD pre-install updates, and Nginx Proxy Manager certificate handling.
The catch: 22 open issues suggest ongoing stability gaps, and reliance on community maintenance means script quality and update frequency can vary across services.

Previously in The Times “covered” — Sep 16

Use Cases
  • Homelab operators deploy Home Assistant in under five minutes
  • Media enthusiasts install Jellyfin with automated GPU passthrough
  • Developers test databases like PostgreSQL via single-command containers

Source: community-scripts/ProxmoxVE — based on the README and release notes.

reconFTW v4.1 expands subdomain hunting with SRV and PTR sweeps 🔗

Latest release adds TLS certificate harvesting from raw IPs for deeper asset discovery

six2dez/reconftw · Shell · ▲ 4 in 1d Est. 2020

The six2dez/reconftw project released v4.1 on March 6, 2026, introducing automated SRV record enumeration (`_ldap.

_tcp, _sip._tcp) via dnsx to uncover service infrastructure hosts. It now performs PTR sweeps over ASN-discovered CIDR ranges and attempts DNS zone transfers on delegated nameservers. A new tls_ip_pivots` function harvests TLS certificates from raw IPs, using passive extraction, SNI brute-forcing, and reverse PTR probing to find hidden subdomains. Outputs feed into httpx for delta-only probing. Built in Shell, it supports Docker, Terraform, and Ansible deployments with AX Framework for distributed scanning.
The catch:** Despite active development, the project has one open issue and relies on a complex web of external tools, creating potential friction in air-gapped or tightly controlled environments.

Use Cases
  • Penetration testers automate subdomain enumeration for client engagements
  • Bug bounty hunters expand attack surface with passive and active recon
  • Security teams integrate reconFTW into CI/CD pipelines for continuous asset discovery

Source: six2dez/reconftw — based on the README and release notes.

Quick Hits

berty Enables secure peer-to-peer messaging that functions offline or without internet, cellular, or network trust for private communication 9.3k
infisical Provides an open-source platform to manage secrets, certificates, and privileged access with centralized control and automation 29.5k
trufflehog Detects, validates, and analyzes leaked credentials across codebases and repositories to prevent security breaches 28.1k
hosts Aggregates and extends curated hosts files to block unwanted domains, with optional filters for adult content, social media, and more 31.1k
opencti Delivers a comprehensive cyber threat intelligence platform for collecting, correlating, and acting on threat data in real time 10.1k
mitmproxy Offers an interactive TLS-capable HTTP proxy to inspect, modify, and debug traffic for security testing and development workflows 45.2k

nlohmann/json Adds Byte-Position Tracking to Improve JSON Debugging 🔗

New diagnostic macro helps developers pinpoint parse errors in large or nested JSON payloads

nlohmann/json · C++ · ▲ 2 in 1d Est. 2013 · Latest: v3.12.0

The nlohmann/json library, a widely used header-only JSON parser for modern C++, has introduced a new diagnostic feature in its v3.12.

0 release that records byte positions during parsing. Enabled via the JSON_DIAGNOSTIC_POSITIONS macro, the update adds member functions to query where values originated in the input stream and includes this data in exception messages. This helps developers quickly locate syntax errors in complex JSON files, a common pain point when debugging configuration or API payloads.

The release also enhances conversion macros for arbitrary types, making them templated to work with json, ordered_json, and other basic_json specializations. Derived classes can now use these macros seamlessly, improving flexibility for custom type integration. All changes are backward-compatible, continuing the project’s focus on trivial integration via a single json.hpp header and zero dependencies.

Despite its maturity—over 13 years old and used in countless projects—the library maintains rigorous testing, including OSS-Fuzz fuzzing and full sanitizer coverage. It supports binary formats like BSON, CBOR, and MessagePack, and remains compatible with C++11.

The catch: While feature-rich, the library’s reliance on exceptions for error handling may not suit performance-critical or embedded systems where deterministic behavior is required, and compile times can increase due to its extensive header-only template metaprogramming.

Previously in The Times “covered” — Sep 16

Use Cases
  • Debug JSON config files in C++ backend services
  • Serialize STL containers to JSON for REST APIs
  • Parse binary JSON formats like MessagePack in IoT applications

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

More Stories

CodeGraph v1.6.0 Adds GitHub Copilot Support and One-Click Setup 🔗

Fixes disk-space leaks and improves context accuracy for AI coding agents

colbymchenry/codegraph · C · 72.2k stars 8mo old

CodeGraph’s latest release extends compatibility to GitHub Copilot across VS Code, CLI, and JetBrains IDEs via codegraph install. The --yes --init flag now enables agent wiring and project indexing in a single command, streamlining onboarding for CI and fresh environments.

Context precision improved with codegraph_explore, which avoids redundant output and prioritizes relevant code over generated files or type shims. Indexing stability is enhanced for large C/C++, Swift Vapor, and monorepo workloads, eliminating prior drift and crash risks. A persistent disk-space leak from interrupted sessions is resolved, with write-ahead logs now auto-trimmed on restart.
The catch: Indexing remains a manual step (codegraph init) after CLI setup, requiring user action to activate per-project knowledge graphs.

Previously in The Times “covered” — Sep 14

Use Cases
  • AI agents get accurate, token-efficient code context
  • Teams enforce consistent test coverage via PR impact analysis
  • Developers avoid context repetition in large monorepos

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

Llama.cpp adds multi-address server binding and image output support 🔗

Release v0.5.0 improves CUDA convolution and Metal MoE fusion for faster LLM inference

ggml-org/llama.cpp · C++ · ▲ 102 in 1d Est. 2023

The llama.cpp project released v0.

5.0, enabling its REST API server to bind to multiple TCP addresses and UNIX sockets via comma-separated --host input. Function-call outputs now support image generation, expanding multimodal capabilities. Backend gains include CUDA conv2d acceleration via implicit GEMM and Metal optimizations for MoE and SSM_CONV layers. Built on ggml, it maintains broad hardware acceleration across Apple Silicon, x86 with AVX/AVX2/AVX512/AMX, RISC-V, and NVIDIA/AMD GPUs. Quantization down to 1.5-bit integers reduces memory use. The catch: Despite wide hardware support, complex model serving still requires significant tuning to avoid latency spikes under concurrent load.

Previously in The Times “covered” — Sep 14

Use Cases
  • Developers deploy local LLMs with GPU acceleration
  • Teams run multimodal AI serving via REST API
  • Engineers quantize models for edge device inference

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

Scrcpy v4.1 Adds VP8/VP9 Support and Fixes Camera Mirroring Constraints 🔗

Update improves video encoding flexibility and resolves size calculation errors for Android screen mirroring

Genymobile/scrcpy · C · 150.5k stars Est. 2017

Scrcpy v4.1 introduces support for VP8 and VP9 video encoders, expanding compression options beyond H.

  1. The release fixes flawed size constraint handling in flex display mode, which previously caused incorrect resolutions during camera capture and aspect ratio scaling. It also updates FFmpeg to 8.1.2, SDL to 3.4.12, and libusb to 1.0.30, alongside gamepad detection fixes and terminal title updates while running. The tool continues to mirror Android devices via USB or TCP/IP without root, offering low-latency screen sharing and input control on Linux, Windows, and macOS.
    The catch: Despite broad compatibility, camera mirroring and HID input simulation remain limited to Android 12+ and specific hardware configurations, reducing utility on older or locked-down devices.
Use Cases
  • Developers testing apps on physical Android devices via desktop
  • Users recording gameplay or tutorials directly from phone screens
  • Linux users mirroring phone as a webcam for video calls

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

Quick Hits

go Provides a robust, efficient, and portable programming language for building scalable backend systems, cloud services, and high-performance applications. 139.1k
ollama Enables quick deployment and interaction with diverse large language models via a simple local API for AI experimentation and integration. 181.8k
lazygit Offers an intuitive terminal-based interface to visualize and manage Git repositories with keyboard-driven efficiency. 82.7k
browser Delivers a lightweight, Zig-built headless browser optimized for AI-driven automation, scraping, and testing workflows. 35.6k
fzf Empowers users to rapidly filter and select items from lists using fuzzy matching in the terminal with minimal keystrokes. 83.3k

RealSense SDK adds GPU frame access for Jetson CUDA workflows 🔗

Zero-copy depth streaming eliminates host-device copies on integrated NVIDIA platforms

realsenseai/librealsense · C++ · ▲ 1 in 1d Est. 2015 · Latest: v2.58.4

The latest release of librealsense v2.58.

4 introduces rs2_get_frame_gpu_data, enabling direct CUDA/TensorRT consumption of depth frames without host-mediated copying on Jetson devices. This zero-copy GPU frame extension targets integrated-GPU platforms where memory bandwidth is critical, reducing latency in perception pipelines. Builders can now bypass traditional CPU-GPU transfers when processing RealSense streams on NVIDIA Jetson Orin and similar SoCs.

The update also adds official support for Ubuntu 26.04 and ROS2 Lyrical, expands D401 GMSL dual-RGB camera controls via dedicated v4l2 nodes, and exposes device hardware time through rs2_get_device_time_ms. Perception Studio now ships new closed-source libraries, including close-range performance enhancements (min-z) for D400 series depth sensing.

Platform improvements include optional system-provided nlohmann_json via USE_EXTERNAL_NLOHMANN_JSON, fixed pybind11 v3.0.2 compatibility, and JetPack 7.2 (L4T R39.2) support in the L4T patch script. The C++ and web-based viewers gained searchable control filters and viewport grid overlays for easier debugging.

The catch: Zero-copy GPU frames require Jetson platforms with unified memory architecture; discrete GPU setups still incur host-device copies, limiting cross-platform performance gains.

Previously in The Times “covered” — Sep 26

Use Cases
  • Robotics teams fusing depth with CUDA inference
  • Drone developers optimizing perception latency
  • Industrial vision systems using Jetson Orin for edge AI

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

More Stories

IBM's z/OS resource hub shares tools for mainframe builders 🔗

Repository hosts REXX samples, JCL, and SMF utilities for z/OS learning and automation

IBM/IBM-Z-zOS · REXX · 466 stars Est. 2017

IBM/IBM-Z-zOS provides downloadable z/OS files not included in the product, organized by type and release level. Builders access REXX scripts, JCL examples, SMF tools, and workflow aids for education, migration, and system automation.

The repo supports z/OSMF, HCD, RMF, and WLM exploration with materials like PDF guides and DLOG formatters. Last updated a day ago, it sees active use despite no IBM support or warranties.
The catch: Materials lack IBM service support, requiring users to validate content before live-system deployment.

Use Cases
  • Mainframe admins automate z/OS tasks using REXX and JCL samples
  • Learners explore zOSMF and WLM capabilities via provided tools
  • Teams migrate workloads using z/OS Requirements-Satisfied and HCD resources

Source: IBM/IBM-Z-zOS — based on the README and release notes.

Raspberry Pi Flight Tracker gains real-time image upload API for custom displays 🔗

Major rewrite shifts stable releases to main branch, legacy code moved to master

ColinWaddell/FlightTracker · Python · 204 stars Est. 2021

ColinWaddell/FlightTracker now accepts POSTed images via /api/image to display custom graphics, animations, or loops on its 64x32 RGB LED matrix. The June 2026 rewrite promotes v2 to the main branch, with master archiving v1.

FlightTracker OS offers a browser-based installer for Raspberry Pi 32/64-bit images, automating setup via systemd. A 3D-printable case and optional GPIO loading LED are available. The catch: The display shows only text or simple graphics; complex visuals or video remain unsupported due to hardware and Python rendering limits.

Use Cases
  • Hobbyists tracking live flights overhead
  • Makers displaying custom LED matrix art
  • Educators demonstrating real-time ADS-B data

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

GitHub project detects GPU tiers for adaptive web graphics 🔗

Migrated to @pmndrs/detect-gpu with offline benchmark support via local data serving

pmndrs/detect-gpu · TypeScript · 1.2k stars Est. 2018

The pmndrs/detect-gpu project classifies user GPUs into performance tiers using WebGL benchmark scores, enabling developers to adjust graphical intensity in apps like Three.js or PixiJS based on real hardware capability.

It returns tiers 0–3, where higher tiers support more demanding visual workloads, and falls back to non-WebGL experiences if needed. Benchmark data is pulled from UNPKG by default but can be self-hosted for offline or CSP-restricted environments by extracting benchmarks.tar.gz and pointing getGPUTier to a custom URL. The tool uses normalized framerate scores to assign tiers: ≥15 fps (tier 1), ≥30 fps (tier 2), ≥60 fps (tier 3). The catch: Benchmark data relies on gfxbench.com, which stopped updating in December 2025, raising questions about long-term accuracy for newer GPUs despite current data remaining valid for existing hardware.

Use Cases
  • Game developers adjusting texture quality per device
  • 3D visualization apps scaling complexity for mobile
  • Web designers enabling WebGL fallbacks gracefully

Source: pmndrs/detect-gpu — based on the project README.

Quick Hits

fanpico Fanpico lets builders programmatically control PC fan speeds via PWM for precise thermal management and quiet operation. 237
streamdeck Stream Deck SDK enables custom plugin development to extend hardware button functionality with tailored workflows and integrations. 263
openflight OpenFlight provides an open-source flight dynamics model for simulating aircraft behavior in research, training, and avionics testing. 1.1k
SonosESP SonosESP turns an ESP32-P4 into a touchscreen Sonos controller with album art, lyrics, multi-room sync, weather, and standby clock — flashable via browser. 59
hal HAL (Hardware Analyzer) offers a C++ framework for reverse-engineering and analyzing hardware behavior through side-channel and fault injection testing. 830

Magpie Upscaler Adds Auto-Hide Cursor and Top-On-Zoom Behavior 🔗

Latest release removes manual window priority toggle, ensures scaled output stays foreground

Blinue/Magpie · HLSL · ▲ 20 in 1d Est. 2021 · Latest: v0.12.1

Magpie, the HLSL-based window upscaler for Windows 10/11, has updated its scaling engine to prevent interruptions from pop-up dialogs during gameplay or app use. Version 0.

12.1 introduces automatic cursor hiding when idle, with a user-configurable delay, reducing visual clutter in fullscreen modes. The update also removes the former “Keep scaled window on top” option, instead enforcing that the upscaled output always remains above the source window — fixing a long-standing issue where legacy apps would render over the scaled view.
Built on cppwinrt and DirectX 11, Magpie continues to bundle Anime4K, FSR, and CRT shaders within a WinUI interface that adapts to system themes. Multi-monitor setups are supported, and the tool now better handles high-DPI compatibility by advising users to override scaling settings for older, non-HiDPI-aware executables.
The project maintains active localization via Weblate and follows the all-contributors spec, with recent commits showing sustained community involvement.
The catch: Magpie requires manual configuration per application and does not automate profile switching, limiting its utility in dynamic multi-app workflows where frequent reconfiguration breaks immersion.

Previously in The Times “covered” — Sep 26

Use Cases
  • Gamers upscaling retro titles to modern displays
  • Developers testing UI scaling across resolutions
  • Streamers enhancing legacy game visuals for broadcast

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

More Stories

Warcraft-GoA Team Updates CK3 Mod with Azeroth Lore 🔗

Global contributors advance Crusader Kings 3 integration using GLSL for visual enhancements

The Warcraft-Guardians-of-Azeroth-2 project maintains a development build for its Crusader Kings 3 mod set in the Warcraft universe. Last updated September 27, 2026, the team continues refining the adaptation of Azeroth’s history from classic strategy games and WoW Classic.

Contributors use GLSL to implement graphical improvements within the CK3 framework. The mod requires manual installation and testing via Discord coordination, with builds labeled as Alpha and prone to save-breaking updates. The catch: With 446 open issues and no recent breaking changes, stability remains a concern for players relying on consistent saves.

Use Cases
  • Modders testing CK3 compatibility with Warcraft assets
  • Players seeking Azeroth-themed grand strategy experiences
  • Developers contributing GLSL-based visual upgrades to Paradox mods

Source: Warcraft-GoA-Development-Team/Warcraft-Guardians-of-Azeroth-2 — based on the project README.

Godot Dialogue Manager v4.1.0 adds CSV export and UID path support 🔗

The nonlinear dialogue tool for Godot 4.6+ gains practical workflow upgrades in latest release

nathanhoad/godot_dialogue_manager · GDScript · 3.9k stars Est. 2022

Nathan Hoad’s Dialogue Manager 4 remains a key addon for Godot developers needing branching conversations. The v4.

1.0 release introduces CSV export for dialogue data and allows UIDs in imported paths, improving localization and asset pipeline flexibility. It continues to offer a stateless editor and runtime, with C# wrapper support and MIT licensing. Used across 275 forks, it sees active maintenance despite its age.
The catch: Open issues persist, and the project shows no signs of major architectural evolution after nearly five years.

Use Cases
  • Indie devs creating RPG dialogue trees in Godot 4.6
  • Localization teams exporting dialogue via CSV for translation
  • C# Godot projects integrating dialogue via the official wrapper

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

Quick Hits

luanti Luanti lets builders create and mod voxel-based games with flexible, open-source tools for custom worlds and gameplay. 13.6k
tracy Tracy provides real-time frame profiling to help developers optimize game performance by visualizing CPU/GPU bottlenecks. 16.8k
nakama Nakama delivers a scalable backend for multiplayer games with matchmaking, leaderboards, chat, and social systems out of the box. 13.4k
GDevelop GDevelop enables anyone to build 2D/3D/multiplayer games visually or with code, no deep programming required. 26.9k
Revelation Revelation enhances Minecraft Java Edition with dynamic, exploratory shaders that transform lighting, water, and atmosphere. 651
OpenSpeedy OpenSpeedy lets players and testers adjust game speed in real time for debugging, speedrunning, or accessibility tuning. 18.3k
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