Design
Preset
Background
Text
Font
Size
Width
Account Pricing

The Git Times

AI Models
Claude Fable 5.1 $50/M —GPT-6 Astra $50/M —Gemini 3.1 Pro Preview $12/M —Grok 4.6 $6/M —DeepSeek V4 Pro 0813 $3.16/M ▼5%Qwen3.8 Max (0902) $6/M —Kimi K3 $13.28/M ▼11%
Full Markets →

Pake Turns Web Pages Into Native Desktop Apps Using Rust and Tauri 🔗

Latest release fixes download handling and window behavior across macOS, Windows, and Linux

tw93/Pake · Rust · 61.4k stars Est. 2022 · Latest: V3.16.3

Pake enables developers and power users to convert any webpage into a standalone desktop application with a single command, eliminating the need for heavy frameworks like Electron. Built with Rust and the Tauri framework, it produces lightweight apps—typically under 10MB on disk—that launch faster and consume less memory than traditional web wrappers.

The tool supports custom icons, window settings, ad removal, and immersive modes, and works across macOS, Windows, and Linux. Recent updates in V3.16.3 improve reliability: external links from email frames no longer hijack the app, download directories can be explicitly set with clear error handling, and video fullscreen on Windows now preserves player controls and captions. Linux builds targeting multiple package formats now compile once, allowing concurrent CLI builds without overwriting installed packages. macOS gains HTTP Basic authentication support and native tab grouping for additional windows.

The catch: While Pake excels at simple web-to-desktop conversion, it offers limited debugging tools and lacks built-in support for complex backend integration or native module access, making it less suitable for developers needing deep OS-level interaction beyond what a web view can provide.

Why this leads today Pake simplifies desktop app creation by converting any webpage into a native application with one command, reducing reliance on heavy frameworks and improving workflow efficiency for developers and users alike.

Use Cases
  • Developers packaging internal tools as desktop apps
  • Users creating distraction-free YouTube Music clients
  • Teams deploying lightweight intranet portals across OSes

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

More on the Front Page

Recurrent Looped Transformer Enables Infinite-Depth Reasoning Across Tokens 🔗

Combines causal encoder memory with recurrent decoder feedback for scalable latent computation

yifanzhang-pro/recurrent-looped-tranformer · HTML · 526 stars 1d old

Yifan Zhang’s Recurrent Looped Transformer (RLT) introduces a novel architecture where latent computation flows across every token in a sequence. A causal encoder builds global key–value memory from the input, while a recurrent decoder merges this memory with sliding-window attention and feedback from its own prior hidden state.

This design allows each processed token to extend a recurrent path through the decoder, enabling temporal computation depth that scales with sequence length—without increasing per-token compute. The model uses 48 encoder and 48 decoder layers, with sliding-window attention (SWA) caches and recurrent states preserved across prompt–response boundaries. Training and inference benefit from model–hardware co-design: parallel encoder work, memory reuse, and activation checkpointing improve efficiency. Crucially, the same complete-state transition supports pretraining, supervised fine-tuning, sampling, and RL replay, unifying diverse workflows under one recurrent framework. The “infinite depth” refers to an extensible temporal path over tokens, not unbounded computation per step—a distinction aimed at enabling longer reasoning chains in language models.

The catch: RLT remains a research prototype with no published benchmarks on reasoning tasks or hardware throughput, and its real-world efficacy in scaling RL or long-context understanding is unproven at scale.

Use Cases
  • Researchers studying long-horizon reasoning in transformers
  • Engineers designing recurrent architectures for scalable inference
  • Teams exploring unified pretraining and RL pipelines with shared state dynamics

Source: yifanzhang-pro/recurrent-looped-tranformer — based on the project README.

Omnigent v0.13.0 adds GitHub panel and Canvas for agent workflows 🔗

Real-time PR tracking and drag-and-drop session boards unify multi-agent collaboration

omnigent-ai/omnigent · Python · 9.9k stars 3mo old

Omnigent’s latest release integrates a live GitHub panel showing PR summaries, comments, and CI checks within agent sessions, refreshing as code pushes. The new Canvas page offers a per-project board of draggable session cards with worktree branches and PR links, letting users start sessions or view transcripts directly.

These features deepen its role as a meta-harness for orchestrating Claude Code, Codex, Cursor, and custom agents across devices.
The catch: Sandbox providers like Modal and Daytona require separate signups, and policy enforcement relies on user-defined YAML, creating setup friction for teams seeking zero-config governance.

Use Cases
  • Developers debug code using multiple AI agents in sync
  • Teams review agent-generated PRs with live CI feedback
  • Engineers orchestrate custom agents via YAML-defined workflows

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

mjbatch Enables Thousands of Parallel MuJoCo Simulations on CPU for RL and Control 🔗

Python library uses C++ thread pool with GIL released for live NumPy access across batches

kevinzakka/mjbatch · Python · 408 stars 3d old

mjbatch lets developers run large-scale MuJoCo simulations efficiently on CPU by leveraging a C++ thread pool that bypasses the Python GIL. It provides live NumPy views of simulation states and controls, supports per-simulation model parameters, and includes utilities like set_const to recompute derived constants.

The library targets reinforcement learning, MPC, system identification, and hardware co-design, with demos showing a Go1 quadruped learning to walk in under a minute on an M1 laptop. Built for Python 3.10–3.14 and free-threaded 3.14t, it offers wheels for Linux and macOS.
The catch: As a v0.1.0 release with four open issues, real-world scalability beyond demo workloads remains unproven at larger scales.

Use Cases
  • Train RL policies for quadruped locomotion using PPO
  • Implement MPC and iLQR controllers for manipulator tasks
  • Perform system identification on robotic arms via synthetic data fitting

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

Mural turns iPhone into conversational language tutor with AI orb for language practice 🔗

Uses your OpenAI key and SwiftUI to adapt lessons from spoken replies

Chuloo/mural · Swift · 285 stars 1d old

Mural is a native iPhone app that teaches languages through conversation with an animated orb. Built in SwiftUI, it uses local SwiftData storage and connects directly to OpenAI via your API key.

The app adjusts difficulty based on your spoken responses, offering real-time feedback and spaced repetition. No account or Mac is needed to run it—just an iPhone, Xcode 26+, and an OpenAI project with GPT-Live-1 or GPT-5.6 Luna access.
To install, clone the repo, configure signing in Xcode, enable Developer Mode on iOS, and enter your API key in Settings. The app requires microphone access and an internet connection.
The catch: It demands an OpenAI API key with billing, making ongoing use costly without free credit.

Use Cases
  • Language learners practice speaking with AI feedback
  • iOS developers study SwiftUI and OpenAI integration
  • Self-paced students build vocabulary through dialogue

Source: Chuloo/mural — based on the project README.

Mac Duo brings iPhone lid-tilt effect to MacBooks via menu bar 🔗

Uses GPU rendering and ScreenCaptureKit for real-time perspective, blur, and fade

sumimakito/Mac-Duo · Swift · 793 stars 3d old

Mac Duo adds the iPhone Duo visual effect to MacBooks by tilting, blurring, and fading screen content as the lid closes. Built in Swift, it leverages Metal for GPU rendering and ScreenCaptureKit to capture live display output.

Controls live in the menu bar, with adjustable perspective to match viewing position. The app requires macOS 14+, a compatible lid angle sensor, and Screen Recording permission. It works on both Apple Silicon and Intel Macs, applying the effect only to the built-in display. The catch: Only MacBooks with a built-in lid angle sensor supported by macOS can use the effect, excluding many older models and external displays.

Use Cases
  • Developers testing UI responsiveness during lid transitions
  • Designers showcasing motion effects in macOS presentations
  • Users seeking immersive laptop interactions with built-in sensors

Source: sumimakito/Mac-Duo — based on the project README.

Code-drawn explainer videos turn any topic into shareable learning clips 🔗

Built with Remotion and TypeScript, outputs sync voiceover, subtitles, and progress bars

Vincentwei1021/anything2explainer · TypeScript · 1.2k stars 5d old

Vincentwei1021/anything2explainer is a Claude Code / Codex skill that generates narrated explainer videos from a topic or document. Every frame is rendered in code using Remotion, with TTS voiceover, word-aligned subtitles, and a chapter progress bar.

Outputs are 1280×720 H.264 MP4s in English or Chinese, featuring a black-canvas motion-graphics style. The project includes a compilable template, primitives library, storyboarding tools, and QC tooling for reliable agent-driven video generation.
The catch: Output quality depends on the clarity of input topics and requires manual tuning for complex narratives.

Use Cases
  • Developers creating quick overviews of vector databases or RAG systems
  • Educators turning lesson notes into shareable explainer videos for students
  • Technical writers converting documentation into visual summaries for onboarding

Source: Vincentwei1021/anything2explainer — based on the project README.

Open Source LLMs Become Modular Skill Engines for Real-World Tasks 🔗

Projects now treat language models as interchangeable components in agent pipelines, data extractors, and domain-specific automation tools.

Trendllm-tools
openai/NavierStokesAndEulerVincentwei1021/anything2explainerSpaceDudem/text-humanizermizzlelover/gongwen-gbt9704-skillyifanzhang-pro/recurrent-looped-tranformereternityspring/reelbench-skillskruzovic7/ai-data-extractoropenonion/connectonionwiltodelta/remove-ai-watermarksomnigent-ai/omnigentdatawhalechina/self-llmExtelligence-ai/bagelj3ssie/osmedeustw93/Pakejustrach/codedbddalcu/mlx-serve

The open source LLM ecosystem is shifting from monolithic model releases to composable, skill-based tooling that wraps models in purpose-built interfaces. Rather than chasing raw benchmark scores, developers are packaging LLMs into discrete capabilities: anything2explainer turns prompts into narrated motion graphics via Remotion, while ai-data-extractor parses chat logs from Claude Code, Cursor, and Aider to mine developer intent.

Frameworks like omnigent and connectonion enable real-time agent orchestration, letting users swap between Claude Code, Codex, or custom models without rewriting workflows. Specialized skills are emerging too: gongwen-gbt9704-skill auto-generates GB/T 9704-compliant Chinese government documents in editable DOCX, and text-humanizer rewrites AI output to evade detection by Turnitin and GPTZero. Even infrastructure is adapting: mlx-serve provides a Python-free, Zig-based LLM server for Apple Silicon with native tool calling, and bagel uses MCP to query robotics and IoT data through natural language. This pattern reflects a maturation where LLMs are no longer endpoints but pluggable engines — integrated into pipelines for video generation, document automation, data extraction, and cross-agent collaboration. The catch: Much of this tooling remains fragmented across languages and agent protocols, with overlapping functionality and limited interoperability; many skills are brittle, tightly coupled to specific model versions or APIs, and lack standardized evaluation, making large-scale adoption uncertain despite rapid experimentation.

Use Cases
  • Developers extract insights from AI coding assistant histories
  • Teams generate compliant government documents from prompts
  • Engineers orchestrate multiple AI agents in real-time workflows

Open Source Builds Unified Data Infrastructure for AI Systems 🔗

Projects converge on schema, catalog, and real-time data layers to power next-gen agentic and embodied AI workflows

Trenddata-infra
kruzovic7/ai-data-extractorPhyzicalorg/Phyzical_orgopen-circle/valibotapache/gravitinowekan/wekanschemacrawler/SchemaCrawlerscikit-learn/scikit-learnLightning-AI/pytorch-lightningfastai/fastbookExtelligence-ai/bagelusememos/memosvalkey-io/valkeytaosdata/TDengineprotocolbuffers/protobufSanderMertens/flecsGodotNuts/GodotFirebase

A clear pattern is emerging in open source around foundational data infrastructure tailored for AI systems. Rather than isolated tools, projects are converging on composable layers that handle data validation, discovery, movement, and reduction at scale.

open-circle/valibot provides type-safe schema validation in TypeScript, ensuring structural integrity across data pipelines. apache/gravitino delivers a geo-distributed, federated metadata lake, acting as a unified catalog for disparate data sources critical in multi-region AI training. For real-time workloads, valkey-io/valkey and taosdata/TDengine offer high-performance key-value and time-series storage optimized for low-latency AI inference and IoT telemetry. On the edge, Extelligence-ai/bagel uses an MCP server to query robotics and IoT data in plain English while intelligently reducing data volume at the source—keeping only what matters for downstream agents. Embodied AI stacks get fuel from Phyzicalorg/Phyzical_org, which supplies browser-teleoperation trajectory data with onchain provenance and elizaOS readiness, enabling reproducible agent training. Meanwhile, kruzovic7/ai-data-extractor normalizes chat histories from AI coding assistants like Cursor and Aider, turning interaction logs into structured training data. Together, these repos signal a shift toward modular, interoperable data infra where schema, catalog, storage, and edge processing are decoupled yet composable—enabling AI systems to source, validate, and act on data with minimal friction.

The catch: Much of this infrastructure remains early-stage, with limited cross-project integration and unclear governance models; adopting multiple tools risks fragmentation before standards solidify, and real-world validation in large-scale agentic systems is still sparse.

Use Cases
  • AI agents validate structured inputs using schema libraries
  • Robotics stacks query live sensor data via natural language
  • Federated teams share metadata across distributed data lakes

AI-Powered Local-First Tools Redefine Web Framework Boundaries 🔗

Open source shifts from monolithic stacks to modular, privacy-centric utilities

Trendweb-frameworks
Vincentwei1021/anything2explainertruanayangi-com/truanayangineovim/neovimBasedHardware/omimdn/translated-contentvercel-labs/emulatentop/ntopngusememos/memospocketbase/pocketbasejstrieb/github-statsddalcu/mlx-serveMewPurPur/GodSVG

A clear pattern emerges in open source: frameworks are dissolving into specialized, AI-augmented tools that prioritize local execution and user sovereignty. Projects like Vincentwei1021/anything2explainer generate narrated motion graphics from text prompts using Remotion and TTS — all client-side, no backend.

Similarly, truanayangi-com/truanayangi runs fully offline with automatic cookie storage, eliminating login or cloud dependence. AI agents such as BasedHardware/omi interpret screen and audio input to suggest actions, while ddalcu/mlx-serve delivers native LLM inference on Apple Silicon via Zig and MLX, bypassing Python entirely. Even documentation efforts like mdn/translated-content reflect this shift — community-driven, markdown-based localization without central orchestration. These tools reject monolithic frameworks in favor of composable, privacy-first units that run locally, often in a single binary or via pnpm. The trend signals a move toward infrastructure as disposable, intent-driven scripts rather than enduring platforms.
The catch: Much of this remains experimental — local AI agents struggle with context retention, Zig and MLX tooling lacks broad ecosystem support, and truly useful AI-driven video generation is still brittle and resource-heavy, risking a gap between demo appeal and daily utility.

Use Cases
  • Developers build explainer videos from prompts
  • Teams run offline note tools with zero login
  • Users get real-time screen-aware AI suggestions
  • Engineers serve LLMs natively on Apple Silicon
  • Contributors localize docs in native language
  • Engineers emulate APIs in CI without network
  • Designers edit SVGs across desktop and web
  • SysAdmins monitor traffic with Lua-based dashboards
  • Creators generate GitHub stats from private repos
  • Deploy self-hosted backends in a single file
  • Run apps with automatic local cookie persistence
  • Edit structured SVGs with code-driven precision

Quick Hits

lunel A lightweight Python tool for generating realistic synthetic lunar data to test and validate space mission simulations and algorithms. 564
Edge0 An AI-powered edge computing framework enabling real-time inference on resource-constrained devices with minimal latency and power usage. 1.6k
YOINK A Python-based DeFi exploit detection and analysis toolkit for identifying vulnerabilities in smart contracts and blockchain protocols. 355
reelbench-skills A curated collection of HTML-based learning resources and practical tools for mastering AI video generation, editing, and deployment techniques. 523
bank-sampah A PHP application for managing waste collection and recycling logistics, optimizing routes and tracking bin fill levels in urban environments. 646
viserys-agent A JavaScript agent framework for building autonomous AI assistants that interact with web interfaces, APIs, and local tools to complete complex tasks. 628
ai-data-extractor A free, open-source extractor that parses and structures AI coding assistant chat histories from Claude Code, Cursor, Windsurf, and more for analysis or reuse. 390
rune A Go-based professional development environment offering isolated, reproducible workspaces with integrated tooling for streamlined coding, testing, and deployment. 463
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.

OpenClaw v2026.9.4 sharpens agent harnesses for cross-platform AI control 🔗

Adds deterministic policy enforcement and plugin-swappable models without reconfiguration

openclaw/openclaw · TypeScript · 389.6k stars 9mo old · Latest: v2026.9.4

OpenClaw’s latest release refines its core architecture: a trusted gateway managing untrusted agent execution through deterministic policies. Users can now swap AI models—Claude, Codex, or local LLMs—as plugins without altering workflows or reconfiguring integrations across Discord, Slack, Teams, WhatsApp, and 20+ other channels.

The Gateway provisions Node.js 24.16+ or 26.1+ automatically, or uses existing installs via npm/pnpm/Bun. State, memory, and credentials stay on-device; only anonymous feature stats and version checks phone home by default, both fully disabling via update.checkOnStart: false.

This release incorporates 20 direct commits and feedback from 294 contributors, focusing on stability in the agent harness layer and installer robustness for macOS, Linux, and Windows. No hosted service or paid tier exists; the OpenClaw Foundation stewards the project as a 501(c)(3).

The catch: Reliance on user-provided Node.js runtimes means environment drift or outdated versions (below Node 24.16) can break onboarding, requiring manual intervention despite automated provisioning claims.

Previously in The Times “covered” — Sep 4

Use Cases
  • Developers test local LLMs via Discord without exposing data
  • Teams deploy shared AI assistants on Linux using existing npm setups
  • Individuals run persistent AI workflows across WhatsApp and macOS natively

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

More Stories

Microsoft's MCP Curriculum Adds Cross-Language AI Workflow Examples 🔗

New Rust and TypeScript modules expand practical model-context-protocol training for developers

microsoft/mcp-for-beginners · Jupyter Notebook · 17.2k stars Est. 2025

Microsoft's mcp-for-beginners repository now includes hands-on exercises in Rust and TypeScript alongside existing .NET, Java, and Python examples.

The curriculum teaches developers to build modular AI workflows using the Model Context Protocol for secure service orchestration. Each language track covers session setup, client-server communication, and security practices through Jupyter Notebook lessons. Recent commits show ongoing maintenance with multilingual support expanded to 50+ translations. The catch: The project's heavy reliance on Jupyter Notebooks may limit adoption in production-focused teams preferring script-based or compiled workflows.

Use Cases
  • Backend engineers learn MCP service orchestration
  • AI developers build cross-language model integrations
  • Students practice secure AI-tool communication patterns

Source: microsoft/mcp-for-beginners — based on the project README.

MCP server collection grows despite stalled development activity 🔗

Forks exceed 16,000 as open issues climb to 2,377 with last commit one day ago

punkpeye/awesome-mcp-servers · Unknown · 94.9k stars Est. 2024

The punkpeye/awesome-mcp-servers repository aggregates Model Context Protocol server implementations for AI tooling. While no new release has appeared recently, community forks continue to multiply, indicating sustained interest in MCP server patterns.

The project serves as a reference catalog rather than active software, listing servers across various languages and frameworks. The catch: With 2,377 open issues and no clear maintenance signal, builders must vet individual servers for stability and security before production use.

Previously in The Times “covered” — Aug 29

Use Cases
  • Developers discover MCP server examples for custom AI agent tooling
  • Teams evaluate MCP implementations across different runtime environments
  • Architects compare server designs for scalable context protocol adoption

Source: punkpeye/awesome-mcp-servers — based on the project README.

Quick Hits

cs249r_book Teaches how to build scalable machine learning systems from data pipelines to deployment, with hands-on Python examples. 28.2k
fastbook Provides interactive Jupyter notebooks for learning practical deep learning with fastai, covering vision, NLP, and tabular data. 25.3k
superpowers Offers an agentic framework that turns shell scripts into reusable, composable skills for automating complex development workflows. 286.2k
gemini-cli Brings Google’s Gemini AI model into your terminal as a CLI agent for coding, querying, and automating tasks with natural language. 107k
open-webui Delivers a clean, self-hosted web UI for chatting with local and remote LLMs — supporting Ollama, OpenAI, and more — without setup hassle. 151.9k
caveman Reduces token usage by 65% through a Claude Code skill that rewrites prompts in terse, caveman-style language while preserving meaning. 105.4k

ArduPilot Plane-4.7.1 Release Boosts VTOL Stability for Autonomous Flight 🔗

Update refines fixed-wing and tilt-rotor control after 13 years of open-source evolution

ArduPilot/ardupilot · C++ · 15.9k stars Est. 2013 · Latest: Plane-4.7.1

The ArduPilot team released Plane-4.7.

1 on September 4, 2026, marking the latest stable update for fixed-wing and VTOL aircraft. This release focuses on refining attitude estimation and servo output smoothing, particularly benefiting tilt-rotor transitions where mechanical lag previously caused oscillation. Built on over a decade of community-driven C++ development, the software continues to support diverse frames—from traditional airplanes to hexarotors and underwater vehicles—via a unified MAVLink-based architecture. Contributors like Andrew Tridgell and Francisco Ferreira maintain core subsystems, ensuring patches undergo rigorous review before merging. The project’s GitHub activity remains robust, with the last commit just zero days ago and over 21,000 forks indicating widespread adoption across academia, industry, and hobbyist builders.
The catch: Despite its maturity, ArduPilot’s monolithic codebase presents a steep learning curve for newcomers, and real-time performance tuning often requires deep familiarity with PID loops and sensor fusion—barriers that may deter rapid prototyping compared to more modular alternatives.

Previously in The Times “covered” — Sep 3

Use Cases
  • Autonomous cargo delivery with VTOL drones
  • Underwater pipeline inspection using ArduSub
  • Agricultural field mapping with fixed-wing planes

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

More Stories

Sesame Robot Adds Voice Control via New Python Companion App 🔗

Open-source quadruped now supports remote WiFi control and expressive OLED face animations

dorianborian/sesame-robot · C · 4.4k stars 9mo old

The sesame-robot project recently updated its Python-based Companion App, enabling voice control and advanced interactions for the ESP32-powered quadruped. Users can now trigger pre-programmed emotes like waving or dancing through voice commands, leveraging the robot’s 128x64 OLED display for reactive facial expressions.

Built with 8 servos and 3D-printed PLA parts, the robot connects via WiFi to a JSON API for remote operation using Python, JavaScript, or a Serial CLI. The Sesame Studio animation composer lets users design custom movements without deep coding. The catch: The project relies on a 3D printer and basic soldering skills, posing a barrier for makers without access to fabrication tools or electronics experience.

Use Cases
  • Hobbyists building expressive walking robots
  • Educators teaching robotics and embedded systems
  • Developers prototyping voice-controlled IoT devices

Source: dorianborian/sesame-robot — based on the project README.

RKO-LIO Boosts Single-Threaded LiDAR-Inertial Odometry Performance 🔗

v0.4.0 improves failure recovery and reduces drift on long sensor gaps

PRBonn/rko_lio · C++ · 651 stars Est. 2025

The latest release of PRBonn/rko_lio focuses on single-threaded efficiency, changing the default max_num_threads to 1 for better ROS-side performance. It adds robust failure recovery: registration errors now reset odometry to a new identity frame instead of crashing, and IMU-based initialization allows recovery from gaps of 1 second or more.

These changes, detailed in PRs #157–#174, improve voxel search and linear system handling. The system still requires extrinsic calibration between IMU and LiDAR, specified via config or data.
The catch: Despite performance gains, the project remains early-stage with limited real-world validation beyond four tested platforms.

Use Cases
  • Robotics teams fusing LiDAR and IMU for SLAM
  • Developers processing ROS1/ROS2 bags for odometry
  • Researchers testing drift correction in varied environments

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

Bagel MCP Server Enables Plain-English Robotics Data Queries with Deterministic SQL 🔗

Latest release adds dual MCP transports and Codex plugin for seamless LLM integration

Extelligence-ai/bagel · Python · 396 stars Est. 2025

Bagel lets builders query robotics, drone, and IoT data using plain English, converting questions into auditable DuckDB SQL queries—no LLM guesswork. Its edge data reduction pipeline retains only relevant windows, like "10s around every hard brake," reducing bandwidth while preserving critical signals.

The MCP server now serves both streamable HTTP and legacy SSE by default, with a bundled plugin supporting Claude Code and OpenAI Codex natively. Dockerized demos work out of the box with ROS2 and PX4 sample logs, enabling instant health checks.
The catch: Open issues (23) and narrow domain focus may limit adoption for teams needing broader IoT protocol support or faster issue resolution.

Use Cases
  • Drone engineers audit IMU thermal spikes via natural language
  • Robotics teams trigger edge-based event detection without custom code
  • IoT analysts run fleet-wide SQL queries using plain English prompts

Source: Extelligence-ai/bagel — based on the README and release notes.

Quick Hits

RESPLE Computes precise 6-DoF motion estimates from LiDAR data using recursive spline-based odometry for robust navigation. 233
ontology-development-kit Streamlines ontology creation, versioning, and deployment with Dockerized lifecycle management tools. 374
ogma Generates verified runtime monitors in Haskell to ensure safety-critical correctness in flight and robotic systems. 580
ros2_controllers Provides reusable, ROS2-compliant robotic controllers that integrate seamlessly with ros2_control for diverse actuation. 831
dora Enables low-latency, composable AI robotic apps via distributed dataflow pipelines modeled as directed graphs. 3.9k

Authentik 2026.8.2 patches SAML and OIDC flaws in self-hosted IdP 🔗

Fixes critical redirect validation and token handling in enterprise SSO deployments

goauthentik/authentik · Python · 25.5k stars Est. 2019 · Latest: version/2026.8.2

The goauthentik/authentik project released version 2026.8.

2 on September 14, addressing two security flaws in its self-hosted identity provider. The patch resolves a SAML redirect validation bypass that could allow unauthorized access under specific SP-initiated flows, and corrects an OIDC token handling issue where expired refresh tokens were improperly accepted during silent authentication. These fixes apply to deployments using Docker Compose, Kubernetes via Helm chart, or AWS CloudFormation templates. Authentik continues to support SAML 2.0, OAuth2, OIDC, LDAP, and RADIUS, targeting teams replacing commercial IdPs like Okta or Entra ID in air-gapped or regulated environments. The update includes no breaking changes but updates internal dependencies to address CVEs in cryptographic libraries.
The catch: Despite active maintenance, the project’s enterprise features remain behind a closed-source license, limiting full auditability for security-sensitive teams.

Previously in The Times “covered” — Aug 29

Use Cases
  • Enterprises replacing Okta with self-hosted SSO
  • Dev teams securing Kubernetes workloads via OIDC
  • Labs testing SAML integration with legacy apps

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

More Stories

HackTricks wiki updates search index for Japanese security researchers 🔗

Automated ja release improves accessibility for pentesters using local STM Cyber mirror

HackTricks-wiki/hacktricks · CSS · 12.3k stars Est. 2020

HackTricks-wiki/hacktricks published a searchindex-ja release, automating Japanese-language indexing for its CTF and pentesting technique repository. The project remains a reference for offensive security knowledge, with local copies served via STM Cyber’s mirror at http://localhost:3337.

Contributors include security researchers from Intigriti and Modern Security, linking to AI security training and bug bounty programs. Despite steady updates, the wiki’s CSS-only structure limits interactive tooling and dynamic content filtering.
The catch: The static CSS format hinders real-time collaboration and advanced search compared to database-driven alternatives.

Use Cases
  • Pentesters study CTF exploit techniques offline
  • Security teams audit apps using indexed hacking tricks
  • Researchers access Japanese-translated pentesting references

Source: HackTricks-wiki/hacktricks — based on the README and release notes.

BBOT 3.0.2 fixes YARA rule gaps and boosts HTTP decoding for recon 🔗

Update raises process pool limits and parallelizes tests with pytest-xdist

blacklanternsecurity/bbot · Python · 10.6k stars Est. 2022

BBOT 3.0.

2 patches a bug where custom YARA rules matched but emitted no output, improving rule reliability. The release upgrades blasthttp to surface decode errors on HTTP responses, preventing undecoded bodies from being misread as content. Process pool max_tasks_per_child increases from 25 to 250 for heavier workloads, and the test suite now runs in parallel via pytest-xdist. Dependencies like nuclei and trufflehog are updated to latest versions.
The catch: Open issues remain at 51, and the tool’s Python-heavy stack may strain air-gapped or minimal environments despite its recursive depth.

Use Cases
  • Security teams mapping subdomains via recursive DNS brute-force
  • Bug hunters enumerating attack surfaces with YARA and NLP mutations
  • Pentesters exporting findings to Neo4j, Splunk, or CSV for correlation

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

Quick Hits

airgeddon Audits wireless networks using a multi-use bash script for Linux systems to identify vulnerabilities and improve security posture. 8k
sniffnet Monitors network traffic comfortably with Rust-powered real-time insights for developers and sysadmins seeking visibility without overhead. 41.1k
hosts Consolidates and extends trusted hosts files to block ads, malware, and unwanted content, with optional category-specific filters for enhanced privacy. 31.1k
ImHex Provides a retina-friendly hex editor for reverse engineers and programmers to inspect binary data efficiently, even during late-night debugging sessions. 54.8k
Reverse-Engineering Teaches comprehensive reverse engineering across x86, x64, ARM, AVR, and RISC-V architectures through free, hands-on tutorials for builders and learners. 14.3k
Cybersecurity-Projects Building 70 Projects ranging from beginner to advanced so anyone can — learn from, build upon, use as a reference, or even copy directly. Gamified Cybersecurity learning 👇 7k

Bun v1.4.2 sharpens JavaScript toolchain with faster installs and TypeScript 6 support 🔗

The all-in-one Rust runtime now patches security gaps and refines workspace handling for monorepos

oven-sh/bun · Rust · 96k stars Est. 2021 · Latest: bun-v1.4.2

Bun v1.4.

2 upgrades its integrated toolchain, focusing on reliability and developer velocity. The release includes a security scanner that audits dependencies during bun install, flagging known vulnerabilities in real time. Workspace management sees tighter integration, with improved hoisting logic and lockfile consistency for monorepo setups. TypeScript 6 support lands fully, enabling newer syntax and faster type checking without external tooling.

Performance remains central: bun add and bun install leverage a global cache and parallel fetching to cut install times, often outperforming npm and pnpm in cold starts. The built-in test runner now supports snapshot testing with DOM APIs, easing frontend component validation. Hot Module Replacement (HMR) in the bundler gained stability for React and Vue projects, reducing full-reload cycles during dev.

Bun still ships as a single executable, replacing Node.js, bundler, test runner, and package manager in one Rust-powered binary. It uses JavaScriptCore for execution, keeping startup latency low and memory footprint lean compared to V8-based alternatives.

The catch: Bun’s Node.js compatibility layer, while extensive, still lacks full support for native addons and certain core modules like vm and worker_threads, limiting drop-in use for some legacy or native-heavy Node applications.

Previously in The Times “covered” — Sep 2

Use Cases
  • Migrate a React app from Node.js to Bun for faster startup
  • Run TypeScript 6 monorepos with built-in testing and bundling
  • Audit and install JS dependencies with integrated security scanning

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

More Stories

Pure C engine runs trillion-parameter MoE models on consumer hardware 🔗

Streams experts from disk, treating storage, RAM, and VRAM as a unified inference hierarchy

JustVugg/colibri · C · 29.8k stars 2mo old

JustVugg/colibri is a single-file C inference engine that runs frontier Mixture-of-Experts models — from 744B to 2.8T parameters — without dependencies, by streaming experts directly from storage.

It treats VRAM, RAM, and disk as a unified memory tier, enabling full model execution on heterogeneous hardware like consumer CPUs and GPUs. The engine supports nine model families, including Kimi K3 and DeepSeek V4.1 Flash, with a web dashboard showing live expert routing and token metrics. Built for reproducibility, it avoids silent precision changes and prioritizes semantic correctness over speed guarantees.
The catch: Performance degrades significantly when fast memory is insufficient, and the project explicitly offers no SLA on inference speed, trading predictability for accessibility.

Use Cases
  • Researchers run 2.8T-parameter models on a single workstation
  • Developers deploy vision-enabled MoE models using only CPU and SSD
  • Teams test speculative decoding and expert routing on commodity hardware

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

Clash Verge Rev delivers Tauri-based proxy GUI with Rust performance 🔗

Latest release fixes macOS tray icon glitches, subscription handling, and memory leaks across platforms

clash-verge-rev/clash-verge-rev · Rust · 144.2k stars Est. 2023

Clash Verge Rev offers a cross-platform GUI for Clash.Meta (mihomo) proxy management, built with Tauri and Rust for Windows, macOS, and Linux.

The v2.5.2 update resolves macOS tray icon flickering, improves subscription error messaging for outdated TLS protocols, and fixes memory leaks in frontend connections. It supports custom themes, CSS injection, TUN mode, and WebDAV-backed configuration sync. Despite active development with 424 open issues, the project emphasizes local-only data storage and no telemetry.
The catch: Reliance on Tauri’s webview introduces potential rendering instability under memory pressure, as seen in recent macOS WebView crashes requiring process restarts.

Use Cases
  • Developers testing proxy rules across Windows, macOS, Linux
  • Users managing Clash.Meta subscriptions with custom scripting
  • Teams deploying self-hosted proxy nodes with visual rule editing

Source: clash-verge-rev/clash-verge-rev — based on the README and release notes.

Zen Browser Adds Cross-Device Tab Sync and Shareable Spaces 🔗

Built on Firefox 155/156, it targets productivity with split views and folders.

zen-browser/desktop · C++ · 44.4k stars Est. 2024

Zen Browser, a Firefox-based open-source project in C++, now lets users sync normal tabs, Spaces, and folders across devices via Settings > Sync. Users can also share Spaces, folders, or split views through shareable links from the context menu.

The latest release improves scrolling and video playback performance while reducing battery usage. Despite 685 open issues and no new platform support, the project maintains active development with a commit just 0 days ago.
The catch: It remains dependent on Firefox’s release cycle, limiting independent feature velocity.

Use Cases
  • Developers testing web apps across synchronized workspaces
  • Researchers organizing tabs into shareable topic-based Spaces
  • Writers minimizing distractions with Zen’s focus-oriented interface

Source: zen-browser/desktop — based on the README and release notes.

Quick Hits

openinterpreter Acts as a coding agent for open models like Kimi K3 and GLM 5.3, enabling local, private code generation and editing without API reliance. 68.3k
rtk Reduces LLM token usage by 60–90% on dev commands via a zero-dependency Rust CLI proxy that proxies and optimizes prompts. 80.2k
memos Provides a self-hosted, Markdown-native note-taking tool for instant capture, offering full data control and lightweight performance. 63k
fx Functions as a Unix-like coding agent in Zig, allowing natural language-driven code manipulation directly in the terminal. 2.9k

PypelineC adds automatic pipelining to Python-based hardware description language 🔗

New Python frontend simplifies FPGA design by eliminating blocking/nonblocking assignment complexity

JulianKemmerer/PipelineC · Python · 747 stars Est. 2018

JulianKemmerer/PipelineC introduces Pypeline, a Python-based hardware description language that enables automatic pipelining for pure functions without side effects. The language constructs allow developers to describe combinatorial logic using familiar Python syntax, which the compiler then transforms into pipelined hardware.

This approach mirrors techniques from Intel's variable latency Hyper-Pipelining and Xilinx's retiming, while sharing design goals with Google's XLS Project and DFiantHDL. Pypeline generates human-readable VHDL compatible with all HDL simulators, including Modelsim, and supports instant native Python simulation for rapid verification. For performance-critical workflows, PipelineC offers compiled C-based simulations and seamless integration with software-side C for data unpacking tasks like de/serializing byte arrays from host FPGA interfaces. Conversion to Verilog is included for toolchain compatibility with Verilator. The project provides timing feedback from synthesis reports to help identify critical paths that resist automatic pipelining—a feature particularly useful for newcomers to digital logic design. Despite eight years of development, Pypeline remains a work in progress toward full feature parity with the legacy PipelineC implementation.
The catch: Pypeline lacks certain features present in PipelineC, requiring developers to verify critical functionality before relying on it for complex designs.

Previously in The Times “covered” — Aug 30

Use Cases
  • FPGA designers accelerating data pipelines with Python syntax
  • Hardware beginners learning pipelining without HDL assignment complexity
  • Engineers integrating FPGA accelerators with software-side C data paths

Source: JulianKemmerer/PipelineC — based on the project README.

More Stories

DeskHop v0.78 fixes mouse jumps and USB bugs for seamless KVM switching 🔗

Release adds LED sync and NKRO fallback, improving reliability on Raspberry Pi Pico

hrvach/deskhop · C · 7.9k stars Est. 2023

DeskHop enables instant keyboard/mouse switching between computers via hardware shortcuts or mouse drag, acting as a USB HID host on RP2040. The v0.

78 update resolves spontaneous mouse jumps from composite keyboards and fixes USB disk mount issues, while adding LED synchronization and improved NKRO handling. Built with bundled Pico SDK and TinyUSB, it requires no software on target machines and supports cross-OS switching. The catch: 119 open issues suggest ongoing stability challenges, and reliance on specific hotkeys may disrupt existing workflows.

Use Cases
  • Developers switching between Mac and Linux workstations
  • DIY builders creating low-latency hardware KVMs
  • Raspberry Pi Pico users implementing USB host projects

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

Silhouette Card Maker Adds Borderless Templates for Faster Card Cutting 🔗

Version 3.0.0 enables 3x3 layouts on letter and A4 paper for standard TCG cards

Alan-Cha/silhouette-card-maker · Python · 207 stars Est. 2025

Alan-Cha/silhouette-card-maker now supports borderless templates, allowing users to fit nine standard-sized cards per sheet on letter or A4 paper using Silhouette cutting machines. The update reworks corner extension logic, separating --extend_corners from --extend_edges, and adds bleed control via --extend_bleed.

New plugins for Arkham Horror, Lord of the Rings, and Keyforge LCGs expand compatibility beyond Magic: The Gathering and Pokémon. Built in Python, the tool generates PDF layouts and cutting templates for DIY card proxies and custom games, emphasizing casual play only.
The catch: Proxies must be clearly identifiable and cannot be used in sanctioned tournaments or sold without full copyright ownership, limiting commercial use.

Use Cases
  • Tabletop designers create custom card games using Silhouette cutters
  • Players produce proxies for playtesting Magic: The Gathering decks
  • Hobbyists generate Pokémon or Yu-Gi-Oh! card sheets for casual matches

Source: Alan-Cha/silhouette-card-maker — based on the README and release notes.

P1P2MQTT Bridge Enables Direct Daikin Heat Pump Control via Home Assistant 🔗

Latest release adds automatic climate entity configuration for v0.9.46+

Arnold-n/P1P2MQTT · C++ · 491 stars Est. 2019

The Arnold-n/P1P2MQTT project lets builders monitor and control Daikin/Rotex Altherma heat pumps through the 2-wire P1/P2 thermostat interface, bypassing cloud APIs. It uses an ESP8266 and ATmega328P-based bridge to translate HBS protocols—including Daikin DIII-NET, Mitsubishi M-Net, and others—into MQTT for Home Assistant integration.

Since v0.9.46, climate controls auto-configure in HA, simplifying setup for Altherma full electric, hybrid, Sky Air, and VRV systems. The bridge reads power and data over existing P1/P2 wires, enabling local temperature adjustments and energy monitoring without vendor lock-in.
The catch: 64 open issues suggest unresolved stability or compatibility gaps, particularly for non-Daikin HBS variants like Hitachi H-link or Panasonic SIII-Net.

Use Cases
  • Retrofit Daikin Altherma heat pump with local MQTT control
  • Monitor Mitsubishi M-Net HVAC systems via Home Assistant
  • Reduce energy costs by optimizing Daikin heat pump schedules

Source: Arnold-n/P1P2MQTT — based on the project README.

Quick Hits

HIDman Adapts USB devices for legacy computers, enabling modern peripherals to work on outdated systems without driver headaches. 426
WLED-wemos-shield Provides a universal ESP8266/ESP32 shield for WLED firmware, simplifying smart lighting control with plug-and-play hardware integration. 562
hal Analyzes hardware behavior in real time to detect vulnerabilities and side-channel leaks in embedded systems and chips. 827
IceNav-v3 Delivers offline GPS navigation with multi-GNSS support and OSM maps on ESP32, guiding users without cellular or internet dependency. 456
OpenWeedLocator Uses low-cost image processing to detect weeds in crops and fallow fields, enabling precision agriculture without expensive sensors. 497

Wicked Engine adds HEIC support and clipboard screenshots for faster workflows 🔗

Latest release fixes ocean tiling, paint tool crashes, and MSAA rendering issues

turanszkij/WickedEngine · C++ · 7.2k stars Est. 2015 · Latest: v0.72.113

Wicked Engine, a C++-based 3D engine with DirectX 12, Vulkan, and Metal backends, now supports HEIC image import and automatically copies screenshots to the clipboard. The v0.

72.113 release optimizes GUI rendering by eliminating unnecessary shader switches for image and font renderers. Fixes address a sculpting tool crash when multiple objects target the same mesh, ocean tiling artifacts at distance, and MSAA-related crashes during camera feed rendering. Editor improvements include restored window state on startup and performance gains for metadata visualizer rendering. Built with an entity-component-system architecture, it integrates Jolt Physics, Lua scripting, and glTF/VRM model loading. Developers can compile via Visual Studio, CMake, or Xcode for Windows, Linux, macOS, iOS, Xbox Series X|S, and PlayStation 5 targets. Samples and the Editor_Windows project demonstrate usage as a standalone tool or C++ framework.
The catch: With 115 open issues and a decade-old codebase, stability for cutting-edge ray tracing or large-scale production use remains unproven despite active commits.

Use Cases
  • Build cross-platform 3D editors with Lua scripting
  • Learn modern graphics pipelines using PBR and ray tracing
  • Prototype games with Jolt Physics and glTF/VRM asset loading

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

More Stories

Pumpkin Rust Server Adds Bedrock Support in Latest Dev Release 🔗

Project reaches 0.1.0-dev+26.2-26.45 with dual-edition compatibility and performance tuning

Pumpkin-MC/Pumpkin · Rust · 11.2k stars Est. 2024

Pumpkin-MC/Pumpkin released version 0.1.

0-dev+26.2-26.45, advancing its Rust-based Minecraft server toward 1.0.0. The update adds Bedrock Edition support (marked W.I.P in README) alongside Java Edition, leveraging Rust’s multi-threading for speed. Recent commits optimized release profiles, fixed packet deserialization, and added default gamemode configuration. With 799 forks and active development — last commit 0 days ago — the project targets vanilla mechanics while enabling plugin extensibility. The catch: Bedrock support remains incomplete, and 376 open questions indicate unresolved stability and feature gaps before production use.

Previously in The Times “covered” — Aug 29

Use Cases
  • Host low-latency Java Edition servers with custom plugins
  • Test Bedrock Edition compatibility in dev environments
  • Learn Rust-based game server architecture via contribution

Source: Pumpkin-MC/Pumpkin — based on the README and release notes.

Godot AI v4.1.0 stabilizes MCP server reloads on Windows 🔗

Fixes unblock v3 upgrades with active clients and prevent editor restarts

hi-godot/godot-ai · GDScript · 2.4k stars 5mo old

The latest release improves Windows startup performance and stabilizes plugin reloads by reducing lifecycle overhead. It now activates verified plugins without requiring a Godot editor restart, streamlining updates for users with active MCP clients like Claude Code or Cursor.

These changes address prior upgrade blockers where running AI assistants would halt the update process. The MCP server continues to bridge Godot 4.7+ with AI tools, enabling scene building, node editing, and signal wiring through 46 tools and 120+ operations.
The catch: Open issues remain at 27, with no indication of long-term stability under heavy AI-driven scene generation workloads.

Previously in The Times “covered” — Aug 29

Use Cases
  • Game designers iterate scenes via Claude Code prompts
  • Developers refactor GDScript using AI-assisted node edits
  • Technical artists automate material and particle system configuration

Source: hi-godot/godot-ai — based on the README and release notes.

Quick Hits

flame Build 2D and 3D games in Flutter with high-performance rendering and physics using Flame Engine. 10.8k
Fyrox Create cross-platform 2D and 3D games in Rust with Fyrox’s modern, feature-rich engine and editor. 9.6k
stride Develop stunning cross-platform games in C# with Stride’s powerful rendering, physics, and asset pipeline. 7.8k
GodSVG Edit structured SVG graphics precisely across desktop and web with GodSVG’s intuitive vector tooling. 2.7k
flecs Build high-performance games and simulations in C/C++ using Flecs’ lightning-fast entity-component-system architecture. 8.7k
OpenSpeedy Modify game speed in real time for any browser-based game with OpenSpeedy’s lightweight TypeScript injector. 17.7k
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