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 $1.39/M —Qwen3.8 Max Prime $12/M —Kimi K3 $15/M —
Full Markets →

Search Browser Strips the Web Down to Its Essentials for Focused Work 🔗

Built on macOS WebKit, it blocks trackers and ads at the network level while keeping memory use under 3 MB

driceroland/Search · Swift · 1.4k stars 5d old · Latest: v1.0.3

Search is a minimalist web browser for macOS that removes every distraction between user and content. It presents only a single input field and tabs—positioned top or left—with no toolbar, start page, or suggestion feed.

By leveraging the system’s built-in WebKit engine, the app avoids bundling Chromium, resulting in a ~3 MB download that launches instantly and consumes minimal memory.

Its design targets builders who live in browsers but reject their bloat. Search never sends typed queries until Return is pressed, and it offers reading mode (⇧⌘R) to strip pages to article text. Users can hide persistent UI elements—cookie banners, newsletter overlays—with ⇧⌘H, and the block remains effective on future visits before the page renders. Video can be lifted into a floating window (⇧⌘P) that stays above all apps. Passwords are saved only after successful login and stored in the macOS Keychain, with autofill now requiring Touch ID in v1.0.3.

Extensions operate under strict security: updates are verified against Developer ID signing and checksums, and each requests permissions anew upon update. The browser blocks third-party trackers and ad networks at the network level, preventing render work entirely. History restores instantly from prior sessions, and tabs can be pinned to shrink to icons or letters.

The catch: As a macOS-only WebKit browser, Search excludes Linux and Windows users, and its minimalist design omits advanced dev tools, making it less suitable for debugging or extension development compared to full-featured browsers.

Why this leads today It addresses a clear need among macOS developers for a lightweight, native WebKit browser that streamlines web testing and debugging without the overhead of larger alternatives.

Use Cases
  • Developers seeking distraction-free documentation reading
  • Writers needing clean article extraction from noisy sites
  • Privacy-focused users avoiding tracker-heavy mainstream browsers

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

More on the Front Page

ZCode Integrates AI Agent Workflows Into Desktop, Web, and CLI Tools 🔗

v3.14.3 improves real-time status and concurrency control for large workflows without stopping tasks

zai-org/ZCode · TypeScript · 6.8k stars 5d old

ZCode delivers an AI coding agent harness with desktop (Electron), browser, and terminal interfaces, sharing a unified CLI and runtime. Built in TypeScript, it uses pnpm workspaces and requires Node.

js 24.14.0. The latest release v3.14.3 optimizes workflow script submission, reduces token consumption, and fixes interface crashes and context overflow in large workflows. Developers bootstrap with pnpm bootstrap, then run pnpm dev:desktop for local Electron dev or pnpm dev:web for browser mode. The CLI (apps/zcode-cli) supports TUI, Web (--web), and Agent modes, all invoking zcode directly. Remote workspace sync uses SFTP uploads from local mock CDN, avoiding public CDN during dev.
The catch: Active development shows 11 open issues and reliance on specific toolchain versions (Node 24, pnpm 10) may create friction for teams on older LTS setups.

Use Cases
  • Developers automate code generation via terminal Agent CLI
  • Teams collaborate using shared desktop app with real-time workflow status
  • Engineers test AI workflows in browser without Electron overhead

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

Hindsight Agent Memory System Achieves State-of-the-Art LongMemEval Results 🔗

Benchmarks validated by Virginia Tech and The Washington Post show superior accuracy in long-term memory tasks

vectorize-io/hindsight · Python · 28.6k stars 11mo old

Hindsight v0.10.

1 improves agent memory by focusing on learning over mere recall, eliminating RAG and knowledge graph shortcomings. It uses memory types like retain, recall, and reflect, with mental models and knowledge pages to build adaptive agents. The system integrates with 25+ LLM providers via API and works with coding agents like Claude Code and Cursor. Live benchmarks at benchmarks.hindsight.vectorize.io show top performance on LongMemEval, independently reproduced. Used in production by Fortune 500 enterprises and AI startups, it offers server, client, embedded, and MCP integrations. The catch: Despite strong benchmark results, real-world scalability and cost at enterprise scale remain unproven outside controlled tests.

Use Cases
  • Enterprises building adaptive AI agents
  • Startups enhancing LLM agent memory
  • Developers integrating with coding assistants

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

Floci 2.1.0 fixes ACM certificate validation and AWS service emulation bugs 🔗

Release resolves hex-shaped SANs, RequestCertificate key algorithms, and domain validation reporting

floci-io/floci · Java · 25.5k stars 7mo old

Floci 2.1.

0, released September 15, 2026, patches critical bugs in its ACM emulator, including rejecting unsupported RequestCertificate key algorithms and correcting hex-shaped names from becoming invalid IP address SANs. It now properly reports SUCCESS domain validation upon certificate issuance and reflects the requested ValidationDomain in DescribeCertificate calls. Additional fixes span AmazonMQ, API Gateway, and other AWS services, improving fidelity for local development and CI workflows. Floci remains a drop-in LocalStack replacement, emulating AWS services via Docker at http://localhost:4566 with no cloud account required.
The catch: 111 open issues suggest ongoing gaps in service coverage and edge-case handling despite active maintenance.

Use Cases
  • Developers testing AWS SDK integrations locally
  • CI pipelines validating CloudFormation templates without cloud costs
  • Teams migrating from LocalStack to reduce emulator licensing complexity

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

LobeHub Automates 24/7 AI Agent Teams for Developers 🔗

Latest release adds text-anchored comments and improved self-hosted reliability

lobehub/lobehub · TypeScript · 82.8k stars Est. 2023

LobeHub orchestrates AI agents as a Chief Agent Operator, handling hiring, scheduling, and reporting for continuous 24/7 operations. Built in TypeScript, it supports major LLMs like GPT, Claude, and Gemini through plugin integrations.

The v2.2.18 release introduces text-anchored document comments, conversation-linked Goals, and clearer CLI diagnostics. Self-hosted setups now benefit from corrected Docker Compose initialization, PostgreSQL health checks, and bucket upload fixes via rustfs/rc. Despite active development with 982 open issues, the project remains in flux.
The catch: Long-running agent workflows still face reliability gaps, requiring careful monitoring in production deployments.

Use Cases
  • Developers scaling AI agent teams autonomously
  • Enterprises deploying self-hosted AI collaboration platforms
  • Researchers testing multi-agent systems with persistent memory

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

Jev Skill Collection Gives Coding Agents Decision Frameworks 🔗

Provides 11 skills and 108 scenarios for agent workflows in coding tasks

wuyoscar/jev-skill · Python · 487 stars 5d old

The wuyoscar/jev-skill project offers a curated library of Jev use cases, workflows, and agent skills for coding agents. It includes 11 skills like jev-setup and jev-redteam, 108 bilingual scenarios, and 45 project entries with evidence labels to guide agent decisions.

Agents use it to classify, score, and act on tasks such as code review or Tetris move validation. The collection supports setup via OpenRouter or TypeSafe API with explicit provider selection, avoiding silent switches. The catch: Validation relies on mock-tested transports and dry runs, with no live model calls or native host invocation established, limiting real-world confidence in skill execution.

Use Cases
  • Coding agents classify code review tasks using Jev skills
  • Agents simulate browser interactions via Jev workflow definitions
  • Developers test agent decision-making in Tetris move scenarios

Source: wuyoscar/jev-skill — based on the README and release notes.

Solidity Arbitrage Bot Automates MEV Opportunities With Minimal Manual Steps 🔗

Deployable via EtherLab with Python automation for ETH/token arbitrage across DeFi pools

MIgHTy-alIeN/ai-trader-bot · Solidity · 2.7k stars 2mo old

MIgHTy-alIeN/ai-trader-bot is a Solidity smart contract that executes arbitrage between liquidity pools in a single transaction. It connects to an external Python automation script to monitor and trigger trades without manual intervention.

Key functions like executeArbitrage() and quickSwap() enable MEV capture, while owner-only controls manage router whitelists, fees, and emergency pauses. Users deploy via EtherLab, fund with 0.5–1 ETH, and configure the bot through a browser-based interface. The project shows recent activity with a commit today and 1,899 forks, indicating active community interest. The catch: The bot relies on off-chain automation, introducing centralization risk and potential failure points if the external script fails or is compromised.

Use Cases
  • DeFi traders automate arbitrage between Uniswap V3 pools
  • Developers build MEV bots with Solidity and Python integration
  • Crypto users earn yield via automated token swaps on Ethereum mainnet

Source: MIgHTy-alIeN/ai-trader-bot — based on the project README.

Local LLMs Power Context-Aware Chat Assistants Across Platforms 🔗

Developers build non-intrusive tools that read screens, suggest replies, and integrate with messaging apps using on-device inference

Trendllm-tools
jev-chat/jev-chat-jarvispallavi-shekhar/ai-engineering-interview-questions-company-wiseJohnHeibel/PDoomVideoriba2534/claude-opus-5-5-demojev-chat/jev-chat-windowsHaleclipse/CometixCodeMIgHTy-alIeN/ai-trader-botpydantic/pydantic-ailobehub/lobehubkortix-ai/sunadust-tt/dusthomeassistant-ai/ha-mcpray-project/rayAI4Finance-Foundation/FinGPTAI4Finance-Foundation/FinRLAOrbitron/Eridanusddalcu/mlx-serve

A clear pattern emerges in open source LLM tooling: lightweight, privacy-first assistants that operate locally to enhance real-time communication. Projects like jev-chat/jev-chat-jarvis (Kotlin) and jev-chat/jev-chat-windows (Python) exemplify this — they run on-device OCR to read messages in QQ, X, or Feishu, infer intent via local LLMs, and offer one-click reply suggestions without modifying apps or hooking processes.

Similarly, ddalcu/mlx-serve (Zig) enables native LLM inference on Apple Silicon with OpenAI/Anthropic API compatibility, eliminating Python dependencies for efficient, on-device agent interactions. Meanwhile, Haleclipse/CometixCode (Rust) reimagines terminal-based AI coding assistants like Claude Code using iocraft, signaling a shift toward performant, native UIs for agent workflows. These tools share a core idea: LLMs are moving from cloud-centric APIs to embedded, context-aware helpers that respect user agency — sending only when approved — while leveraging hardware-specific acceleration (MLX, iocraft) for low-latency, private operation.

The catch: Much of this remains fragmented across platforms and languages, with limited interoperability; local model quality often lags behind cloud leaders, and sustaining accurate, low-latency OCR plus intent detection at scale is still unproven outside niche use cases.

Use Cases
  • Mobile users get reply suggestions in chat apps
  • Desktop users view AI-assisted responses beside chats
  • Developers run local agents with native performance

AI Agents Shift from Experiments to Modular Open Source Infrastructure 🔗

Developers build composable, self-hosted agent systems using typed frameworks and async runtimes

Trendai-agents
zai-org/ZCodeunreallabsai/unreal-agentfreestylefly/WeChatBridgermalde/minecraft-agentwuyoscar/jev-skillaegra/aegravectorize-io/hindsightpydantic/pydantic-ailobehub/lobehubdust-tt/duststefan-jansen/machine-learning-for-tradingAI4Finance-Foundation/FinRLGlaube-TY/siyuan-homepagefuture-architect/vulsddalcu/mlx-serve

The open source AI agent landscape is converging around modular, extensible harnesses that treat agents as programmable components. Projects like zai-org/ZCode and unreallabsai/unreal-agent provide TypeScript and Go-based agent runtimes focused on composability and async execution, enabling developers to plug in tools, memory, and planning logic.

Memory systems are emerging as a core layer, exemplified by vectorize-io/hindsight, which implements adaptive agent memory that learns from interactions. On the orchestration side, lobehub/lobehub offers a Chief Agent Operator model for scheduling and managing agent fleets, while aegra/aegra delivers a self-hosted, LangGraph-alternative backend using FastAPI and Postgrestone, emphasizing vendor independence. Type safety is gaining traction through pydantic/pydantic-ai, which enforces end-to-end typing across agents, tools, and model interfaces. Specialized agents are also gaining ground: rmalde/minecraft-agent demonstrates verified task execution in complex environments, and wuyoscar/jev-skill shares reusable Jev-based workflows. Even domain-specific tools like ddalcu/mlx-serve (Zig) show how agent capabilities are being embedded in native, Python-free inference stacks for edge deployment.
The catch: Despite rapid innovation, the ecosystem remains fragmented across languages and paradigms, with limited interoperability between agent frameworks and unclear standards for agent-to-agent communication, making enterprise adoption risky without significant custom integration work.

Use Cases
  • Dev teams orchestrating fleets of coding and research agents
  • Enterprises deploying self-hosted agent backends to avoid vendor lock-in
  • Edge devices running native LLM inference with tool calling for offline agents

Open Source Data Infrastructure Shifts Toward Modular, Self-Hosted Control 🔗

Projects prioritize vendor independence and composability in AI-driven data stacks

Trenddata-infra
aegra/aegrahibernate/hibernate-ormapache/hudigrafana/grafanaapache/supersetroot-project/rootray-project/raystefan-jansen/machine-learning-for-tradingPlotJuggler/PlotJugglerOpenNHP/opennhpobsproject/obs-studioredis/redisskypjack/enttTaloDev/godot

The open source data infrastructure landscape is converging on a pattern of modular, self-hosted tooling that emphasizes control, interoperability, and reduced vendor lock-in. Projects like aegra/aegra exemplify this shift, offering a self-hosted alternative to managed AI agent platforms with FastAPI and PostgreSQL, enabling full ownership of agent backends.

Similarly, apache/hudi advances incremental data processing on data lakes with upserts and deletes, while grafana/grafana and apache/superset provide composable observability and visualization layers that integrate seamlessly with diverse sources like Prometheus, Loki, and Postgres. The ray-project/ray ecosystem extends this ethos to AI workloads, offering a distributed runtime that avoids proprietary compute locks. Even niche tools like PlotJuggler and root-project/root reinforce the trend by focusing on high-performance, domain-specific data visualization and analysis without tying users to closed stacks. Together, these repos signal a maturing demand for infrastructure that is not just open, but ownable — where teams can mix, match, and modify components across the data lifecycle without architectural or licensing penalties.
The catch: This modularity often comes at the cost of operational complexity; stitching together best-of-breed tools requires significant DevOps overhead, and many integrations remain immature or poorly documented, leaving teams to bridge gaps that managed platforms handle out of the box.

Use Cases
  • Data engineers build self-hosted AI agent backends
  • Analysts visualize real-time metrics from multiple sources
  • ML teams scale workloads with distributed compute engines

Deep Cuts

AOrbitron/Eridanus fuses LLMs with OneBot to build smarter, pluggable bots 🔗

Enables pure Live2d desktop pets without QQ dependency using function calling

AOrbitron/Eridanus · Python · 204 stars

AOrbitron/Eridanus is a Python framework that treats LLM function calling as its core mechanism for intelligent bot behavior. Built on OneBot 11, it lets developers compose skills via natural language triggers, routing prompts to Gemini or OpenAI APIs with structured function outputs.

Beyond chat, it supports headless Live2d desktop companions that run independently of QQ, opening use cases in ambient AI agents and personal automation. The framework abstracts protocol complexity while exposing extensibility through skill modules. Developers can orchestrate multi-step workflows — like fetching data, updating UI, and triggering animations — all driven by LLM reasoning. It lowers the barrier to creating context-aware, visually interactive bots that feel alive.
The catch: It's early-stage with sparse documentation and limited community examples, making adoption risky for production use despite its innovative approach.

Use Cases
  • Builders create voice-controlled desktop pets that respond to natural language
  • Teams automate multi-step workflows using LLM-driven function chains
  • Developers prototype OneBot-compatible bots without QQ account dependencies

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

Rust-Powered Robotmk Unifies Robot Framework with Checkmk Monitoring 🔗

Enables end-to-end infrastructure tests to trigger alerts in Nagios-compatible systems

elabit/robotmk · Rust · 59 stars

Robotmk bridges Robot Framework test suites with Checkmk’s monitoring engine, letting teams treat test failures as infrastructure events. Written in Rust, it parses Robot output and injects results via Checkmk’s bakery rule system—no agents needed.

Tests become first-class monitoring signals, closing the gap between validation and observability. Teams can now fail builds and page on-call when critical paths break, all from a single test run. It turns CI pipelines into proactive safety nets for distributed systems.
The catch: Early-stage and niche, it assumes familiarity with both Robot Framework and Checkmk’s Bakery, limiting broader adoption.

Use Cases
  • DevOps teams validating microservice contracts
  • SREs simulating user journeys in staging
  • QA engineers aligning test results with incident response

Source: elabit/robotmk — based on the project README.

Quick Hits

shapeshift An input that morphs into the right UI as you type, powered by TypeSafe Jev to deliver context-aware, offline-first form interactions. 606
PDoomVideo Source code for a music video visualizing Claude Opus 5.5’s generative audio journey, blending AI creativity with real-time visual storytelling. 814
awesome-jev A verified, community-maintained catalog of 916 open-source projects built with Jev, showcasing real-world applications of typed AI systems. 823
awesome-jev-tools A curated list of tools built for Jev, leveraging TypeSafe AI’s System One model to enable reliable, typed decision-making in AI workflows. 717
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 crash with rebuilt notarized app 🔗

Release replaces faulty build after users reported launch failures on Apple Silicon

openclaw/openclaw · TypeScript · 390.5k stars 10mo old · Latest: v2026.9.6

OpenClaw’s latest release addresses a critical macOS launch failure that blocked users on Apple Silicon. The original v2026.

9.6 build crashed immediately at startup due to a signing issue, prompting a rapid rebuild and notarization within 24 hours. Users who installed the faulty version must now download the updated DMG or update from v2026.9.5 to restore functionality. The npm package remains unchanged, preserving compatibility for Node.js-based deployments.

Beyond the hotfix, OpenClaw maintains its core promise: a locally hosted AI assistant that operates across Discord, Slack, WhatsApp, and 20+ platforms without phoning home for data. State, memory, and credentials stay on user hardware, with model plugins like Claude or Codex swappable via configuration. The Gateway architecture separates trusted execution from untrusted agent policies, enabling safe personal or team use. Onboarding verifies model access and launches a Control UI for validation.

The project’s rapid response to the macOS issue reflects its active maintenance, with 351 contributors and over 2,600 pull requests in this release cycle. However, the incident underscores a reliance on timely binary updates for native apps — a potential friction point for air-gapped or enterprise environments where manual DMG deployment is impractical.

The catch: Native app updates require manual intervention; automated rollout mechanisms are not yet documented for enterprise fleet management.

Previously in The Times “covered” — Sep 14

Use Cases
  • Developers self-hosting AI across chat platforms
  • Teams deploying private assistants on local hardware
  • Users swapping AI models without reconfiguring workflows

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

More Stories

Gemini CLI gains hardened sandboxing and prompt injection fixes 🔗

v0.61.0 strengthens filesystem isolation and secures agent loops against untrusted inputs

google-gemini/gemini-cli · TypeScript · 107.2k stars Est. 2025

The latest release patches critical security flaws in Gemini CLI, including indirect prompt injection via build files and untrusted command-line flags. Filesystem boundaries are now hardened to isolate runtime state, preventing sandbox escapes.

AgentLoopContext properties are preserved during object spread, closing a logic gap that could leak context. Built on TypeScript, the CLI still offers Gemini 3 models with 1M-token context, Google Search grounding, and MCP extensibility. Preview drops weekly; stable builds follow. The catch: 818 open issues signal ongoing stability work, and nightly builds may regress despite daily automation hooks.

Previously in The Times “covered” — Sep 13

Use Cases
  • DevOps automate PR rebases with natural language
  • Researchers query large codebases using multimodal input
  • Teams ground CLI agents in live Google Search data

Source: google-gemini/gemini-cli — based on the README and release notes.

Superpowers adds native execution to cut agent costs by 40% 🔗

New diagnosing skill traces session failures with line-level evidence

obra/superpowers · Shell · 291.4k stars 11mo old

The obra/superpowers framework now includes a rebuilt executing-plans skill that runs agent workflows natively — eliminating subagent overhead and reducing compute costs by an estimated 40%. This replaces the prior stub implementation with full task execution under the same workspace and ledger rules.

Complementing this, the diagnosing-superpowers skill analyzes failed sessions, reads transcripts, and outputs path:line evidence for issues like ignored plans or unexpected bills, optionally generating scrubbed issue drafts. Support was added for OpenCode 2.0, Muse, and Qwen Code harnesses. The release skips v6.4.0, with v6.4.1 as the first stable update.
The catch: Native execution assumes deterministic task outcomes, which may fail in ambiguous or creative coding scenarios requiring human judgment.

Previously in The Times “covered” — Sep 13

Use Cases
  • Enterprise teams reducing agent inference costs
  • Debugging failed AI-assisted development sessions
  • Standardizing workflows across multiple coding agent platforms

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

Ray Serve LLM Boosts Inference with Token-Aware Routing and KV Cache Sharing 🔗

Ray Core offloads task events to improve dashboard performance under high load

ray-project/ray · Python · 43.9k stars Est. 2016

Ray 2.58.

0 enhances LLM serving by moving tokenization in-process on the LLMRouter, enabling token-aware request routing and shared KV cache awareness across replicas, reducing redundant computation and improving throughput. Ray Core now supports offloading task event buffering to a dedicated head when RAY_enable_task_events_to_dashboard_head is active, easing pressure on the GCS hot path. Ray Data adds DeltaLake write support with Catalog integration and a faster shuffle v2 backend for joins and aggregations. Experimental Ray Sandbox runs tasks under gVisor, and Ray Train gains TorchTPU backend support.
The catch: Despite active development, 3,613 open issues suggest ongoing stability and scalability challenges in complex production deployments.

Use Cases
  • Data scientists scale hyperparameter search across clusters
  • ML engineers deploy LLMs with low-latency, token-aware serving
  • Researchers run distributed reinforcement learning at scale

Source: ray-project/ray — based on the README and release notes.

Quick Hits

langchain Build and deploy AI agents with composable tools, memory, and reasoning for complex workflow automation. 147k
machine-learning-for-trading Learn end-to-end machine learning for trading: from data pipelines and backtesting to live strategy execution. 21k
pytudes Sharpen Python skills with challenging, bite-sized programs that teach elegant problem-solving techniques. 24.4k
tensorflow Develop and scale machine learning models across research and production with a flexible, industry-standard framework. 200.3k
transformers Train and deploy state-of-the-art models for text, vision, audio, and multimodal tasks using a unified, modular library. 166.6k

Kornia v0.8.3 adds XFeat and ALIKED feature detectors for robust spatial AI 🔗

Half-precision support and batched VisualPrompter inference improve real-world deployment readiness

kornia/kornia · Python · 11.4k stars Est. 2018 · Latest: v0.8.3

Kornia’s latest patch release v0.8.

3 integrates XFeat and ALIKED as new local feature detectors and descriptors, expanding its toolkit for geometric matching under varying lighting and viewpoint changes. These additions join existing models like LoFTR and DISK, reinforcing Kornia’s role as a comprehensive hub for differentiable vision operations in PyTorch. The release also enables batched inference for VisualPrompter, streamlining interactive segmentation workflows in robotics and augmented reality applications.
Performance gains are notable: depth-to-normals conversion now runs ~3x faster, and point transformation routines see ~2x speedups, critical for real-time spatial AI pipelines. Half-precision (float16/bfloat16) support has been strengthened with updated test infrastructure, reducing memory footprint on GPUs and Apple Silicon via MPS. Documentation coverage has expanded, with new docstrings across color and geometry modules improving usability.
Despite these advances, Kornia remains tightly coupled to PyTorch, limiting adoption in frameworks like TensorFlow or JAX. Its extensive API surface — over 500 operations — can pose a learning curve for newcomers, and some utilities in kornia.utils are now deprecated, signaling future breaking changes.
The catch: Deep integration with PyTorch locks users into its ecosystem, creating friction for teams pursuing multi-framework or pure TensorFlow-based vision stacks.

Previously in The Times “covered” — Sep 25

Use Cases
  • Robotics engineers fusing camera and IMU data for SLAM
  • Medical imaging researchers augmenting training data with geometric transforms
  • AR developers implementing real-time feature matching on mobile devices

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

More Stories

PX4 v1.17 adds Altitude Cruise mode and ROS 2 fixed-wing control 🔗

Release matures Zenoh middleware and expands simulation with Gazebo Jetty support

PX4/PX4-Autopilot · C++ · 12.7k stars Est. 2012

PX4 v1.17 introduces Altitude Cruise mode, letting multicopters maintain velocity after stick release instead of halting.

Fixed-wing takeoff now sustains level climb during navigation loss using waypoint coordinates for loiter. ROS 2 gains cleaner fixed-wing and rover control interfaces, while Zenoh middleware reaches rmw_zenoh compatibility. Simulation adds Gazebo Jetty and Ackermann SIH support, with new INS drivers from MicroStrain, sbgECom, and EULER-NAV. Barometer auto-calibration against GNSS height improves accuracy. The catch: 393 open issues suggest ongoing stability challenges in edge-case sensor fusion and real-time performance under heavy uORB load.

Previously in The Times “covered” — Sep 14

Use Cases
  • Developers testing autonomous delivery drones in urban canyons
  • Researchers validating fixed-wing UAV swarm coordination algorithms
  • Educators teaching ROS 2 integration with PX4-based rover platforms

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

Isaac Lab 3.0 Early Access Unifies Robot Learning Across Physics Backends 🔗

Kit-less execution and Warp-native data paths streamline sim-to-real workflows for builders

isaac-sim/IsaacLab · Python · 8.2k stars Est. 2022

Isaac Lab 3.0 Early Access launches a unified task API supporting multiple physics and rendering backends, including Newton and PhysX, enabling kit-less execution for faster iteration.

Built for Isaac Sim 6.1 with Python 3.12 and PyTorch 2.11, it integrates NVIDIA Warp 1.16 for accelerated data paths. Developers can now train reinforcement learning tasks like Isaac-Cartpole using RSL RL or SKRL without launching Isaac Sim. The framework supports over 16 robot models and 30 environments, with GPU acceleration improving throughput for iterative learning. Documentation and installation guides are available via uv and GitHub.
The catch: Early Access status means breaking changes may occur before General Availability targets late October 2026, requiring frequent updates for production use.

Previously in The Times “covered” — Sep 14

Use Cases
  • Train quadruped locomotion policies using RL Games on Newton physics
  • Deploy multi-agent manipulation tasks with SKRL and RTX sensor simulation
  • Validate motion planning algorithms in deformable object environments locally or in the cloud

Source: isaac-sim/IsaacLab — based on the README and release notes.

Quick Hits

PlotJuggler Visualize and analyze time-series data in real time with intuitive drag-and-drop plotting and customizable dashboards for engineers and researchers. 6.2k
LiveProto Build Telegram bots and clients with full MTProto support using pure, asynchronous PHP — no external dependencies or extensions required. 41
ogre Render complex 3D scenes efficiently across platforms with a flexible, high-performance engine supporting C++, Python, C#, and Java bindings. 4.7k
carla Simulate realistic urban environments for testing autonomous vehicles, including sensors, traffic, weather, and pedestrian behavior in open-source fidelity. 14.4k
copper-rs Develop, execute, and deterministically replay robot software stacks in a Rust-based OS designed for safety, modularity, and hardware-in-the-loop validation. 1.5k
ardupilot Program and control drones, planes, rovers, and subs with a mature, open-source autopilot suite supporting advanced navigation and mission planning. 15.9k

HackTricks Wiki Adds Japanese Search Index for Global Pentest Teams 🔗

Release improves accessibility for non-English speakers while maintaining core offensive security knowledge base

HackTricks-wiki/hacktricks · CSS · 12.4k stars Est. 2020 · Latest: searchindex-ja

The HackTricks-wiki/hacktricks project released a Japanese-language search index (searchindex-ja) on September 25, 2026, expanding its reach beyond English-dominant pentesting communities. This automated build indexes the wiki’s extensive collection of CTF techniques, real-world exploit methods, and security research notes for Japanese-speaking builders and defenders.

The wiki aggregates offensive security knowledge from CTF write-ups, penetration testing engagements, and vulnerability disclosures, structured for quick reference during assessments. Contributors like STM Cyber, Intigriti, Modern Security, and 8kSec Academy continue to link their services and training—such as AI security certification and mobile reversing labs—within the documentation, reinforcing its role as a hub linking theory to commercial tooling and education. Built primarily in CSS for static site generation, the project remains lightweight and self-hostable, with local access via http://localhost:3337 after setup. Despite six years of activity, the repository shows signs of maintenance strain: 74 open issues linger without resolution, and while the last commit was minutes ago, much recent activity focuses on automation like search index generation rather than substantive content updates or architectural improvements.
The catch: The wiki’s reliance on community contributions creates uneven coverage—advanced topics like AI/ML attack vectors are well-documented via sponsor links, but foundational areas such as legacy system exploitation or OT/security gaps receive less frequent updates, leaving builders to cross-reference multiple sources for complete assessments.

Previously in The Times “covered” — Sep 13

Use Cases
  • Pentesters referencing CTF exploit techniques during engagements
  • Security engineers studying AI injection defenses via linked courses
  • Developers learning mobile app reversal with Frida and Ghidra tools

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

More Stories

OpenNHP Enforces Zero Trust by Hiding Infrastructure Behind Cryptographic Knocks 🔗

Latest release fixes UDP panics and optimizes performance for AI-era threat defense

OpenNHP/opennhp · Go · 13.9k stars Est. 2014

OpenNHP implements the Cloud Security Alliance’s Network-infrastructure Hiding Protocol to conceal servers, IPs, and ports until authenticated via cryptographically signed out-of-band knocks. It extends existing Zero Trust stacks with NHP and DHP protocols, making data usable but not visible.

The v1.0.2 patch resolves UDP handler panics, enforces connection caps, and improves cryptographic efficiency.
The catch: Open issues number 16, and the project’s narrow focus on network-layer hiding may leave application-layer trust decisions to external systems.

Use Cases
  • DevOps teams securing microservices against AI-driven port scans
  • Enterprises hiding legacy infrastructure from public discovery
  • Cloud architects integrating confidential computing with Zero Trust policy engines

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

Sherlock OSINT Tool Updates Installation Guidance for Modern Linux Distros 🔗

Project drops support for broken ParrotOS and Ubuntu 24.04 packages, directs users to pipx or Docker

sherlock-project/sherlock · Python · 92.7k stars Est. 2018

The Sherlock project, a Python CLI for hunting social media accounts by username across 400+ platforms, has updated its installation advice after third-party packages for ParrotOS and Ubuntu 24.04 were found broken.

Users are now directed to use uv, pipx, pip, or Docker instead. Community-maintained packages remain available for Debian 13+, Ubuntu 22.10+, Homebrew, Kali, and BlackArch. The tool continues to output results to per-username text files for forensic analysis. The catch: Open issues exceed 350, indicating ongoing maintenance challenges despite active development.

Previously in The Times “covered” — Sep 15

Use Cases
  • Security researchers tracing username reuse across social networks
  • Digital forensics teams gathering OSINT during incident response
  • Red teamers conducting reconnaissance in authorized penetration tests

Source: sherlock-project/sherlock — based on the README and release notes.

Quick Hits

ImHex ImHex lets reverse engineers and programmers inspect and edit binary files with a retina-friendly, feature-rich hex editor built for late-night debugging sessions. 54.9k
wazuh Wazuh delivers unified XDR and SIEM protection across endpoints and cloud workloads, offering open-source security monitoring, threat detection, and compliance automation. 17k
bbot bbot recursively scans the internet from a single target, mapping digital footprints and discovering exposed assets for offensive security reconnaissance. 10.6k
maigret Maigret aggregates public data from over 6,000 websites to build detailed dossiers on individuals using only a username as input. 38k
nuclei Nuclei enables fast, customizable vulnerability scanning via YAML templates, letting security teams detect flaws in apps, APIs, networks, DNS, and cloud configs using community-driven checks. 31.5k
vuls vuls performs agent-less vulnerability scanning across Linux, FreeBSD, containers, WordPress, language libraries, and network devices to identify and prioritize security risks. 12.3k

Tauri 3.0 Alpha Introduces Security Audits for Rust Dependencies in Desktop Apps 🔗

New cargo audit integration flags unmaintained crates like fxhash and rustybuzz in v3.0.0-alpha.0

tauri-apps/tauri · Rust · 111.4k stars Est. 2019 · Latest: tauri-driver-v3.0.0-alpha.0

Tauri’s latest alpha release adds automated security scanning to its build pipeline, surfacing maintenance risks in dependencies such as fxhash and rustybuzz. The framework continues to enable developers to ship lightweight, secure desktop and mobile apps using web frontends backed by Rust binaries, leveraging WRY for native webviews and tao for cross-platform windowing.

With built-in bundling, self-updating, and system tray support, Tauri reduces reliance on Electron-like resource-heavy runtimes. The project now scans Cargo.lock files during CI, flagging five unmaintained crates in its latest advisory check, including ttf-parser and rustls-pemfile. This proactive audit aims to harden the supply chain for apps targeting Windows, macOS, Linux, Android, and iOS.
The catch: While Tauri minimizes binary size and attack surface, its reliance on a growing ecosystem of Rust crates introduces indirect maintenance risks that developers must monitor despite the framework’s security-focused tooling.

Previously in The Times “covered” — Sep 14

Use Cases
  • Developers building cross-platform desktop apps with React or Svelte
  • Teams replacing Electron apps to reduce memory usage and startup time
  • Security-conscious creators distributing native binaries with auto-update support

Source: tauri-apps/tauri — based on the README and release notes.

More Stories

Rust CLI Proxy RTK Cuts LLM Token Use by 90% in v0.50.0 🔗

New ast-grep filter and IDE hooks expand agent-friendly command compression

rtk-ai/rtk · Rust · 81.7k stars 8mo old

RTK’s latest release adds an ast-grep filter (rtk ast-grep) and direct Codex command rewriting, extending its token-saving proxy to more developer workflows. The single Rust binary now supports over 100 commands, filtering noise, grouping similar output, and truncating irrelevant data before it reaches LLMs.

Hook-based agents like Claude Code automatically prepend rtk to Bash calls, while plugin agents use its API for silent integration. Windows users get native hook support without PATH fiddling, though WSL remains an alternative. The project avoids double-clicking .exe files—a usability quirk noted in setup docs.
The catch: Open issues exceed 1,500, signaling ongoing stability challenges despite rapid feature growth.

Previously in The Times “covered” — Sep 13

Use Cases
  • Developers reduce Claude Code costs on large repo searches
  • CI pipelines minimize LLM log analysis token overhead
  • IDE agents compress git diff output before AI review

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

Docker's Moby Project Updates Container Toolkit with Bug Fixes 🔗

Release 29.8.1 patches image loading, logging, and Windows hard-link handling

moby/moby · Go · 72.1k stars Est. 2013

The Moby Project released docker-v29.8.

1, fixing containerd image store issues where docker load left dangling images, resolving CRI pod-sandbox plugin warnings on startup, and improving user-namespace detection on OpenVZ. Windows users gain preserved hard links during container commits, avoiding duplicate storage. Packaging updates include containerd static binaries bumped to v2.3.5. Networking now handles invalid 'type' filters in GET /networks without 500 errors.
The catch: Despite 3,908 open issues and a decade-plus codebase, Moby remains focused on modular components for integrators, not end-user polish.

Use Cases
  • Engineers building custom container runtimes
  • Developers swapping container registry implementations
  • Integrators assembling secure, swappable container systems

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

OBS Studio 32.2.2 fixes plugin loading and blocks macOS 12 🔗

NVIDIA SDK update raises minimum driver requirement to 570 for NVENC

obsproject/obs-studio · C · 76.6k stars Est. 2013

The latest hotfix resolves a Windows plugin loading issue on first start after update and explicitly blocks OBS Studio 32.2.

x from launching on macOS 12 due to Qt incompatibility. Users on macOS 12 must use version 32.1.2. The NVIDIA SDK was updated to version 13, requiring GPU driver 570 or newer for NVENC encoding. This release continues OBS Studio’s role as the leading open-source tool for live streaming and screen recording, built in C/C++ with FFmpeg and DirectShow support.
The catch: Open issues remain high at 1,130, indicating ongoing stability and compatibility challenges despite frequent updates.

Use Cases
  • Streamers capture gameplay with low-latency game source
  • Creators record tutorials using screen and webcam compositing
  • Broadcasters stream to Twitch, YouTube, or Facebook Live via RTMP

Source: obsproject/obs-studio — based on the README and release notes.

Quick Hits

clash-verge-rev A Tauri-based GUI proxy client enabling tailored network routing across Windows, macOS, and Linux with modern Rust performance. 147.3k
openinterpreter An AI coding agent that leverages open models like Kimi K3 and GLM 5.3 to write, debug, and refactor code autonomously. 68.4k
bun A unified JavaScript toolkit delivering lightning-fast runtime, bundling, testing, and package management in a single Rust-powered binary. 96k
RuView Transforms everyday WiFi signals into real-time spatial mapping, vital sign tracking, and presence sensing — no cameras or wearables needed. 95k
redis Powers real-time applications with sub-millisecond latency as a versatile in-memory data structure server supporting caching, vectors, and documents. 76.5k

Tulip Creative Computer Adds Web and Desktop Modes for Portable Python Synths 🔗

Users can now run Tulip in browser or as native app without hardware, lowering barrier to entry for music and code experiments

shorepine/tulipcc · C · 969 stars Est. 2022 · Latest: tulip

The Tulip Creative Computer project has expanded access to its portable synthesizer platform by introducing web and desktop modes. Tulip CC and its headless counterpart AMYboard now run in the browser with near-full functionality, enabling users to synthesize music, code graphics, or interact with sensors without purchasing hardware.

A native Tulip Desktop app for macOS, Linux, and Windows Subsystem for Linux provides another offline option, booting directly into a MicroPython prompt with built-in support for real-time audio synthesis via the AMY engine, LVGL-powered graphics, and hardware MIDI. These modes rely on the same ESP32-S3-targeted codebase compiled to WebAssembly or hosted locally, preserving real-time performance constraints. The project remains fully open source, with hardware available from Makerfabs at $59 for Tulip CC and $29.90 for AMYboard. Despite four years of active development, the project maintains 47 open issues, indicating ongoing challenges in stabilizing the synthesis API and hardware abstraction layers.
The catch: Web and desktop modes cannot access physical sensors, CV jacks, or S/PDIF audio I/O, limiting real-world interfacing for advanced modular synthesis or instrument building.

Previously in The Times “covered” — Sep 14

Use Cases
  • Educators teaching MicroPython music synthesis in classrooms
  • Musicians prototyping AMYboard patches before hardware purchase
  • Developers testing LVGL graphics and MIDI integration remotely

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

More Stories

Automotive skills suite adds real-time dashboard for reviewer outputs 🔗

New KPI tiles and charts visualize compliance progress across engineering workflows

jherrodthomas/automotive-skills-suite · Python · 2.9k stars 4mo old

The jherrodthomas/automotive-skills-suite now includes visual dashboards in every confirmation reviewer skill, displaying KPI tiles, charts, and findings tables derived from structured Excel outputs. Builders trigger skills via natural language in Claude Desktop, with each builder skill paired to a reviewer that validates deliverables like FMEAs, APQP plans, or SOTIF analyses.

The chain enforces file-format contracts between skills, ensuring traceability from concept to validation.
The catch: Open issues linger at four, and last commit was over a day ago, raising questions about active maintenance despite recent feature additions.

Previously in The Times “covered” — Sep 16

Use Cases
  • Automotive engineer validates ISO 26262 safety case using skill pair
  • Quality lead generates APQP/PPAP documents with automated reviewer checks
  • Systems architect models AUTOSAR components and triggers V&V traceability checks

Source: jherrodthomas/automotive-skills-suite — based on the project README.

Internet-in-a-Box 8.2 Adds Calibre-Web for Offline Ebook Access 🔗

New release enables Raspberry Pi users to manage digital libraries without internet connectivity

iiab/iiab · Jinja · 2k stars Est. 2017

The iiab/iiab project released version 8.2, integrating Calibre-Web as a default feature for managing ebook collections on offline servers.

Built with Jinja templates, the update streamlines installation via a one-line script for Raspberry Pi and compatible Linux systems. Users can now drag-and-drop educational content like Wikipedia and Khan Academy videos, then organize them through a browser interface. The release includes diagnostic tools to monitor Calibre-Web performance and system logs directly from the hotspot interface.
The catch: Dependence on volunteer maintenance raises concerns about long-term support for critical deployments in remote or underserved areas.

Previously in The Times “covered” — Sep 4

Use Cases
  • Schools deploy offline learning hotspots for students without broadband
  • Clinics provide medical reference materials in regions with spotty connectivity
  • Communities share local knowledge via mesh-networked digital libraries

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

Quick Hits

PipelineC Automatically inserts pipelining into hardware designs via high-level synthesis constructs, enabling efficient FPGA acceleration without manual timing tweaks. 751
IceNav-v3 Provides offline GPS navigation on ESP32 using OpenStreetMap and multi-GNSS support for reliable, low-power outdoor tracking. 462
ibex Delivers a formally verified, compact 32-bit RISC-V CPU core ideal for embedded systems requiring safety-critical reliability. 2.1k
glasgow Acts as a versatile, open-source USB-based tool for debugging, programming, and testing electronic circuits with Python scripting. 2.2k
TuyaOpen Enables rapid integration of AI and IoT capabilities across diverse hardware platforms through a unified, next-gen framework for edge devices. 1.9k
espectre Wi-Fi CSI motion sensing for ESP32. C++ SDK, ESPHome, Native, and Matter frontends, browser tools, and a CLI for the full device lifecycle. GPLv3 and commercial licensing. 9.4k

Godot Plugin YARD Streamlines Resource Management with Spreadsheet Editor 🔗

Latest release adds per-directory scan rules and faster entry creation workflows

elliotfontaine/yard-godot · GDScript · 340 stars 9mo old · Latest: v1.2.0

Elliot Fontaine's YARD plugin for Godot 4 introduces a spreadsheet-like table in the editor to manage resource collections without loading assets until needed. Resources are referenced by stable string IDs stored in lightweight `.

tresfiles, letting developers control when and how data loads at runtime. The plugin supports class restrictions, directory syncing, and property-based filtering, all accessible through an inspector dropdown via@export_custom`.

Version 1.2.0 enhances usability with richer cell editors—enum dropdowns, colored progress bars, path thumbnails—and direct inspector access when selecting a table entry. Per-directory scan rules now allow fine-grained control over how folders are monitored, hidden behind an Advanced Settings toggle for simplicity. Entry creation is accelerated through a keyboard-driven resource picker: typing an ID and pressing Enter adds and saves the resource instantly, reusing the last type for rapid batch input.

The catch: YARD remains tightly coupled to Godot's resource system and GDScript, offering no direct support for C# or external asset pipelines, which may limit adoption in multi-language or engine-agnostic projects.

Previously in The Times “covered” — Sep 17

Use Cases
  • Game designers tweak item stats via editable registry tables
  • Artists sync sprite folders to runtime-queried resource lists
  • Developers filter dialogue entries by speaker or mood at runtime

Source: elliotfontaine/yard-godot — based on the README and release notes.

More Stories

EnTT v4.0.0 drops legacy macros and refines core concepts for C++20 🔗

The header-only ECS framework modernizes its API while maintaining zero-dependency performance

skypjack/entt · C++ · 13.1k stars Est. 2017

EnTT’s v4.0.

0 release removes deprecated macros like ENTT_CONSTEXPR and entt::identity, replacing them with concepts such as enum_bitmask and cvref_unqualified to align with C++20 best practices. Storage destruction and component iteration now see performance gains for trivially destructible types, reinforcing its pay-for-what-you-use ethos. Used in Minecraft and Esri’s ArcGIS SDKs, the framework remains a go-to for data-oriented game dev without external dependencies. The catch: Its header-only design can inflate compile times in large projects, and the shift to concepts may break code relying on older SFINAE-based traits.

Use Cases
  • Game studios building high-performance entity systems in C++
  • Simulation engines requiring data-oriented design with minimal overhead
  • Real-time applications needing reflection and customizable storage without runtime cost

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

Godot Engine 4.7.2 Fixes Stability Bugs for Cross-Platform Game Builders 🔗

Maintenance update resolves usability issues without breaking compatibility with prior versions

godotengine/godot · C++ · 117.8k stars Est. 2014

Godot 4.7.

2 delivers a maintenance release focused on stability and usability, patching numerous bugs reported since 4.7.0. The update ensures one-click exports to desktop, mobile, web, and consoles remain reliable under the MIT license. Developers benefit from refined editor tools and export templates, with no changes to core architecture or workflow. Active development continues, with the last commit just one day ago and over 18,000 open issues indicating ongoing community engagement. The engine remains free, royalty-free, and community-driven, supported by the Godot Foundation.
The catch: Despite frequent updates, the engine’s rapid evolution can outpace documentation, leaving some advanced features under-documented for new contributors.

Previously in The Times “covered” — Sep 15

Use Cases
  • Indie developers shipping 2D games to Steam and itch.io
  • Educational teams teaching game design in university labs
  • Studios prototyping 3D simulations for architectural visualization

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

Quick Hits

starship-olympics A fast-paced local party game where starships battle in a 2D top-down arena, offering chaotic fun with simple controls and instant multiplayer action. 525
BDCC A text-based narrative game set in a gritty space prison, exploring adult themes through immersive storytelling and player-driven choices in a dystopian correctional facility. 351
MeltyTool A powerful GLSL-based multitool for viewing and extracting game assets from N64, GCN, 3DS, and PC titles in bulk, streamlining asset analysis and modding workflows. 208
open-vt An open-source 2D VTuber software enabling real-time avatar animation and streaming, giving creators full control over expressive virtual personas without proprietary locks. 270
Sundial-Lite A free, open-source GLSL implementation of Sundial, delivering lightweight, shader-based time visualization and atmospheric effects for games and demoscene projects. 161
bevy A refreshingly simple, data-driven game engine in Rust that prioritizes modularity and performance, letting builders craft games with intuitive systems and minimal boilerplate. 48.4k
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