Muster is an open-source macOS menu bar app for receiving and responding to agent notifications from Claude Code, Codex, and Cursor in real time. It supports global shortcuts such as ⌥↩ to allow, ⌥⌫ to deny, ⌥↓ to expand details, and ⌥1–⌥9 to select options. All communications are limited to 127.0.0.1, with no code or data uploads. It requires macOS 14+, has a size of 2 MB, uses approximately 20 MB of memory, is licensed under MIT, and is open-sourced on GitHub.
AWS Anchor Browser has officially integrated with Amazon Bedrock AgentCore Payments, allowing AI agents to autonomously handle HTTP 402 Payment Required responses within browser sessions; by enabling the session.x402 switch, agents can complete payment signing, retries, and content retrieval without modifying Playwright/Puppeteer code; payments are managed by AgentCore regarding wallets, budgets, and credentials, ensuring private keys do not enter the runtime environment.
Tern is a native macOS SSH client written in Rust and GPUI, with a size of only 76 MB (compared to Tabby's 580 MB). It supports saving hosts, SFTP, split screens, code snippets, port forwarding, and two-factor login. Passwords are sourced from SSH keys, password manager commands like gopass, or age-encrypted vaults, with manual input supported via ⌘\. Configuration and data are synchronized through private Git repositories, fully encrypted locally throughout the process. The app is signed and notarized, installable via Homebrew or DMG. The current version is 0.1.3, with Windows/Linux preview versions available on the Releases page.
BitRun v0.2 is a zero-compilation, single-binary local inference tool that supports running Microsoft BitNet b1.58 and other 1.58-bit ternary LLMs (such as bitnet-70m, bitnet-2b, llama3-8b) on CPUs across macOS, Windows, and Linux without requiring a C++ compiler or Docker; it provides model pulling (bitrun pull), interactive chat (bitrun run), OpenAI-compatible REST service (bitrun serve), and Modelfile custom persona features; real-world tests show pure CPU execution on Apple M4 Pro, Intel i5-12600K, and WSL2 with 0% GPU utilization; the project is based on llama.cpp and BitNet.cpp, is MIT open-source, and code is hosted on GitHub.
Microsoft announced the establishment of Xbox Labs, a small team comprising game developers, engineers, and scientists, aimed at exploring practices such as AI-model-supported game decompilation, cross-platform recompilation, and player co-creation; Xbox CEO Asha Sharma noted that games and programming share characteristics suitable for agents, and that players spend over 72000000000 hours gaming annually, approximately five times the global annual coding hours; the team will open-source tools and share results, while emphasizing its commitment to IP and community responsibility.
PanamaKVM is an experimental, embeddable hypervisor written in Java 25 that directly calls Linux KVM via the Foreign Function and Memory API. It supports direct booting of x86 Linux bzImage kernels and exposes a serial console without requiring project-specific JNI libraries. Currently, it can boot an Alpine kernel into a BusyBox smoke test environment or an interactive shell. It implements only single vCPU and contiguous memory from 64 MiB to 3 GiB, with no support yet for block devices, networking, SMP, firmware booting, or production-grade isolation. The project provides a portable emulator, Gradle build scripts, JUnit tests, JaCoCo coverage reports, and CLI tools. It is licensed under MIT, with source code hosted on GitHub.
Vectorize has released Hindsight 1.0 and Hindsight Memlake. The former is an open-source AI agent memory system that ranks first on all datasets in the Agent Memory Benchmark; the latter is a storage engine written in Rust, capable of managing 100 million memories in object storage with p99 latency under 100 ms, requiring no database, consensus cluster, or master node.
Ronan Gaillard released Factory DS, a Factorio-like factory building game developed specifically for the Nintendo DS (2004 hardware, 4 MB RAM), supporting dual-screen operation, French and English interfaces, with a ROM size of 1.9 MB.
Hatoba is an open-source desktop SSH client that supports AI assistants (requiring users to configure their own OpenAI or Anthropic API), end-to-end encrypted synchronization (via user-deployed Cloudflare Worker + D1 database), offline usage, and a Windows-first native UI.
a16z investor Collin McCune analyzes the substantive progress in U.S. crypto regulation over the past three years: The federal stablecoin law (GENIUS) has been enacted; the IRS withdrew its DeFi broker rule, the SEC rescinded SAB 121, and FinCEN withdrew its unhosted wallet rule; the SEC and CFTC have released frameworks, proposals, and guidance on token issuance, asset custody, and on-chain stock trading; and the House passed two bipartisan comprehensive market structure bills, FIT21 (2024) and Clarity (2025).
GitHub released a fine-tuned ModernBERT classifier model dedicated to secret detection, replacing the previous LLM pipeline. In secret scanning, it achieved an 11.5% higher repository-owner-weighted acceptance rate, an 86% lower median scan execution time, and over 98% reduction in service costs. The model works with regular expressions to form a two-stage pipeline: the first stage uses regex with high recall to generate candidate secrets, and the second stage uses ModernBERT combined with file paths, variable names, and context to determine whether they are real secrets.
Deslop is an open-source TypeScript Lint rule set based on dmmulroy/anti-slop with extended new rules, all published under the deslop/* namespace; it supports npm installation and is compatible with the Oxlint JS plugin runtime on Bun/Deno/Node.js; it includes recommended rules for type safety, code quality, and testing, as well as several opt-in rules such as no-chained-type-assertions, no-array-filter-map, and no-module-mocking; the project uses the MIT license.
AURORA Lifeline converts official cyclone bulletins from the India Meteorological Department (IMD) into district-level emergency action plans, identifying which Primary Health Centers (PHCs), hospitals, and villages will be cut off due to road disruptions, assessing the probability and duration of disconnection, and providing recommendations for prioritizing material transfers; Gemini parses the bulletins, calls over 1000 WeatherNext and ECMWF storm track forecasts to cover each road, and automatically fills in all numerical values via code; officials can approve all outputs; the project is open-source but has not released an official version.
AWS published a tutorial introducing how to automate the activation of AWS Marketplace Private Pricing Offerings (MPPO) for Amazon Bedrock using managed entitlements.
Liquid AI released LFM2.5-24B-A2B, a MoE reasoning model with 24B parameters but only 2B active parameters per token, designed specifically for private edge-side agents; it supports a 131K context window and a 128K vocabulary, explicitly generates chain-of-thought, and leads competitors such as Qwen3.6-35B-A3B on 5 tool-use benchmarks including LiveClaw; it offers multi-backend support via llama.cpp, MLX, vLLM, etc., achieving measured speeds of 184 tok/s (on M5 Max) and 21K tok/s (on B300 GPU); the model has been open-sourced on Hugging Face.
Google Play has introduced a new feature that allows users to create private vibe-coded apps via Google AI Studio and run them directly on their own Android devices using standard Google accounts, without requiring public listing. These apps do not appear in search results or list pages and are restricted to use by the creator only. Initially, this is limited to users aged 18 and older in the United States; subsequent updates will support sharing with trusted small groups and provide a path to convert the app to a public listing.
Lowlight is a macOS application that allows users to adjust the keyboard backlight brightness of Apple Silicon MacBooks below the system's default minimum value. It provides menu bar slider control, supporting 2–20 levels of "below minimum" steps (default 10 levels); implemented via private CoreBrightness API, requires macOS 26, tested on M5 MacBook Air, licensed under MIT.
Satlyt has released software capable of running AI models on satellites. Its code will be launched aboard a SpaceX rocket, conducting its first in-orbit validation jointly with the Google Project Suncatcher space data center prototype. Previously, the software deployed Google DeepMind's Gemma model on Momentus satellites, reducing data transmission caused by software errors by over 60%. This mission is supported by hardware from Indian startup TakeMe2Space, serving three clients: NASA, Stellerian, and itself.
Trezor announced that its authorized dealer qualification will terminate on January 1, 2027, and it will no longer supply new units, stating this is a preventive measure; there is currently no evidence indicating its devices have been tampered with. This decision follows Ledger's discovery of an LTE implant module embedded in hardware wallets sold by Southeast Asian distributor CryptoBilis, which was designed to steal recovery phrases. Ledger has not verified the suspected theft amount exceeding $86,000,000 traced by Specter.
Orca Security, in collaboration with AWS, launched Discovery AI, a natural language security query agent running on Amazon Bedrock that converts user questions into structured Discovery filters supported by its Sonar query system.
Google DeepMind released Gemini 4 Argon, a general-purpose physical reasoning model designed specifically for robotics, achieving 87% accuracy on the Embodied Reasoning (ERQA) benchmark through pure reasoning.
Google DeepMind has released Gemini 4 Argon, available to developers, enterprises, researchers, and consumers via API, Google Antigravity, the Gemini app, and Search AI Mode.