Agent8 is an open-source project that enables AIX 7.1, HP-UX, and Solaris systems running Java 8 to continue serving as Jenkins nodes by downgrading the official Jenkins agent.jar to Java 8 bytecode and handling compatibility issues such as serialVersionUID, OpenJ9, and signature providers. It supports two installation methods: inbound agent and SSH agent. The project has been verified through CI on Linux HotSpot 8/OpenJ9 8/Java 11/17, and basic self-checks and connections have been completed on real machines for Solaris, HP-UX Itanium, and AIX 7.1; build functionality remains to be confirmed. The project is hosted on GitHub under the MIT license.
fin is a locally running open-source CLI tool that connects bank, credit card, and brokerage accounts via Plaid, with experimental support for parsing Amazon and Costco bills; it stores data in a local DuckDB, offering SQL queries, a browser UI (fin ui), Google Sheets synchronization, and natural language Agent reporting features; all keys are stored in the macOS Keychain, no sensitive information is uploaded, it uses the MIT license, and utilizes the free Plaid trial plan.
pi-cowboy is an open-source multi-agent orchestration tool that supports running isolated sub-agents (such as Codex, Claude Code, OpenCode, Pi, etc.) in parallel within Herdr. Each agent uses a copy-on-write Git worktree and features real-time visualization, manual intervention, templated workflows, and fine-grained concurrency control. The project is licensed under MIT and hosted on GitHub.
MiniMax officially released the M3.1 Flash model, positioned as a lightweight model optimized for Coding/Agent tasks, supporting a 1 million token context window with an output speed of 200+ TPS; input price is 1 yuan per million tokens, output is 4 yuan, and cache read is 0.1 yuan, reducing the comprehensive usage cost for Agentic-type tasks by 70% compared to the previous generation; it performs prominently in tests such as Agents’ Last Exam, Automation-Bench v1.0.6, and SWE Refore Bench, matching or surpassing DeepSeek V4.1 Flash-1017 on multiple metrics.
Xingchen General Artificial Intelligence Lab of China Telecom Artificial Intelligence Technology Co., Ltd. has open-sourced the Xing4.0-29B-Intent model, which supports integrated intent classification and parameter extraction, outputting JSON with Schema constraints; the model is fine-tuned based on the Xingchen 29B base, trained using 2.753 million SFT corpus entries, achieving a Micro-F1 score of 70.27 in zero-shot cold-start evaluation and an Overall total score of 58 on the unified leaderboard; the model, dataset, and inference SDK are available on GitHub, Hugging Face, ModelScope, and MoLe.
Conway Research has released its first personal AI assistant, Underdog, which runs entirely on user devices and supports offline use. Mac/Windows beta versions have been launched, featuring local email and calendar processing, meeting note summarization, local image generation (based on Qwen-Image Flash and Qwen-Image-2.1-Turbo), and Stripe local payments. Its model Saluki 27B is fine-tuned from Qwen3.8 27B, and its inference engine Husky is 4.5 times faster than MLX on MacBook. The product is currently in invite-only testing, initially free with no ads.
Gecit is a Windows-Mac keyboard and mouse sharing tool released by an independent developer. It uses the QUIC protocol within a LAN to enable pointer sliding switching, bidirectional clipboard (including images), media key control, and file drag-and-drop (bidirectional support in the Pro version).
Ranet3 is a high-performance policy-routing IPsec mesh SD-WAN implementation supporting standard protocols (IKEv2/Ed25519, Babel, SRv6) that runs in user space on Linux and macOS; benchmarks from 2026-10-15 show unidirectional throughput of 8.39 Gbit/s (95% of 10 GbE line rate), higher than wireguard-go, Kernel WireGuard, and Kernel IPsec, with differentiated performance in CPU overhead and p99 latency.
According to an Allium report, stablecoin payments reached $482 billion in the first nine months of 2026, up 45% year-on-year, but circulating supply increased only slightly by 0.4% to $305 billion; during the same period, the number of transactions jumped 71% to over 1 billion.
long2alfred released Dragon Card, a local memory card framework based on Flask + SQLite. It supports custom HTML/CSS/JS card faces, full-list scheduling, flexible review goals, and dynamic UI; the code is open source, requires no account or cloud services, and does not provide mobile support. The demo page is at https://dragoncard.top/, and the GitHub address is https://github.com/crazy-dragon/dragon_card.
Four AI agents based on Google Gemini 4 Argon collaborated fully autonomously in a randomly generated world. Without assigned roles, strategic guidance, or game-specific tools, they self-organized tasks, warned of threats, recovered teammates' items, and shared error logs, taking 27 hours and 14 minutes to defeat the Ender Dragon. During this time, they wrote 2,783 JavaScript programs, exchanged 229 messages, and died 605 times. The entire run was billed at $1,620 according to standard Gemini 4 Argon pricing effective from 2027.
SciForge is a local-first deep learning research agent tool that splits the research workflow into four layers: Skill, Tool, Pipeline, and GUI. It includes 24 methodological Skills, 46 deterministic Tools, 7 Pipeline Recipes, and 8 local web pages; the tech stack is Python 3.12 + Typer + FastAPI + native frontend, with default support for deepseek-flash / deepseek-v4-pro and compatibility with the OpenAI API; the project contains 355 offline tests, CI coverage for Python 3.10/3.12, an MIT license, and the open-source address is https://github.com/sddzlsc/SciForge.
CallSitter is an open-source phone gatekeeper tool that connects to real phone lines via Telnyx, driven by OpenAI gpt-live-1 (referred to in the text as GPT-6 Luna) for real-time agent responses; the Jev model classifies call transcription streams (Scam/Sales/Genuine/Urgent), executing differentiated responses according to config/call-flow.yaml; it supports Telegram real-time monitoring and human intervention (SUPERVISOR commands, Take over button), itemized cost statistics (Live per-second fees, Luna tokens, Jev input tokens), and rolling limit controls; the code is Apache-2.0 open source, hosted on GitHub.
TonyGong released DLNA Mirror Flow V1.1, a Windows desktop application that supports real-time pushing of computer screen and audio to HTTP-TS streams and discovers DLNA players on the local area network for mirroring; streaming latency is approximately 3–4 seconds, while slide mode screen capture image casting latency is 1 second; completely free with no ads, providing a Baidu Netdisk download link.
Developers of Spore & Skog released a forest wildlife sound effect synthesis system that runs in real time in the browser, written entirely in Rust, based on CPAL audio output and Bevy game engine integration; the system does not use pre-recorded audio clips but dynamically generates unique calls for each creature, relying on three core tools: noise sources, state-variable filters, and time envelopes, supports weather-driven call weight scheduling, and open-sources all code (MIT license).
LayoutKeeper releases versions 0.6.7 and 0.6.8, supporting scene differentiation by monitor EDID and saving/restoring window layouts across desktops. Version 0.6.7 rewrites the auto-restore logic and adapts to Chrome multi-profiles; version 0.6.8 switches to system notifications and removes erroneous features such as deleting desktops upon cable disconnection. It only supports a single external display; cross-desktop restoration requires macOS 26+ and relies on SkyLight private interfaces.
VAST has open-sourced the real-time world model EdenLive, which supports continuous generation at over 24 FPS at 720P resolution on a single consumer-grade RTX 5090 GPU. It features arbitrary reference input, real-time rendering in any art style, and minute-level sustained operation capabilities; model weights, inference code, technical reports, and an end-to-end Demo have all been made public.
In 2026, WeChat launched the native AI agent Xiaowei, with an entry point marked by a green dot in the top-left corner of the first screen. It supports invocation from 16 locations including Official Accounts, Moments, chat images, and Scan QR Code, and can execute system-level tasks such as sending messages, modifying settings, and calling mini programs. Based on the WeLM model, Xiaowei focuses on privacy protection and cost control. Unlike Yuanbao's external friend form, it does not directly take over payment or publishing actions; all transactions and social expressions require final user confirmation.
jyeric released an open-source Android app called "Don't Track," which uses accessibility permissions to automatically parse and remove tracking parameters from shared links after users click the share button or a floating window. It supports card-style sharing on WeChat and QQ; rules are maintained separately, with current WeChat sharing limited to 100 times/day; power consumption is approximately 0.1%/day.
Portlivo is a native macOS port management tool that displays all running development servers and ports on the current Mac, annotating the associated project, startup command, runtime duration, and response status. It supports one-click opening, restarting, terminating, viewing logs, or sharing any service; no account is required, and it operates on a one-time payment model.
Researchers Pan et al. proposed the Taste-Bench benchmark to evaluate AI agents' ability to make high-quality intermediate decisions (i.e., 'taste') in long-term tasks, filling the gap left by existing benchmarks that focus only on final outcomes while ignoring process quality.