Cachety is an open-source tool that performs ID document watermarking entirely within the browser, supporting PDFs and various image formats (including HEIC). All processing occurs in user device memory with no file uploads. It offers customizable watermark text, information redaction, AES-256 encryption, metadata removal, and unique tracking identifiers. The tool is free, requires no account, includes no tracking, uses the AGPL license, and features publicly verifiable code.
Rhci is a lightweight Haskell REPL implemented in Ruby, fully compatible with GHCi semantics, including Hindley-Milner type inference, lazy evaluation, precise output formatting (such as Double display, Int overflow), etc.
THX-01 is an open-source decision model with 322M parameters. According to the author, it surpasses Jev on some benchmarks; it supports local execution on mobile phones and provides a free API (api.hal-x.ai/docs/thx-01); the model weights are released on Hugging Face (huggingface.co/doofz/THX-01), and installation via pip install thx01 is also supported.
LLMxRay is an open-source local large language model observability tool that supports real-time token streaming rendering with confidence coloring, response quality gating (repetition/refusal/garbled text detection), multi-model temperature comparison, surgical benchmarking based on actual logprobs, RAG retrieval visualization, embedding vector analysis, decision model probability output, Ollama multi-protocol response difference observation (native/OpenAI/Anthropic), cache reuse diagnostics (new in v0.6.0, measured prefill speedup of 3.5 times), and a visual tool-building canvas; it can be launched with one click via npx or Docker, depends on local Ollama (0.20+, recommended 0.33.3+), keeps all data local, and requires no cloud services or API keys.
The author's team developed and open-sourced Indexio, a tool written in Rust. By pre-indexing codebases and providing on-demand data via MCP, it replaces high-frequency regex searches. Real-world tests show a 50%–60% reduction in token consumption, with query latency dropping from several seconds to near-instantaneous. Support for symbol lookup is weak; if there is no match, it automatically falls back to regex. The project was inspired by CocoIndex, and the code is hosted on GitHub.
NTXM Browser Mini v0.1.2 is a native Chromium desktop browser with a built-in REST API, supporting Windows x64 and macOS Apple Silicon. It allows direct control of windows, tabs, navigation, and page content reading via HTTP requests to http://127.0.0.1:9090; it supports optional Cloudflare tunnels for remote invocation, requiring no installation of Selenium or Playwright; the project is distributed as closed-source, providing only binary installers and documentation.
INVAR has launched an analytics tool that connects PostHog, Stripe, HubSpot, and Slack. It supports read-only integration within 15 minutes and outputs the first attribution finding within 24 hours. The tool automatically correlates user behavior data (e.g., declining activation rates) with team conversation data (e.g., recurring Salesforce field mapping issues in Slack), ranks root causes using four confidence labels—Correlation/Fact/Hypothesis/Tested—and enables initiating approvals from within Slack, auto-generating Jira/Linear tickets, and tracking changes in target metrics.
Takes for Mac is a free, open-source Mac application that supports teleprompter scrolling synchronized with speech, one-click recording of camera and screen, and saving each recorded clip as a standard folder; it integrates Claude Code (requiring users to have their own Claude plan) to automatically edit clips, generate vertical videos, and draft publishing copy, while also supporting notes on video frames, which the system converts into rules and remembers.
Dropstone CLI is a coding agent tool that can read code repositories, perform cross-file edits, run commands and tests, and self-check before completion. It supports Fast/Pro/Heavy model selection, plan execution, command permission control, skill extensions, custom agents, parallel sub-agents, and MCP tools. It automatically creates Pull Requests after GitHub Issue comments, with all planning features available in the free version.
The author of Xiao Pen released the macOS native menu bar voice assistant Xiao Pen 1.1.0, open-sourced under the MIT license. It supports wake word triggering, streaming Q&A, local volume/reminder execution, and Bluetooth headset shortcut interruption. It can integrate with OpenAI/Anthropic compatible APIs, ChatGPT subscriptions, or local Ollama. Text-to-speech supports system voices or locally running Qwen3-TTS. By default, it uses Qwen for speech recognition, including approximately 2.5 seconds of pre-recording, and can be switched to local recognition. Requires macOS 14+ and Apple Silicon; the installation package is not yet notarized.
OpenClear (formerly PosPay) is an open-source, self-hosted positive pay and ACH anti-fraud screening platform designed for small and medium-sized banks. It supports multi-tenancy, can be extended to other payment networks, and integrates pluggable OCR with an ML-assisted anomaly review feedback loop.
vLLM Semantic Router has launched and open-sourced the Decision 3.0 series of multimodal decision models, including six variants: d3, d3-flash, d3-mini, d3-nano, d3-lite, and d3-edge, with parameter scales ranging from 0.6B to 27B. The weights have been released on Hugging Face. The flagship model d3 is based on Qwen3.8-27B, supports multi-image and video inputs, does not generate long text, and only returns structured judgments and probabilities. It ranks first in both the text leaderboard (64.7 points) and the visual leaderboard (71.6 points) on the Jev Decision Index 0.3.1. Local testing shows median latency of 55 milliseconds for text requests, 273 milliseconds for image-text requests, and 553 milliseconds for 10-second video requests on a single AMD MI325X card. d3-lite can process 1000 PRs on a Mac M5 Pro via Core ML, with a total time of approximately 2 minutes.
Zhipu is conducting gray testing of the GLM-5.3 HighSpeed model within the AI programming tool ZCode, with a measured output speed exceeding 490 tokens/s; this model has a context window of 1 million tokens and a maximum single output of 131072 tokens; previously, GLM-5.1-highspeed achieved 400 tokens/s, with optimizations relying on the three-layer synergy of the TileRT team's inference engine, scheduling system, and infrastructure.
The author deployed Let's Encrypt certificates to the HP Color LaserJet M553 (FutureSmart 3 firmware 2309118_002275) printer, enabling valid certificates for its HTTPS management page and IPPS/AirPrint. Certbot runs in an Alpine LXC container on Proxmox to perform DNS-01 validation, and headless Chromium automates the upload of the .pfx certificate (must be RSA 2048, -legacy encoding), with automatic renewal every 60 days.
Vals AI tested single-agent and team configurations of GPT-6 Sol and Claude Opus 5.5 on Vibe Code Bench, finding that team costs were 1.8–5.1 times those of single agents, with a statistically significant improvement of 7.3 points observed only under medium reasoning intensity for GPT-6 Sol; Anthropic's internal tests also showed that gains from Opus 5.5 multi-agents in knowledge base and theorem proving tasks saturated rapidly; OpenAI researcher Noam Brown and developer Eric Provencher noted that multi-agents primarily improve speed rather than quality, and involve coordination overhead (coordination tax).
Tencent Yuanqi issued a service shutdown announcement on October 8, 2026, stating that due to business adjustments, all services would be fully closed on November 9, 2026, meaning previously created agents will no longer be usable; starting October 29, the creation of new agents and distribution of API Keys will cease, with a data export transition period until January 9, 2027; the announcement also recommends migrating to Tencent Cloud ADP or WeChat Official Account intelligent replies.
SupoClip automatically converts long videos into vertical short videos with subtitles, supporting pasting YouTube links or uploading files; the AI handles transcription, selects high-virality clips of 15–60 seconds, crops to 9:16 aspect ratio, burns in synchronized subtitles, and adds hook titles; it supports manual trimming, splitting, and merging, exporting watermark-free content compatible with TikTok, Reels, and Shorts; it is open-source (AGPL-3.0), available for free self-hosting, and also offers a web version, iPhone app, REST API, and MCP server.
iterm-agents is a small panel tool running in the iTerm2 Toolbelt, used to centrally display session statuses for multiple coding agents such as Claude Code, Codex, Cursor Agent, Devin, and Kimi Code. It supports grouping by Git repository, search, tags, and completion marking; clicking jumps to the corresponding pane. The code is open source on GitHub.
channelerh launched https://bulin47ai.io, a small project that synthesizes user-uploaded photos into Bulin 47 freestyle videos based on the Pixverse Swap model; the entire site is built with TanStack Start, deployed on Cloudflare Workers using D1 and R2, with payments handled via Stripe one-time charges; the author reviewed practical issues such as non-writable disks, CDN transfer storage, preview abuse, and Google login configuration.
Epoch AI released the InnovationEval benchmark test, requiring Claude Fable 5 and GPT-5.6 Sol to independently invent a new training method superior to GRPO under constraints of no internet access and 3000 hours of compute, comparing it with the human-designed SDPO reference method.
Yk developed and released Sealwire, a multi-agent collaboration tool supporting five Harnesses: Claude Code, Codex, Cursor, OpenCode, and Pi. It offers cross-agent review (/review), goal-driven execution (/goal), and task delegation (/delegate). The default is local privacy mode; Cloud mode supports remote approval via mobile PWA with end-to-end encryption. Currently only macOS/Linux are supported. Source code is open under the PolyForm Internal Use + Personal Use license.
Independent developer hanliang introduced that his open-source project MatrixMedia received a CVE-2026-90711 fix PR submitted by the AI bot anupamme, which was successfully merged. He then established the AGENTS.md specification: all PRs must include the four-part structure of phenomenon → root cause → fix approach → verification; AI-generated content must declare the tools used; formatting the entire repository, pure dependency upgrades, and refactoring without behavioral changes are prohibited. The project gained 949 stars and 9 external PRs within 2 months of launch.
BetterWispr is a free, open-source voice dictation tool for macOS 14+ (Apple Silicon recommended). It supports launching local speech recognition in any application via the ⌥ Space shortcut. All models run on-device, requiring no accounts or cloud fallback. Built-in models include Parakeet, Whisper, and Apple native speech models, supporting multilingual recognition, English filler word cleanup, app-level text style settings (e.g., casual/formal), meeting note transcription by speaker role with local summary generation (relying on Apple Intelligence or Ollama), vocabulary learning, history management, and a fully offline workflow.