This week's ten new repositories do not form one product category. They include an AI news-site framework, agents that work with local games or browsers, a persistent AI workspace, and focused tools for video, writing, floor plans, hardware, and visual explanations. The notes below stick to capabilities, setup paths, and limits stated in each README.
What caught people's attention
The projects cluster around AI agents and workspaces, local creative tools, and controlled browser or device access. KKKKhazix/AIHOT led the snapshot with 5,841 Stars, while the Top 10 held about 23K in total. The list ranks public repositories created during the previous complete calendar week by absolute Stars; it is not GitHub's official Trending list and says nothing by itself about quality or security.
The observation window is 2026-09-28 through 2026-10-04. Star counts are collection-time snapshots, and no listed project was cloned, installed, executed, or security-audited here.
Top 10 projects
1. KKKKhazix/AIHOT
AIHOT is an engine and framework for an industry news site. It collects material from RSS, web lists, JSON APIs, X accounts, WeChat public accounts, and custom pushes, then prefilters, scores items twice independently, writes Chinese headlines and summaries, clusters stories into events, and produces daily, weekly, and monthly briefings. The README also lists RSS, public API, MCP, Agent Markdown, llms.txt, source management, evaluation, and alerting.
Core functionality: It turns multi-source news into events and ranks them by independent-source coverage. Prompts and selection thresholds are public and editable, so the framework can be adapted to legal, finance, or another industry.
Usage and boundaries: The README requires Docker, Node.js 24, PostgreSQL 17, and an OpenAI-compatible model API key. Configure it with node scripts/init-env.ts, run docker compose up -d --build, and open http://localhost:3000; the admin area is /admin. Non-Docker Windows deployment uses WSL2. The repository does not include AIHOT's real source list or operating data, and its name and logo are not granted by the MIT code license.
Attention snapshot: 5,841 Stars · 1,488 forks · about 932.6 Stars/day over its lifetime at collection · TypeScript · MIT. README reviewed at 9acad0c3d768.
2. rehan-remade/universal-modder
universal-modder gives Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, and OpenCode game-modding skills, a um CLI, and a shared knowledge base. Its loop identifies the engine, reads real code, backs up saves, builds code and assets, verifies changes in the running game, records video, and writes a field note for the next agent.
Core functionality: um scan fingerprints Steam, Epic, and Xbox games; um fal, um comfy, um sprite, and um render3d handle assets; um win captures, records, and sends input only to the game; um backup manages saves; and um kb manages field notes.
Usage and boundaries: Install through the relevant agent plugin, npx skills add, or a clone. Python 3.10+, ffmpeg, and Blender for 3D rendering are also required; asset generation can use a fal API key or a local ComfyUI server. The README forbids injection into anti-cheat online games, cheating other players, and bypassing anti-cheat, DRM, or ownership checks. It backs up saves and asks before driving input, installing loaders into game folders, or publishing a PR.
Attention snapshot: 3,339 Stars · 291 forks · about 662.4 Stars/day at collection · Python · MIT. README reviewed at a93ed669e7b6.
3. CopilotKit/OpenDots
OpenDots is a self-hostable AI workspace template. Users create Spaces, pages, and specialist Dots for research, writing, and revision. A Dot can have isolated browser, file, and shell permissions, and a draft can pause for human approval before being saved to an authorized Space. The template also includes calls, background work, and Slack integration.
Core functionality: It includes a page editor with autosave and revision checks, CopilotKit Threads, OpenBot computers, per-Dot tool permissions, background turns, WebRTC speech, and a Slack Channels SDK connection.
Usage and boundaries: The README requires Node.js 24 and npm: run npm ci, copy .env.example, run npm run dev, and open http://127.0.0.1:5173. This is an early template, not a hosted product. Live models, speech, Slack, OpenBot, and cloud schedules need separate service configuration; Slack and spoken compute delegation still need connected-service verification. Shared editing, invitations, file uploads, interactive embeds, and a full goal or event-trigger system are not included.
Attention snapshot: 3,280 Stars · 433 forks · about 593 Stars/day at collection · TypeScript · MIT. README reviewed at c2569bb6a13a2.
4. feder-cr/dots
dots puts an AI agent inside a modified Firefox browser and focuses on what a website actually sees. It keeps screen, fonts, GPU, timezone, and language consistent for each seed, supports persistent profiles and proxy exits, and can expose the browser to Claude Code, Codex, Gemini CLI, or another MCP client through invisible_playwright_mcp.
Core functionality: The agent uses a real browser for navigation, clicking, and typing. The README says it exposes no WebDriver flag, DevTools Protocol, or automation globals to the page, and that --model switches the OpenRouter model.
Usage and boundaries: On Windows or Linux, install uv and run uvx --from git+https://github.com/feder-cr/dots dots --openrouter-key ..., then open http://127.0.0.1:8765. An OpenRouter key is required. The README says the project is not affiliated with OpenAI; because the browser can carry login state, a proxy, and a stable identity, use it only with clear authorization for the target site and account.
Attention snapshot: 2,606 Stars · 437 forks · about 492.9 Stars/day at collection · Python · MIT. README reviewed at 2ff9848c1c8a.
5. nanaism/yomiyasu
yomiyasu is an Agent Skill for polishing Japanese text in Codex, Claude Code, Cursor, and similar environments. It aims to preserve meaning while checking subject-predicate links, modifiers, conditions, exceptions, tone, and information density. Its target material includes technical articles, design documents, specifications, PR copy, and internal reports.
Core functionality: yomiyasu_lint.py checks AI-like phrasing and Markdown emphasis structure, while yomiyasu_diff.py checks meaning and sentence-function changes between the source and revision. The skill supports tech, business, and essay domains.
Usage and boundaries: Install with npx skills add nanaism/yomiyasu, npx openskills install nanaism/yomiyasu, or the Claude Code plugin; a skill-only ZIP is also provided. The README warns that similar Japanese proofreading skills can interfere when enabled together. Its findings are review prompts, not a guarantee of meaning preservation or usability, and referenced external material keeps its own license.
Attention snapshot: 1,425 Stars · 32 forks · about 299.6 Stars/day at collection · Python · MIT. README reviewed at 4075fc34fb03.
6. wy51ai/floorplan-3d
floorplan-3d is a single-file, front-end floor-plan and renovation designer. It places furniture, measures rooms, and edits non-load-bearing walls in an SVG 2D plan, then synchronizes the design with a Three.js 3D scene for overhead viewing or first-person exploration.
Core functionality: It provides more than 60 furniture items, wall snapping, area statistics, flooring materials and waste-adjusted cost estimates, undo/redo, local autosave, PNG and JSON import/export, and furniture editing in 3D.
Usage and boundaries: Clone it and open index.html, or run python3 -m http.server 8000; there is no build step. Three.js loads from jsDelivr, so the first 3D view needs a network connection. Data lives in browser localStorage, while rooms and the furniture library are defined in index.html. The project is MIT-licensed.
Attention snapshot: 1,399 Stars · 287 forks · about 228.4 Stars/day at collection · HTML · MIT. README reviewed at 730ec0923a01.
7. firelex/jeff
Jeff is a 0.8B local model that returns probabilities, an answer, and confidence for a set of options in one forward pass. It is designed to put fast decisions such as inbox guarding, triage, support intents, tool choice, spam, and legal-clause classification in front of a larger model, forwarding uncertain cases to a 27B model.
Core functionality: It ships nine LoRA adapters, an HTTP server, Python and TypeScript clients, choice, yes/no, and score questions, and an adapter kit for custom tasks. Adapters can be switched per request.
Usage and boundaries: The README uses uv sync, downloads the v1.2 base model and selected adapters, then starts jeff-serve; CUDA and Apple-silicon MLX extras are available. Jeff is a classifier, not a planner or multi-step reasoning model, and it supports English text only. Qwen models accept up to 254 options, while v1.2 adapters only work with Jeff-Qwen3.5-0.8B v1.2. Code is MIT, model weights are Apache 2.0, and training data is not released under one project license.
Attention snapshot: 1,374 Stars · 64 forks · about 209.1 Stars/day at collection · Python · MIT. README reviewed at d0173b4ee317.
8. edenfunf/reelmimic
ReelMimic is a local video-style recreation tool. Give it a video file, phone recording, or YouTube link; it analyzes editing rhythm, shot length, transitions, framing, colors, and camera moves, creates a reviewable plan, then has several AI agents produce and review shots in parallel.
Core functionality: It supports up to six parallel agents, before-and-after evidence for fixes, time-coded video comments, Markdown-defined styles, and seven 2D engines: vector/motion graphics, watercolor, crayon, pixel art, paper cut-out, whiteboard, and anime cel.
Usage and boundaries: It requires Node.js 22.18+, Python 3.10+, FFmpeg, Chrome, and a logged-in Claude Code or Codex CLI. Run install.sh or Windows install.bat, then start.sh or start.bat, and open http://localhost:4318. A 30–60 second video usually takes 1–3.5 hours; the project is tested mostly on Windows and makes 2D animation rather than live-action footage. It does not reuse reference footage, characters, logos, or assets, but users remain responsible for rights.
Attention snapshot: 1,278 Stars · 156 forks · about 198.2 Stars/day at collection · JavaScript · MIT. README reviewed at eeb8cce6bf27.
9. facebookincubator/muse-gadget-sdk
Muse Gadgets provides ESP32 and Linux/Raspberry Pi device SDKs for connecting displays, audio, sensors, and actuators to Muse. Devices pair through the iOS or Android Muse app after Developer mode is enabled, using the MuseGadget prefix in Settings > Devices.
Core functionality: The ESP32 SDK supports development boards with displays, audio, and sensors. The Linux SDK turns a Raspberry Pi or Linux device into a Muse gadget that can run custom commands, including system administration or Home Assistant tasks.
Usage and boundaries: Before flashing or pairing, obtain an SDK token and review the Gadget SDK Terms; every gadget needs a token. The README presents this as experimental hardware and warns about bricked boards, voided warranties, brownouts, and cost. The main project is Apache-2.0, but third-party files, build-time dependencies, and the Jollybot avatar keep separate licenses.
Attention snapshot: 1,259 Stars · 232 forks · about 506.8 Stars/day at collection · C · Apache-2.0. README reviewed at 3229892e93c1.
10. QingYunA/answer-me-with-html
Answer me with HTML is an Agent Skill: the agent writes a short Markdown draft, while its CLI turns that draft into a readable HTML page with flowcharts, sequence diagrams, trees, timelines, tables, and annotations. Each page is a single HTML file that opens offline and is easy to share; the same tool can also make narrated explainer videos.
Core functionality: It provides blueprint and shadcn themes, light and dark modes, computed layout, embedded Markdown source, a STE-style writing check, and am render, am patch, am lint, and am config. Video mode can use ElevenLabs, a system voice, or an OpenAI-compatible local TTS service.
Usage and boundaries: Node.js 20+ is required; install with npx skills add QingYunA/answer-me-with-html or the relevant agent plugin, with no npm install step. Video export additionally needs Chrome, ffmpeg, and Node.js 22+. Pages default to ~/.answer-me-with-html/pages/. The README's benchmark reports less waiting but roughly the same cost; strict writing checks cannot replace human review of the content.
Attention snapshot: 1,161 Stars · 81 forks · about 409.7 Stars/day at collection · JavaScript · MIT. README reviewed at 1ca0988ac179.
How to read this list
This list compares only public repositories created during the previous complete calendar week. Stars, forks, creation dates, and languages are GitHub Search API snapshots. Setup requirements, supported platforms, and security boundaries were checked against each repository's pinned README revision. Before adoption, review the license, dependencies, issues, security history, and current maintenance.
Source: GitHub Search API · Collected 2026-10-05 (Asia/Singapore)