Every find: Repos
New skills, plugins, MCP servers, repos and small products, described only from their own README or landing page.
staryone/motlab
A video engine where each scene is a canvas page and an agent skill writes the scenes from your logos, footage, fonts and a brief. It renders to MP4 up to 4K with no AI image or video model, and suits ad producers who want revisions by frame notes.
Video and imageFor creators, developers
Usable todayRun npm install -g @motlab/cli && npx @motlab/cli setup, then hand your agent a folder of references and a direction.
Source: GitHub new repos한국어
YKLovePython/halfsize
A local video compressor that hits a target file size. It computes the bitrate, picks a resolution that fits, encodes in two passes, then reports SSIM and VMAF so you can see whether the video still looks the same.
Usable todayRun npx halfsize wedding.mp4 --to 25mb, replacing the file name and size with your own.
Source: GitHub new repos한국어
jackulau/coding.sh
A local recorder for AI agent sessions. It logs every prompt, model call, tool call, file change and cost in a local SQLite index, with a viewer, a CLI and MCP access. It is MIT licensed, needs no account and has no runtime dependencies.
CodingFor developers
Usable todayNeeds Node 22.16 or newer. Run npx coding.sh, start a new agent session, then run coding.sh to open the viewer.
Source: GitHub new repos한국어
DIYgod/pinky
A desktop app that lets Claude Code, Claude Desktop, Codex, Cursor or a script operate a real iPhone or Android phone over MCP. You watch, approve risky steps and can take over. It asks before payments by default.
Skills and MCPFor developers, creators
Usable todayDownload the app, then connect Claude Code or another agent under Settings > Agents.
Source: GitHub new repos한국어
fendouai/ViralForge
Turns a pasted link or local file into a Viral DNA report covering hook, shot structure, beat grid and caption style. Add your product and script to get an editable VFML project and 10 costed variants for A/B testing.
Video and imageFor creators, business and marketing
Partly usableSee the repo's Quickstart section. The install commands were not available in the source text. The v0.1 studio ships with Chinese-language demo fixtures.
Source: GitHub new repos한국어
IluvatarLabs/sitrep
A live dashboard of every Claude Code and Codex session across machines, showing plan, model, context fill, git state, cost and failures. Pending permissions and questions sit at the top, and you can answer them from the dashboard.
CodingFor developers
Usable todayClone the IluvatarLabs/sitrep repo, then run npm ci and npm run dev. It needs git and Node 24, and the dashboard runs on port 5173.
Source: GitHub new repos한국어
JuneYaooo/rea-video
Breaks a video down shot by shot and judges whether it was filmed, AI-generated or coded, plus how to remake it. Each breakdown gives a report for you and a layered VIDEO.md for your AI agent.
Video and imageFor creators, developers
Usable todayNeeds ffmpeg and Python 3.10+. In Claude Code or Codex, ask it to install JuneYaooo/rea-video from GitHub, then break down a video path. A Gemini or Doubao model is optional.
Source: GitHub new repos한국어
PrunaAI/P-H3-Image-Comfy
ComfyUI files for P-H3-Image, a text-to-image and image-editing model based on MiniMax-H3, in bf16, int8 and w4a8. The smallest set is a 20 GB download that peaks at 21 GB of VRAM at 8 MP, against 143 GB of weights for MiniMax-H3.
Usable todayUpdate ComfyUI to the latest nightly with git pull, then download one file per component into ComfyUI/models with hf download PrunaAI/P-H3-Image-comfy --include, as the model card shows.
Source: Hugging Face trending한국어
nodo-labs-uk/attention-queue (aq)
A terminal UI that gathers every Claude Code session needing input (diffs, edit approvals, permission prompts, questions) into one queue, shown one card at a time. Written in Rust, MIT licensed, for Linux and tmux.
CodingFor developers
Usable todayDownload aq-x86_64-linux.tar.gz from the latest release or run cargo install --path ., then merge the hooks from contrib/settings-hooks.json into ~/.claude/settings.json.
Source: Show HN (AI)한국어
wquguru/fable52-fps-repro
Four code-only Fable 5.2 experiments that produce single-file HTML 3D demos, with prompts, run scripts and billing data published. The FPS demo is 131 KB, generated in 73 turns for about $15.96 in API cost.
CodingFor developers
Usable todayClone the repo, run scripts/run.sh 01-audio prompts/01-audio.md 5m 12, then python3 scripts/metrics.py. Needs Claude Code 2.1.242 or newer and access to claude-fable-5-2.
Source: GitHub new repos한국어
Comfy-Org/MiniMax-H3-Image
Repackaged MiniMax H3 Image model files for ComfyUI with Image Edit and Text to Image workflows. Its license differs from the original MiniMax H3.
Partly usableDownload the listed files into the ComfyUI model folders and load a workflow. Written MiniMax authorization is needed above 1 million USD annualized revenue or for hosted services.
Source: Hugging Face trending한국어
PhoneBox
Turns an old Android phone into a desk panel for a Windows, Linux or macOS computer, over USB or Wi-Fi/Tailscale with no cloud. It shows Claude and Codex quotas, and Claude Code permission prompts appear with Allow/Deny.
Usable todayTry the live browser demo, or follow the README Quick start to install the Android app and the PC agent.
Source: GitHub new repos한국어
igorls/comfyui-h3-editor
A timeline video editor for MiniMax H3 inside ComfyUI. Write shots, render on your own GPU and cut picture and sound together, with cast members referenced by @Name to keep characters and places consistent.
Partly usableInstall it in ComfyUI on a GPU that can run MiniMax H3 models. The text gives no install command.
Source: GitHub new repos한국어
anteloc/ldraw-mecha
An AI agent reverse-engineers the mechanics of an LDraw LEGO model from its .mpd file, then writes Python for Blender that builds rigs and animates the mechanisms.
CodingFor developers, creators
Usable todayClone github.com/anteloc/ldraw-mecha, run ./setup-macos.sh, add the printed env vars to your shell profile. Needs Python 3.12+ and git-lfs; Linux and Windows untested.
Source: Show HN (AI)한국어
weixijia/mas-office
Runs a designed team of Claude Code or Codex sessions on your computer, each role a real session you signed in. Approvals, questions and stuck roles land in one inbox, and answers come back with sources tagged and claims checked.
AutomationFor developers, office workers
Usable todayIn a work folder run npm install mas-office mas-office-runtime (Node 18.14 or later), then mas-office up.
Source: GitHub new repos한국어
voyager-so/vork
A 3D and VFX engine built for AI agents: the agent writes the scene as code and vork renders it on the GPU. It is written in Rust and Slang and runs on Vulkan and Metal on Linux, Windows and macOS.
Video and imageFor creators, developers
Partly usableInstall following the README's Install section (no command shown here), then paste the README prompt to your agent. vork cloud removes the need for a local GPU.
Source: GitHub new repos한국어
Comfy-Org/RelayVSR
Repackaged RelayVSR video upscaling model files for ComfyUI, in 2x and 4x versions with a ready Video Upscale: RelayVSR 2x workflow. The original model is kopper/RelayVSR.
Usable todayDownload the files into ComfyUI/models/diffusion_models and ComfyUI/models/vae, then load the Video Upscale: RelayVSR 2x workflow.
Source: Hugging Face trending한국어
ConwayResearch/Underdog-Woof
A small English text-to-speech model with two voices, Juno and Mira, in a single 40.6 MB file. It runs on a laptop CPU with only numpy, with no GPU, server or API key.
Saving moneyFor creators, developers
Usable todayTry the one-voice version in the browser at underdog.ai/woof, or download the two-voice model and run its numpy runtime.
Source: Hugging Face trending한국어
VideoRebirthAI/hyperflow-h3-flash
An open-weights 14B text and keyframe-to-video model with synced stereo audio, a compressed few-step version of MiniMax-H3. A 5.2 s clip at 6 steps takes 92 s on an RTX 5090. The Q4_K_M GGUF is 9.1 GB.
Video and imageFor creators, developers
Partly usableGet weights from the Hugging Face page and use the ComfyUI templates from the Video-Rebirth/hyperflow-h3-flash repo. Check the licence territory limits first; the US, UK, EU and South Korea need prior authorization.
Source: Hugging Face trending한국어
MiniMaxAI/MiniMax-H3-Image
An open-source image generation and editing model from the MiniMax H3 family. It handles text-to-image and reference-image editing in one checkpoint, at about 16 megapixels with no upscaler.
Video and imageFor creators, developers
Usable todayInstall diffusers, transformers and accelerate, then load it with DiffusionPipeline.from_pretrained. The API and the Hailuo web app are alternatives.
Source: Hugging Face trending한국어
omergocmen/ai-video-edit-workflow
A local workflow where Claude Code, Codex CLI, Gemini CLI or Cursor edits a video from a browser form. It can cut, add a logo, zoom and add music, with output in the out/ folder.
Video and imageFor creators, developers
Partly usableOn Windows, extract the zip, double-click baslat.bat, fill the form and press Edit. Mac steps are in the README, which is in Turkish.
Source: GitHub new repos한국어
bayshier/pixelfx-zh
A library of 244 code-made pixel art effects, including fire, explosions, water, ice and lightning. Effects can be recolored, mirrored, or exported as sprite sheets, and a Claude Code skill makes new ones.
Video and imageFor creators, developers
Partly usableOpen the live gallery linked in the README to play and export effects. The README text gives no install commands for the skill.
Source: GitHub new repos한국어