← All transcripts

GitHub Trending Today #47: 4DAnyone, microduck, boop, PRAXIST, acryl, pgterm, sepia, OpenInstinct Transcript, AI Summary & Key Points

Github Awesome · 2 hours ago · Science & Technology · 14:50 · EN-US

AI Summary

35 trending open-source GitHub projects are presented, spanning 3D reconstruction, robotics, developer agents, self-hosted infrastructure, databases, design workflows, video editing, gaming, observability, simulation, and AI model integrations. The projects emphasize local or self-hosted operation, durable agent context, reproducible workflows, privacy boundaries, and specialized developer tooling.

Key Points

  • 4DAnyone converts a single-person phone video into synchronized virtual viewpoints and uses Gaussian splatting to reconstruct the moving person in 3D.
  • Microduck controls a 25cm biped robot with reinforcement-learning policies running a 50 Hz loop across 15 servos on a single Rockchip chip.
  • Boop is a self-hosted event inbox for sending application events to an iPhone through Apple's push service, with SQLite storage and no hosted relay or account system.
  • Anti-mage evaluates browser-environment coherence using one-way contradictions and does not penalize missing features.
  • EpicInfographics creates scene-based graphics from data, checks color separation and layout, and renders deterministic MP4 or GIF output.
  • PRAXIST runs multiple agents on competing paths toward a metric and carries successful results and failed approaches between generations.
  • Acryl provides durable project events, tasks, artifacts, checkpoints, and handoffs while giving each coding agent its own context projection.
  • pgterm provides a read-only multi-database PostgreSQL interface with diagnostics for locks, query regressions, indexes, vacuum issues, and replication lag.

🔒 27 more in the full analysis

From this video

64 products

4DAnyone ACRYL Anti-Mage ask Boop Breakscale Bun Claude Code Confetti Control Center create-photo-flipbook-ui Drizzle ORM effect-mq Epic Infographics Epic Infographics ESLint Experiential FFmpeg fire-your-seo-agency Fire Your SEO Agency Gemini Gmail Goldie Google Calendar headcount Highball iMessage Jot Lemmalog LM Studio MemoraX Code Microduck Model Context Protocol (MCP) modelprint Ollama OpenAI Codex OpenInstinct Open Pstack Open Steps OpenTIE pgbot pgterm PostgreSQL Praxist Prettier Project SuperDex React Redis Refactoring UI (Claude Code skill) Rust sepia SnagTime SQLite Steam tare Tauri TSRX Vercel AI Gateway Vercel Blob Vite WebRTC whip Wine WolfCut

AI in practice

Used for

With
4DAnyone
How
It synthesizes synchronized views around the subject from a single viewpoint and feeds them into Gaussian splatting.
Outcome
A 3D reconstruction that can be viewed from camera angles that were not filmed.
With
microduck
How
Reinforcement-learning policies run on a single Rockchip through Rust daemons in a 50 Hz loop across 15 servos.
Outcome
The 25cm robot walks, stands back up after being knocked over, and rolls when wheels are attached.
Replaces
Handwritten robot control gates.
With
EpicInfographics
How
The agent selects a visual metaphor, computes chart geometry, checks color palettes for colorblind separation, renders HTML through headless Chromium, and performs pre-flight checks for clipped or colliding text before exporting MP4 or GIF files.
Outcome
Scene-based graphics and deterministic MP4 or GIF output.
Replaces
Building another grid of rounded cards and manually checking animated graphics.

🔒 26 more in the full analysis

Agents

  • EpicInfographics agent skill — Turn numerical data into scene-based graphics and animated exports. 2 held 00:51
  • PRAXIST — Find a better result for runnable code against a supplied metric. 2 held 02:19
  • OpenInstinct — Complete web tasks such as bookings, shopping, Gmail actions, and calendar actions for a user. 2 held 04:25
  • open-pstack — Implement a coding task using an appropriate workflow and validate it as a user would. 2 held 08:40
  • Fire your SEO agency — Audit a website and fix search, crawl, citation, and machine-readable-content issues. 2 held 11:07
  • Whip — Perform coding work through terminal commands, file edits, and delegated background tasks. 2 held 11:32

Business ideas

Turn an ordinary phone video of one person into synthesized camera angles and a 3D reconstruction using Gaussian splatting.

For
People who need 3D or free-viewpoint footage from video captured from only one viewpoint.
Solves
Creates camera views that were not filmed and enables a camera to move around a moment originally recorded from the front.
  • 4DAnyone: generates synchronized views around a person and reconstructs them in 3D.

🔒 The rest of this idea is in the full plan.

Get credits →

Provide the control software for a 25-centimeter biped robot, using reinforcement-learning policies to operate its motors, camera, and gamepad, with alternate brains for different locomotion hardware.

For
Robot builders and researchers working with small legged or wheeled robots.
Solves
Replaces handwritten movement gates with learned policies and allows the same robot platform to stand up, walk, or roll.
  • microduck: operates a 25-centimeter biped robot at a 50 Hz loop across 15 servos and can load a different brain when wheels are added.

🔒 The rest of this idea is in the full plan.

Get credits →

Receive application events on an iPhone through a self-hosted endpoint, with redaction, SQLite storage, device pairing, silencing rules, and retention controls.

For
Developers who need application notifications without a hosted relay or account system.
Solves
Sends app events directly to an iPhone while keeping the event store and notification infrastructure self-hosted.
  • Boop: receives application events through one endpoint and sends them to an iPhone without a hosted relay.

🔒 The rest of this idea is in the full plan.

Get credits →

Links mentioned

🔒 Full analysis locked

Unlock more videos and the full analysis

Buy credits to process more videos. Each run includes the full analysis, not just the summary — and you get access to the locked analysis across the library.

Get credits →

Transcript

Searchable transcript of GitHub Trending Today #47: 4DAnyone, microduck, boop, PRAXIST, acryl, pgterm, sepia, OpenInstinct — Github Awesome (14:50). Search for a phrase, then click its timestamp to jump straight to that moment in the video.

Captions sourced from the original video on YouTube, published by Github Awesome. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.

00:00 Welcome back to GitHub awesome. This is GitHub trending today 47 35 trending open- source projects on GitHub right now. Let's get into it. Ford Anyone takes an ordinary phone video of one person and invents the camera angles you never shot, turning a single viewpoint into a ring of synchronized views around them. Those generated views then feed a Gaussian splatting step that rebuilds the moving person in 3D.

00:27 So you can fly a camera around a moment that was only ever filmed from the front. Microduck is the software brain of a 25cm biped robot that walks using reinforcement learning policies rather than handwritten gates. On a single rock chip chip, a handful of rust dammons run a 50 Hz loop across 15 servos. One owns the motors, one streams the camera over WebRTC, one reads the game pad.

00:52 Knock it over and it stands back up. Put wheels on and it loads a different brain to roll. Boop is a tiny self-hosted inbox that sends events from your app straight to your iPhone. Your code posts to one endpoint. A Go server redacts sensitive fields, stores the event in SQLite, and talks directly to Apple's push service without a hosted relay or account system.

01:16 The web UI handles projects, device pairing, silencing rules, and retention. You must build and sign the iOS app yourself. Anti-age scores how coherent a browser environment is and its whole character is refusing to overclaim. Instead of a block list of tells, it reads bodies of evidence that can only disagree in one direction. A feature that shipped after the version a browser claims is a contradiction.

01:42 But a missing feature is never counted against it because honest machines lack things for a dozen reasons. Epic infographics is an agent skill for turning data into scene-based graphics instead of another grid of rounded cards. It picks a visual metaphor, computes chart geometry from the numbers, checks pallets for colorblind separation, then renders the HTML through headless chromium.

02:08 A pre-flight pass catches clipped or colliding text while the animation tool scrubs CSS frame by frame into deterministic MP4 or GIF output. Praxis is for the moment a project runs, hits a number, and nobody knows what to try next. You hand it the runnable code and one metric, and it sends several agents down competing paths at once. Each generation feeding what worked into the next instead of starting cold.

02:36 Failed approaches come back as evidence you can act on, not silence. It's source available under a fair source license, and it leans on codecs or clawed code to drive it. Acril is building a persistent development workspace where coding agents can change without taking the project's memory with them. Canonical events, tasks, artifacts, checkpoints, and handoffs stay durable while each agent receives its own context projection.

03:03 A Cortis plug-in layer treats agents, terminals, memory providers, and UI surfaces as replaceable capabilities that can be rolled back. PG term is HTTOP for every Postgress database you care about. Each database gets a tab while background checks surface blocked locks, query regressions, unused indexes, vacuum trouble, and replication lag without interrupting the view you're using.

03:26 It delegates diagnostics to PGBot, but keeps a strict readonly boundary. No SQL console, shell, or fix button. Memorax code gives codec and claude code a shared memory layer for engineering work. It separates durable lessons, repository structure, personal preferences, and reusable procedures. Then retrieves only the pieces relevant to a new task. Personal preferences and procedures stay inside the repository, while cloudback search and writeback use memorax.

04:00 CPIA edits AI pros above the word choice layer. For fiction, it starts with narrative architecture. Overly tidy causality, explained themes, linear time, thin character networks, and endings that resolve too neatly. Only then does it fix paragraph flow and surface style. Professional writing gets venue specific rules for release notes, PR replies, postmortems, tickets, and technical articles.

04:25 Open Instinct is a self-hosted personal assistant you text through iMessage with a cloud browser acting on your behalf. It can sign into sites, handle bookings or shopping, and connect to Gmail and calendar while keeping passwords and card details encrypted outside the model's view. Browser screenshots and per user memory live in private Versel blob storage and inference runs through Verscell AI gateway.

04:51 Refactoring UI skill turns Adam Wan and Steve Scogger's design advice into a claude code rulebook for building and fixing interfaces. It gives the agent fixed scales for spacing, type, color, shadows, and radius, plus a review process for hierarchy and contrast. A starter CSS token set covers dark mode and accessibility checks. The repository uses original notes, not the book's text or images.

05:17 Breakcale lets you build a distributed system on a canvas, then drag a traffic slider until it falls over and watch exactly why every number comes from a real discrete event simulation rather than an animation. Cues fill tail latency diverges from the average because service time actually varies and a request that times out keeps burning the database it was waiting on, which is how a retry storm collapses goodput to zero in front of you.

05:46 Ask gives you a normal shell command for asking an AI question without handing it control of your project. It runs your existing codeex clawed code in readonly mode. Prints the answer to stood out then returns you to the prompt. Every turn is saved and ask- C resumes the latest conversation for the current folder. It supports Mac OS and Linux, but the chosen agent must already be installed and logged in.

06:08 Experiential puts hosted bring your own key and local models behind one OpenAI compatible gateway. You can expose stable model aliases, restrict which users or agents may call them, and set spending limits without rewriting every client. It can also learn a custom router from open telemetry traces, then optimize model selection around quality, speed, and cost.

06:32 You can run the data plane locally or use the hosted service. This skill turns a folder of photos into a page turning 3D flipbook website. It edits like a photo editor before it builds anything. It lays your shots out as a contact sheet and actually looks at it so repetition and weak frames are visible. Then keeps only the strongest 11 of 16 in the demo and plans an opener, pauses, and an ending rather than dumping everything in order.

07:03 Goldie gives coding agents a repeatable pipeline for app store screenshots and preview videos. Argent replays flows against an iOS simulator. Then Goldie adds device frames and copy for joining the captures into an H.264 preview. A browser studio lets you adjust templates and saves the design for identical CLI renders while validation checks Apple's upload rules.

07:28 Terry turns Claude Code's local session logs into plain English answers about where your usage went. Ask why you hit a limit, which project consumed the quota, or whether automation was running in the background, and it traces the cause instead of dumping raw token counts. It dduplicates repeated log entries, attributes the cost of repeatedly resent context, and reconstructs rolling usage windows.

07:54 Open TAI runs the original Star Wars Tie Fighter data natively on Windows, Mac OS, and Linux. If you own both supported editions, it can pair the 1995 menus, cutscenes, and adaptive IMU soundtrack with the 1998 flight simulation and 3D assets. Classic and modern graphics modes sit side by side with modern lighting, HDR, and high refresh simulation. Wolf cut is a free video editor built to do what people use Cap Cut for without the watermarks or the pro payw wall.

08:23 A Rust engine does the cutting and encoding while a react and Tory interface handles the timeline so it runs natively rather than as a browser pretending to be an app and it installs as a few megabytes with ffmpeg bundled and no account. Open Pstack brings Lauren Tan's stack, the workflow kit she built shipping code at Cursor to Claude code and codeex.

08:48 It's a plugin rather than a model. You hand Patito mode a task in plain words and it picks a workflow, learns how the code works before touching it, prefers the smallest change that works, and runs the feature like a user would rather than stopping at green tests. Confetti replaces a pile of VS code extensions with one focused tool for configuration files.

09:09 It detects 16 plus formats from both path and content, then adds syntax highlighting and formatting for Nex, Donv, Toml, SSH config, and hosts files. Ambiguous files stay untouched unless detection is confident. It runs locally with no account, network access, or background index. Control center is a local first dashboard for industry news. exact brand mentions, newsletter stories, and audience tracking.

09:37 You configure sources yourself, while collectors save snapshots locally instead of rebuilding each view. AI curation is optional using cloud providers or local models through LM Studio and Alama. Core monitoring still works without one. Project Superex is a simulation stack for dexteras robot manipulation built around a contact first physics engine.

09:59 Its robotics SDK composes controllers, sensors, actuators, and robot definitions, while Studio turns CAD assets into bots, scenes, and task prefabs. Lab exposes a gymnasium style interface for reinforcement learning and control research. Jot turns holding FN into systemwide dictation on Mac OS. Speak, release, and Gemini transcribes, removes filler, follows mid-sentence corrections, then inserts polished text at the current cursor.

10:30 Recordings are written to disk immediately, so crashes and offline sessions can be recovered or retried from history. Your audio goes directly from the Mac to Google's API, but Jot still requires a Gemini key. TSRX is a TypeScript language extension for writing UI components with setup code, JSXshaped markup, template control flow, and scoped styles in one file.

10:50 It parses that source into a framework neutral a then target compilers generate code for react view. The tool chain includes vite and bun integrations plus language server prettier eslint and editor support. Fire your SEO agency is a clawed code skill that audits a site across traditional search answer engines generative search model knowledge and neighbor which gets its own workflow.

11:16 It checks the page without JavaScript, reports crawl and citation problems, then can fix server rendering, sitemaps, metadata, structured data, and machine readable content before scheduling a fresh measurement. Whip is a go- based coding agent harness that ships as one binary with no extra runtime. Its terminal interface wraps a small tool loop for shell commands, file edits, and background sub aents with parallel calls and streaming throughout.

11:47 Model routing works across OpenAI compatible providers while live catalog discovery surfaces newly available models without per model configuration. Highball runs Windows games on Apple silicon without locking you to one wine build. The Mac OS app and CLI assemble wine with several renderer options, create separate bottles, install Steam, and let each game use the renderer that works best.

12:12 Its open compatibility database records verified results and flags kernel anti-che titles before you download them. Effect MQ brings background jobs into effect without forcing your app around a particular Q backend. You define jobs with typed schemas, retries, IDM potency and metadata. Then choose Postgress, Reddus or memory per workload. Workers persist every attempt, recover stalled jobs, renew locks and keep scheduled runs atomic across processes.

12:40 Its Postgress tables live in your Drizzle schema, so Drizzle Kit owns migrations. Model print helps you check what may actually sit behind an OpenAI compatible API alias. It runs browser based probes against endpoints, comparing tokenizer counts, validation errors, response headers, and router metadata side by side. Your API key stays in the tab, and contributors can add deterministic probes as small JavaScript files.

13:10 OpenSteps gives coding agents a reporting layer for people who don't speak and commit hashes. Its skills turn finished work into a plain verdict, simplify questions and instructions, verify another session's claims, and run a premortem before hard to reverse decisions. Hooks request reports after files change and store them outside the repository. Headcount packages Claude Code expertise as an installable company rather than one oversized prompt.

13:38 Its 16 departments contain 143 named skills for engineering and the business around it. You install only the departments a project needs so unrelated instructions stay out of context while namespaced skills avoid collisions. Each department also includes an agent charter for delegation with its own right surface. Lemalog gives LLM agents a memory they can query and audit rather than another pile of embedding matches.

14:06 The Rust engine stores facts with source provenence, applies stratified data log rules incrementally, tracks retractions, and can explain derived answers with proof trees. An MCP server exposes that memory to clawed code and Kimmy CLI, including whatif queries. Snagtime gives you a self-hosted scheduling system with public booking links, time zone handling, custom questions, rescheduling, and calendar conflict checks.

14:33 The local demo runs on SQLite with fake email, calendar, and payments, so you can test the full flow without external accounts. Production adds Google calendar, SMTP, Postgress, and a background worker.