GitHub Trending Today : bolt-slides, loopkit, muscriptor, football-stadium, appless, penecho

[00:00:00]Welcome back to GitHub awesome. This is GitHub trending today number 41. 35 trending open- source projects on GitHub right now. Let's get into it.

[00:00:11]Agent as a router lets a coding agent choose the backend model per task instead of sending everything to the most expensive option. A router uses task by model results, tries a cheaper model first, verifies the answer, and can escalate when needed. The repo includes benchmark data, offline reproductions, and adapters for existing routing tools. The interesting idea is that routing becomes an agentic decision, not a static model name rule.

[00:00:40]Bolt Slides treats a presentation as a React app, so each slide can contain live data, working prototypes, or a full 3D scene instead of another screenshot.

[00:00:49]The bundled agent skill helps tools like codeex or clawed code compose the deck from a prompt. You still get presenter notes, click builds, annotations, thumbnails, and a synced presenter tab.

[00:01:02]It's a better fit for demos when the product itself should stay interactive.

[00:01:07]Loopkit gives coding agents a lightweight plan, act, verify, loop instead of wrapping them in another development framework. It drops files into your repo, loads its 49 focus skills only when their triggers match, and uses a separate verifier to challenge done claims. The skills are plain markdown, so codecs, clawed code, cursor, and other compatible agents can reuse them. Muscriptor turns a mixed audio track into instrument aware MIDI rather than assuming you only recorded a piano. Developed by QI and ML, it streams note events as it processes audio, can restrict decoding to selected instruments, and includes a browser interface with a live piano roll. You can run small, medium, or large model variants locally.

[00:01:55]LLM space is a desktop workbench for building agents when chat logs alone aren't enough to explain what went wrong. You can version prompts and tools, watch every model call and tool run, then replay failed sessions step by step. It also compares performance across runs and keeps threads as local files alongside your API keys. For agent builders, that keeps the whole debugging loop visible without scattering it across separate tools.

[00:02:24]Agent apprenticeship turns completed agent runs into reusable experience so the next task can learn from actual execution rather than another static prompt. It runs local workflow loops with an apprentice agent. Then a mentor model or human evaluates the result and compiles the trace into training material. It supports tools like codeex and claude code with public or private contribution modes.

[00:02:50]Stat view turns seat selection into an explorable 3D stadium so you can see the pitch from a specific chair before committing. The entire venue is procedurally generated with thousands of selectable seats rendered through instancing animated camera flights and firsterson views. It's a concept with simulated tickets and pricing, but it makes the usual flat seating chart feel painfully outdated.

[00:03:16]Kulum adds sound feedback to web interfaces without shipping audio files or asking you to design a tiny sonic brand from scratch. Its 14 interaction sounds are synthesized live through the web audio API with zero runtime dependencies. Add data attributes for hover, press, release, or toggles. then call bind once. It even throttles hover playback and lets your app disable sound because a musical menu gets annoying very quickly.

[00:03:45]Applas imagines a phone with no fixed apps. You ask for something and an AI generates the native interface on demand using Certino components on iOS or Material 3 on Android. Screens stream in as they're produced and likely next views can be prefetched for faster navigation. The actions are simulated though. Orderplaced books, nothing.

[00:04:07]Snap a photo of a pile of clothes on your bed and wardrobe pulls out every garment as its own clean product cutout, no background, then optionally drapes each piece onto a modeled photo of you.

[00:04:19]The app keeps its image library and JSON database local, while bundled codec skills can import entire folders or assemble new outfit ideas. It feels like a personal inventory system built around clothes you actually own.

[00:04:34]Peno gives AI a shared canvas instead of trapping every idea in a chat box. Write an equation, sketch a diagram, or circle part of your work, and it reads both the marks and their spatial relationships before answering beside them. AI results stay as movable drafts until you accept them, while sparse tiles keep the huge canvas lightweight. It's especially useful when translating visual thinking into text would destroy half the context.

[00:05:02]Clat code is a terminal coding agent that routes each request to a different model tier instead of making you choose one model for the whole session. It also indexes your project as a code graph, tracks spending in real time, and can cap session costs. The maintainers report perfect results across their latest 33 task benchmark at a lower cost than several rivals.

[00:05:26]Bloom turns a folder of markdown or MDX into a polished documentation site without making you maintain an Astro project. Run one command and it generates navigation, local search, theming, open graph images, and typed configuration behind the scenes. It also exposes raw markdown, LLM, TX files, and an MCP server for coding agents.

[00:05:48]Sallyport lets an AI agent use credentials without ever placing those credentials inside the agents environment. The Mac app keeps API and SSH keys in an encrypted local vault, receives requested operations over MCP, and executes them after the required approval. Each action is recorded in a signed journal with Touch ID available for confirmation.

[00:06:11]CareerHelper combines a visual resume editor with AI tools you can self-host instead of locking everything inside another subscription service. You get drag and drop sections, 50 templates, PDF or DOCX export, and job description matching that flags keywords and suggests improvements. It can also run mock interviews with adaptive follow-ups and progress reports.

[00:06:35]Image to 3D usually hands you a mesh blob you can't edit or animate. IM23JS rebuilds the object as TypeScript instead. Primitives and shaders with a real hierarchy of pivots, sockets, and colliders underneath. 12 lib Python scripts run validation and gating so the model only spends attention on one sidebyside comparison sheet per pass.

[00:06:58]Holo turns the desk around your MacBook into four tap zones using nothing but the built-in microphone. You tap 10 times in each corner to calibrate and it learns the acoustic signature of that specific desk in that specific spot.

[00:07:12]Move the laptop and the profile is void.

[00:07:15]taps run a shortcut or a shell command and everything stays on the machine with the raw audio thrown away after each 90mise window.

[00:07:26]What happened turns a messy X timeline into a structured briefing about one named topic. It adapts the search window to the story's pace, clusters posts into opinion camps, lays out active debates, and keeps direct post links as receipts.

[00:07:41]The skill also labels sample bias and separates rough sentiment estimates from polling. It only works inside Grock build because it depends on Grock's native exarch and thread tools.

[00:07:53]TU is store is an app store for terminal software, but it doesn't replace the package managers already on your machine. It detects your OS and available tool chains, then ranks compatible install methods across Homebrew, Cargo, UV, Docker, and more.

[00:08:08]Before anything runs, you see the exact command and whether it's verified, scraped, or guessed. That transparency matters when one key install would otherwise mean blindly trusting a shell command.

[00:08:21]Paul turns agent progress into a machine readable registry instead of another stale plan document. Every feature starts as a falsifiable claim with acceptance gates frozen before implementation, then lands as verified or an honest negative. Its standard library Python tools add dependency tracking, scoped edits, searchable memory, and an autonomous experiment loop.

[00:08:45]Cakura grows a different cherry blossom tree inside your terminal every time it runs. Built in C99 with encurses, it animates pedals drifting, settling, and fading while keeping the default refresh rate at 20 FPS. You can adjust wind, pedal density, switch among 15 color palettes, or press one key to regrow the tree. It's basically terminal eye candy with unusually thoughtful controls.

[00:09:12]Toolcraft is a React and TypeScript starter kit for building focused design apps with whichever coding agent you already use. Instead of starting every canvas tool from an empty project, you get a working canvas, configurable controls, and image or video export workflows ready to reshape.

[00:09:29]Generated projects include editable runtime and UI source, so the result can become a real product instead of a fragile AI demo.

[00:09:38]Harness Engineering is Ryan Leopo's field guide for improving coding agents without changing the model. The idea is to shape the repository around the agent so local decisions, quality requirements, and proof checks stay available instead of vanishing between sessions. The collection combines essays, practical playbooks, evaluations, and an agents file that routes tasks to relevant material.

[00:10:04]Vmodal for Flutter gives Android and iOS apps semantic video search through a typed Dart API. Instead of making users scrub through footage, your app can find moments by imagery, spoken words, or text shown on screen. The SDK also handles streamed uploads with live progress and real cancellation while leaving authentication and interface design to your app. Pins scope reviews electrical schematics with the component data sheets open, catching problems a normal ERC can't see. feed it a net list bomb and the relevant PDFs and it builds a graph of the design before checking each IC against pin functions and operating limits. Every finding links back to a data sheet page while deterministic checks handle bomb totals and capacitor derating.

[00:10:53]GC minimal zen poster is a codec skill that turns a loose theme or reference image into a restrained editorial poster, then generates the raster result. Its visual rules are unusually specific. A vertical aged paper canvas, lots of negative space, one small subject, and a single bright color anchor. It also layers in xerox, risograph, or halftone texture, so outputs feel printed and imperfect instead of like glossy ad mock-ups.

[00:11:21]Market Pilot is a market research prototype that keeps AI findings tied to inspectable evidence. Instead of handing you a polished summary with no trail, it collects public signals, masks personal identifiers, and separates direct direct evidence from inference. Research questions run through evidenceaware prompts. Then the report surfaces consensus, disagreement, limitations, and supporting signals.

[00:11:46]Codor gives multiple coding agents one shared channel while each keeps its native session and context. That means you can bring clawed code, codecs or open code into the same local workspace, assign roles and follow their work in chronological order. Sessions survive restarts, messages and run history stay on your machine and private remote access can use tail scale. Open chat cut combines conversational video editing with a real multi-track timeline so an agent's changes remain visible, editable, and undoable. Codeex, Claude Code, or the built-in agent can inspect a project through MCP, then write cuts, captions, effects, and audio directly onto the timeline. Projects and media stay local by default, and Remotion plus FFmpeg handle exports.

[00:12:35]FastctX is a Rust based MCP runtime that gives coding agents structured repository tools instead of forcing them to assemble shell commands for every read or search. It handles file discovery, pageionated output, encoding, and mechanical replacements with stable schemas. Longunning bash jobs can also survive session restarts, but shell access is optin and the server inherits your local permissions.

[00:13:01]Native is a native Mac OS workspace for running MLX models locally on Apple Silicon. Instead of juggling a model server, chat client, and monitoring tools, you get them in one Swift UI app.

[00:13:13]It discovers compatible models from your hugging face cache, tracks token speed and request metrics, and exposes open AI and anthropic compatible local APIs for coding tools.

[00:13:24]Laura Speedrun turns fine-tuning optimization into a controlled race.

[00:13:29]Same task, same model, same L40S GPU, lowest verified wall clock time wins.

[00:13:35]Submissions can change Laura placement, packing, schedules, kernels or data ordering, but they must clear fixed accuracy targets. Every record is rerun three times with fresh seeds in a network blocked sandbox and reports include rejected attempts.

[00:13:50]Nanoex is a minimal Rust library for embedding a codec style coding agent directly inside your own process. It keeps typed conversation history and a persistent responses websocket while exposing turns you can cue, steer, cancel, or fork from exact checkpoints.

[00:14:07]That's useful when an agent is one component of a larger application, not the whole product. Burrow puts a bun native development machine inside a browser tab without sending your workspace to remote sandbox. It combines a shared virtual file system, shell, shell git, package manager, and live server preview. All running client side.

[00:14:28]A small web GPU model can read files, edit code, and run commands while every action stays visible and interruptible.

[00:14:36]Machine genome is an experimental identity and provenence protocol for tracing how models, agents, data sets, and related artifacts connect. Each signed genesis record gets a content addressed gene while parent links build a verifiable lineage graph. Amendments and third-party attestations add evidence without rewriting the original history. And registries expose inclusion proofs and signed checkpoints.