Nimbus scaffolds editable layouts, components, styles, and routes in the repository, then generates agent-readable Markdown twins, llms.txt files, JSON-LD, full-text search, and static output deployable anywhere.
Decimal optical transfer moves files as light by displaying an endless sequence of QR frames in one browser and reconstructing them through another browser's camera.
Ratel reduces agent tool overload by indexing tool schemas and skill metadata, searching them with in-process BM25, and injecting only matching tools or instructions for the current turn.
Backchannel is an open-source userscript and renamed continuation of HNewhere that brings discussions from Hacker News, Bluesky, Reddit, Lemmy, Lobsters, and other selectable sources into a sidebar beside the page or PDF being read. It checks visited pages for matching threads, merges enabled sources into one conversation with source labels, and matches quoted comments to article passages as clickable annotations; enhanced PDF highlighting is supported except in Firefox. Users can vote, reply, and submit through popups on supported source sites, add locally stored notes to pages and PDFs, and use a customizable front page combining content from selected sources. The single-file script runs without a build step, backend, analytics, or telemetry, and is distributed through userscript managers such as Tampermonkey, Violentmonkey, and Userscripts.
camelAI is an AI coding assistant platform built on Cloudflare Workers and Durable Objects. Each chat thread runs a coding agent in its own persistent workspace, retaining chat state and project files while providing workspace-aware tools. The platform connects to APIs, databases, email, Slack, Discord, and other services; supports application builds, notebook analysis, and SQL execution in isolated short-lived sandboxes; and provides previews and publishing through Workers for Platforms. Its agent harness is built on pi's lower-level agent-loop and state-management libraries. The agent uses native file tools and writes JavaScript rather than shell commands; Code Mode executes that JavaScript in fresh V8 isolates with explicit platform and connection methods. Project files are stored in WorkspaceFilesystem Durable Objects, using Durable Object SQLite for small files and R2 for larger ones, while Cloudflare Artifacts provides Git history. Linux sandbox containers are reserved for jobs requiring them, including builds, notebook analysis, and database queries, and credentials remain outside the execution sandbox. The platform can use Anthropic, OpenAI, OpenRouter, Bedrock, and custom model endpoints. Its repository is released under the MIT License and includes local development setup using Node.js, Bun, and a Cloudflare account.
Claude is an AI assistant developed by Anthropic, positioned as 'The AI for Problem Solvers'. It is a general-purpose AI system used for tasks such as generating code prompts, refining requirements, creating advertising strategy and copy, and processing creative content like storyboarding and video prompts.
CodeJury is a terminal-first, knowledge-grounded multi-agent software delivery pipeline for scoping requirements, implementing changes, running tests, and gating pull requests. It divides delivery into six terminal-driven stages and uses isolated development branches, human approval gates, deterministic QA, and ensemble code review. The system indexes a repository into a persistent code graph and uses semantic search to retrieve relevant code during later work. Its review stage sends the same change to independent models from different providers; both must approve it for the change to pass, while a rejection returns the reason to the developer. A configurable panel can use up to six specialist reviewers and a foreperson to reconcile their findings. The project is distributed as a Python package installable with pip and documents support for macOS, Linux, and Windows. Its repository describes a one-command run and operation without an API key.
ctrlb-decompose is a log-processing tool from ctrlb that converts large, repetitive log files into compact, LLM-ready patterns with typed variables, quantiles, anomaly information, severity scores, and example values. Its single-pass pipeline extracts and normalizes timestamps, applies CLP encoding to replace integers, floats, IP addresses, and strings with compact placeholders, clusters log types with the tree-based Drain3 algorithm, and merges wildcard positions with decoded values for semantic typing such as IPv4, UUID, duration, enum, and integer. It accumulates DDSketch quantiles, HyperLogLog cardinality estimates, top-k values, temporal buckets, and reservoir-sampled lines, and provides human-readable and LLM-optimization output modes. The project runs as a CLI, a Rust library, or client-side WebAssembly in the browser; its README states that logs remain on the local machine.
Cynative is an open-source framework and CLI for building security agents with live, read-only access to code, cloud, and runtime infrastructure. It reasons across GitHub, GitLab, AWS, GCP, Azure, and Kubernetes as one system, using frontier models to investigate questions such as exposed resources, privilege escalation paths, infrastructure drift, and leaked credentials. For each investigation, Cynative generates and runs code in an ephemeral sandbox, queries APIs in parallel, cross-checks findings against live evidence, and traces findings to their origins. Its action gate resolves each call to the required IAM actions and applies a read-only policy before attaching credentials; the video also describes an auditable JSONL log of tool calls. Agents are defined as Markdown files containing a description and prompt, and can run interactively, non-interactively, or with piped findings. The project is distributed as a single open-source binary and supports configuring the model provider and model through environment variables.
Decimen Optical Transfer is an open-source browser-based file-transfer tool by Evan Crawley that sends files between devices using only a screen and a camera, without a network connection, pairing, permissions beyond camera access, or a dedicated app. The sender displays an endless stream of animated QR codes; the receiver collects fountain-coded frames in any order and uses Luby transform decoding to reconstruct the payload, so dropped frames affect transfer time rather than correctness. Transfers preserve filenames and media types, accept pasted text, support files up to 64 MB, and verify completed payloads with SHA-256. The web app works offline after its first visit and can be installed on iOS and Android home screens. QR streams can also be exported as APNG or PNG sequences for playback from a file, page, or video. Transmissions are not encrypted: data is readable by any camera pointed at the sending screen. The project is licensed under AGPL-3.0-or-later as of v0.4.0; releases through v0.3.0 remain available under the MIT license.
Drawesome is a drawing toolbar and canvas component for React applications. It provides seven pen types—pencil, pen, brush, fineliner, highlighter, fountain pen, and another configurable pen tool—that vary stroke width according to movement speed or direction, plus an eraser that removes areas of strokes rather than deleting whole strokes. The React component supports configurable tools, colors, sizes, opacity, toolbar placement, appearance, controls, themes, tooltips, keyboard shortcuts, and initial or externally managed strokes. Drawings can be exported as SVG or PNG, downloaded directly, or accessed as stroke data through its ref methods. It is installed with npm, requires React 18 or later as a peer dependency, and has no runtime dependency beyond React.
GeoLook is an open-source, self-hosted platform for end-to-end Generative Engine Optimization (GEO), focused on whether AI engines mention and cite a brand. Its workflow runs from status analysis through diagnosis, strategy, implementation tickets, execution, and verification. The platform samples answers from multiple AI engines and measures mention rate, rank, citation share, competitor visibility, and citation sources. It classifies question-level problems such as absence, low ranking, competitor dominance, and suspected negative mentions; stores browsable sample metadata; and allows human corrections to parser results that recompute metrics and persist across resampling. It can expand question banks using Baidu suggestions and Google autocomplete, with candidates requiring manual review. GeoLook audits sites through an Access → Orientation → Understanding → Quotability dependency chain, covering crawler access, extraction blocks, citation-channel gaps, and brand-fact or messaging inconsistencies. It also performs checks such as robots and WAF/CDN behavior, noindex headers, llms.txt links, hreflang, sitemaps, duplicate content, and passage-level quotability. The action stage produces structured implementation tickets with rationale, ownership, effort, timing, acceptance criteria, risk grades, progress tracking, and regression-based reopening; it also provides content planning, deployable llms.txt and JSON-LD assets, and client-facing reports and execution plans. Subsequent sampling rounds compare results before and after implementation.
An Apache-2.0-licensed open-source Go SDK from Grafana for building AI-powered backends and agents. It provides GenerateText and StreamText APIs for calling language models, returning complete or streamed responses, executing Go functions as tools, generating schema-validated structured output, and running multi-step tool workflows, with approval for consequential tool actions. The SDK supports Anthropic, Amazon Bedrock, OpenAI, OpenAI-compatible APIs, and Grafana's hosted endpoint, along with retries, configurable timeouts, fallbacks, logging, Prometheus metrics, and Agent Observability. It follows the design of Vercel's AI SDK and is wire-compatible with its TypeScript React frontend hooks, including useChat, useCompletion, and useObject, allowing Go backends to stream Server-Sent Events directly to an existing React frontend without a protocol adapter.
hwatu is a Linux visual-verification browser and harness for AI coding agents, implemented as a WebKit daemon. It performs one-call page checks, returns pixel-difference scores and heat maps, exposes animation timing numerically, and runs headless windows that can be handed off to a live human-controlled window. It connects to coding agents through MCP, short CLI commands, or newline-delimited JSON over a Unix socket, and is designed for tiling window managers including Hyprland, sway, niri, and i3. The project distributes a static binary that uses the system's WebKitGTK 6.0 and can also be built from source with Cargo.
Interior Dev is an open-source collection of copy-paste React micro-interactions for the brief period after a click. Each component is distributed as a single file rather than a package and exports a headless `useX` hook containing the behavior without class names alongside an `X` styled example component. The components address layout-preserving transitions, animation restarts, resizing labels, loading jumps, interrupted drags and gesture recovery when focus is lost, keyboard interaction, and `prefers-reduced-motion` behavior; reduced-motion mode preserves the information while skipping the animation. The only dependency is `motion`, and the documentation site can be run with Bun.
kimi-k3-mlx is an MLX port for Apple Silicon of Moonshot’s Kimi K3 native-multimodal mixture-of-experts model. It provides a stock `mlx-lm`-compatible text model definition, a 3D video-capable MoonViT vision tower with a patch-merger projector, and an `mlx-vlm`-shaped wrapper that joins the language and vision components. The model has a 1-million-token context and uses routed and shared experts, Kimi Delta Attention, gated multi-head latent attention, Attention Residuals, latent-dimension experts, and the SiTU-GLU activation; routed experts are distributed in MXFP4 while other weights use bf16. The multimodal processor expands each image placeholder into a feature block rather than performing a same-length one-token-per-image scatter. The repository includes a streaming converter because standard `mlx-lm` conversion would materialize the multi-terabyte bf16 model, along with REAP expert-pruning scripts and expert-overlap analysis using Chinese, English, code, or mixed calibration data. It also contains vision-parity tests against Moonshot’s reference implementation. The repository states that the full model cannot run on a single Mac: even its smallest deployment tier requires hundreds of gigabytes of memory and exceeds the capacity of the largest Apple Silicon machine.
Let's Seal is an open standard, reference implementation, and free service for proving that a file is unaltered, existed by a certain date, and was sealed by a specific certificate. It preserves native signature formats such as PAdES/X.509 signatures for PDFs and uses OpenTimestamps to anchor time to Bitcoin. Each seal is also recorded in an append-only transparency log with a Bitcoin-anchored root, allowing independent auditing. The project provides a hosted web app, the `sealbot` command-line tool, a REST API and SDKs, and a self-hosted engine with its own certificate authority; verification is available online or offline without an account. The repository states that the engine, SDKs, and standard are open source under the Apache-2.0 license.
mirador is an open-source terminal dashboard built with Rust and ratatui for keeping in a terminal tab and glancing at throughout the day. It provides configurable panels for world clocks, a calendar and agenda, weather, an editable task list, notes, a market watchlist, a calculator, RSS news feeds, and live CPU and network charts; panels can be enabled, rearranged, focused, and edited through keyboard controls. Tasks, notes, and the watchlist are stored as hand-editable TOML files rather than in a database or sync service. The dashboard uses keyless public endpoints for weather and market data and configurable RSS feeds for news; the project states that it requires no account, API key, or telemetry and does not contact external services unless requested.
netherite is a from-scratch C and CUDA reimplementation of Minecraft 1.11.2, checked against the original Java game. It provides a batched CUDA reinforcement-learning environment in which multiple worlds step in lockstep on one GPU, with semantic-camera observations; the repository reports a recorded batch reaching 7,200 live worlds. Linux x86_64 supports the full build, CUDA training, and Java-oracle workflow, while macOS provides CPU and Metal game backends without CUDA support. The project does not ship Mojang content and requires a legal Minecraft copy and JDK 8.
Nimbus is an open-source documentation-site scaffolder by Cloudflare that builds Astro sites directly in a repository. It writes layouts, components, styles, routes, content collections, and theme tokens as editable source files, while its plumbing is provided through npm packages. The generated sites include agent-readable Markdown and MDX page twins, llms.txt and llms-full.txt files, JSON-LD, sitemaps, robots.txt, per-page Open Graph images, full-text search, theming, accessible navigation, and optional versioned documentation. Nimbus also provides a registry for adding editable components, utilities, and agent-handoff recipes to an existing project. Sites produce static output for deployment anywhere, with Cloudflare as a first-class target, and include prose, structure, MDX, and configuration validation. The project is built on Astro, Sätteri, Tailwind, and optionally React; its README describes it as work in progress and pre-1.0.
numbat is an open-source endpoint security tool for monitoring AI-agent activity across supported desktop, CLI, IDE, and gateway agents. It combines local hooks and plugins, OTLP/HTTP logs, and on-disk session artifacts, normalizes live and stored activity into a common event model, and evaluates it locally with built-in, sequence-based, or custom YAML CEL rules. The tool can produce NDJSON events, findings, enforcement decisions, indicators, and scan summaries for alerting and forensic reconstruction, including read-only artifact scans, per-session timelines, and portable case bundles with SHA-256 manifests. Optional pre-action blocking is disabled by default and is limited to supported synchronous hooks and explicitly enforced rules. It is distributed as a single cgo-free binary for macOS, Linux, and Windows, with read-only inventory and scanning commands.
OpenAI Codex is an AI coding agent from OpenAI available as a command-line tool (Codex CLI) that helps developers produce software. It can be used alongside Gemini for adversarial audits of software requirements and implementation plans, listed as a supported coding-agent or model option in several projects, and its logs can be joined with task and test evidence. The Codex CLI can also receive and answer requests from the Penako canvas.
OpenSlides is a free, open-source offline desktop application for creating animated code presentations. It turns source code into slides and converts selected lines into presentation steps, using Magic Move transitions to animate changes between code states and stepped highlights with configurable dimming, scaling, and transition timing. It provides light and dark themes, 16 syntax themes, slide search, thumbnails, keyboard navigation, full-screen playback, optional autoplay with per-slide timing, and project operations such as duplication, import, export, and drag-and-drop reordering. Built with Tauri and Svelte 5, it keeps projects on the local computer and provides installers for macOS, Windows, and Linux.
optim-plans is a human-in-the-loop planning plugin for Claude and Codex. It turns repository-change requests into traceable Markdown plans, asks planning questions one at a time, records decisions, applies reviewer or criticizer refinement passes, and requires explicit scope confirmation before writing versioned plan artifacts. Its five public skills cover small, broad, and high-risk planning, diagnosis before planning, and reference analysis before planning. After approval, it hands the accepted plan back to the current agent session for normal implementation. Version 0.3.0 keeps machine state in the Git common directory and public artifacts in repository documentation, but does not include a separate execution engine, manifest, delegated implementation role, verification role, retry loop, or worktree execution gate.
PerceptionBench is a benchmark for evaluating atomic visual perception in multimodal large language models (MLLMs). It isolates perception from reasoning and domain knowledge by defining ten atomic perceptual capabilities, including counting, depth, localization, OCR, and perception-related hallucination, based on an error taxonomy derived from failures across existing benchmarks. The released dataset contains 3,000 verified open-ended questions with short, uniquely determined answers; 1,800 are decomposed from failures in source benchmarks and 1,200 are newly authored using supplementary images. The repository provides the benchmark's code and links to its paper and dataset.
Persona is an open-source, cross-platform desktop character that gives voice conversations an animated visual identity. It listens locally to a selected application's playback process rather than the microphone, using PipeWire playback-stream capture on Linux, WASAPI process-loopback capture on Windows, and a Core Audio process tap on macOS. Imported VRM character models switch between idle and speaking animations, with support for custom VRMA actions, model libraries, click-through desktop-pet behavior, and a local MCP server. Persona does not save audio, produce speech, transcribe content, or send audio over the network. It is distributed as an AppImage or DEB on Linux, an NSIS installer on Windows, and DMG or ZIP packages on macOS.
Qwen Scribe is a private, local transcription and system-wide dictation application for Apple Silicon Macs. It runs Qwen3-ASR through mlx-qwen3-asr on the Mac's Metal GPU, keeping audio and transcripts on the computer without an account, API key, or cloud service. Users can drag audio or video into a local web interface, transcribe with automatic or forced language selection, vocabulary hints, word timestamps, and SRT export, and search or export local transcript history. A menu-bar helper provides push-to-talk dictation in text fields, displaying listening and transcription states in a non-focus-stealing HUD before pasting the result at the cursor. The project requires macOS 14 or newer, Apple Silicon, Python 3.12 or newer, and ffmpeg for non-WAV audio and video. It offers Qwen3-ASR 1.7B and 0.6B model options; model weights are downloaded separately and require roughly 3.4 GB or 1.2 GB of unified memory respectively, in addition to application overhead. It is distributed as a downloadable beta Mac app or can be built from source; the repository describes the project as version 0.2.1-beta.1.
Ratel is an in-process context-engineering platform for AI agents. It catalogs tools, skills, and persistent facts, then searches the catalog on each turn and progressively discloses only the capabilities relevant to the request instead of placing every tool schema and instruction in the context window. Its retrieval supports BM25, semantic, and hybrid ranking without requiring a vector database. The project includes a Rust core, TypeScript and Python SDKs, an MCP server, and a CLI. Tools can be registered directly or ingested from an MCP server; skills can associate instructions with tools, while registered facts are re-injected when they are no longer fresh in the conversation. The repository describes the system as reducing token usage and mitigating tool overload.
Sakura Crossing is an explorable Japanese suburban railway-crossing neighbourhood built as a real-time Three.js 3D scene and rendered to resemble a hand-painted, cel-shaded 2D anime background. The spherical town includes a railway, shopping street, shrine, school, library, residential areas, festival ground, hills and a drainage channel; users can walk, ride an e-bike, call a train, interact with vending machines and other objects, and orbit out to view the planet. Signs, lanterns and other facade details are drawn at runtime with Canvas2D rather than loaded as image assets, while the outlines use a screen-space second difference of depth instead of an edge filter. The repository runs with Node 18 or newer, Three.js and Vite, and provides development and production-serving commands.
Snowflow is a real-time procedural snow-rendering demonstration built with WebGPU, Babylon.js, and hand-written WGSL. It generates terrain, snow deformation, procedural characters, inverse-kinematics foot placement, cloth, surf wakes, water spells, atmosphere, and post-processing at load time without textures, meshes, HDRIs, or animation assets. Its terrain uses a nested-ring geometry clipmap whose vertex shader performs world placement, level-of-detail morphing, and displacement. Layered gradient noise produces the underlying heightfield, while a persistent GPU terrain-state buffer handles additive deformation through scrolling, relaxation, and splatting in a ping-ponged full-screen pass. The demo also provides GPU snow shading, cascaded PCSS shadows, procedural character and cloth systems, and in-browser settings, performance, and debug views. It requires a WebGPU-capable desktop browser and a discrete or recent integrated GPU; it intentionally has no WebGL fallback.
SpinWin is a macOS menubar utility that visually rotates or continuously spins another app’s window without private APIs or disabling System Integrity Protection. It hides the selected window off-screen through the Accessibility API, captures its contents with ScreenCaptureKit, and displays the frames in a transparent, borderless overlay whose CALayer transform applies preset 90-degree rotations, arbitrary angles, or continuous spinning. The overlay is sized to the rotated content’s bounding box to avoid clipping and can be dragged or freely rotated. It requires Screen Recording and Accessibility permissions; the overlay is a live, non-interactive picture, and the project notes limitations involving Mission Control, off-screen rendering throttling, and windows that resize themselves.
Talivia is an open-source, self-hosted revenue-first analytics platform developed by Talivia. Its self-hosted edition combines web analytics, optional Session Replay, website collaboration, shared analytics, import/export, and revenue attribution, connecting website visits with customer revenue from Stripe, LemonSqueezy, Polar, Dodo, Yolfi, or the Manual Payment API. It retains subscription lifecycle events, refunds, disputes, and first- and last-touch attribution. The application requires Node.js, pnpm, and PostgreSQL, or can be run with Docker Compose. Setup involves creating a website, adding Talivia's tracking snippet, optionally enabling Session Replay, and connecting payment data under the website settings. The repository provides the self-hosted edition as a subset of the complete Talivia product; managed Talivia Cloud adds additional integrations.
TanStack Charts is a TypeScript visualization grammar for responsive, accessible, server-rendered application charts, built on granular D3 primitives. Chart definitions are framework-independent: marks consume arrays, objects, tuples, or iterables; channels describe visual encodings; and scale factories infer domains from those channels. Definitions compile into a renderer-neutral keyed scene that can be rendered through React, Preact, Vue, Solid, Svelte, Angular, Lit, Alpine, Octane, the vanilla DOM, static SVG, or an optional Canvas renderer. The system supports layered built-in marks, custom marks using the same scene protocol, responsive layout, automatic guide margins, themes, interaction, animation, accessibility, SVG server-side rendering, hydration, and export. Its marks, transforms, layouts, renderers, and chart-owned interactions are available through independent subpaths, with optional D3-backed capabilities remaining tree-shakeable. The repository documents an unreleased alpha line and says the latest published release is 0.16.0; alpha releases may contain breaking changes between minor versions.
TrueDeck is a terminal-first multi-agent coding workbench for running several coding agents on the same repository in separate live terminal panes. It supports agents including Grok, Codex, Cursor, Claude, and Gemini while retaining their native command-line interfaces rather than presenting them in a chat webview. It provides split panes, tabs, focus control, and workspace restoration, with repository-specific and global memory plus MCP and project-context synchronization handled in the background. Installers and portable builds are distributed through the project's GitHub Releases, and the repository is licensed under the MIT license.
TurboVLA is a compact vision-language-action model and official implementation for robotic manipulation, developed by researchers at Huazhong University of Science and Technology and Huawei Technologies. Instead of routing visual input through a large language model, it independently encodes visual observations and language instructions, exchanges information through lightweight bidirectional vision-language interaction, and uses a compact decoder to predict continuous action chunks. The repository includes training and evaluation code for LIBERO and RoboTwin, with model checkpoints released separately. The project reports 32 Hz inference on an RTX 4090 with less than 1 GB of VRAM; on LIBERO, it reports a 97.7% average success rate, 31.2 ms inference latency, and 0.2 billion parameters.
A formally verified 3D constructive solid geometry (CSG) kernel for intersecting triangle meshes. Implemented in Lean 4, it is checked against a concise formal specification that defines the resulting surface and guarantees stated well-formedness properties for the output triangulation. The project includes a browser demo that imports STL files and runs the compiled Lean code locally without sending mesh data to a server. The kernel is formally verified, but the demo's UI and glue code are not; the implementation prioritizes correctness review over performance and may generate meshes that are finer than necessary.
world-model-optimizer converts OpenTelemetry traces from agent workflows into a routing policy. It scores models on held-out tasks, selects a model for each request, and provides world models for testing changes to prompts, tools, and runtime code.
XY is an open-source Python charting library for web applications, notebooks, and static exports. It supports declarative charts and matplotlib-style conventions, with customization through Python, CSS, or Tailwind; outputs include HTML, PNG, SVG, and PDF. For small charts, XY sends individual points to the browser. For large datasets, its Rust core computes the data needed for the current screen resolution and can represent the result as a density surface. Panning, zooming, hovering, and selections rerun the process for the visible range, while selections can return the original data rows. The library is installable with pip or uv and supports interactions such as pan, zoom, hover, crosshairs, callbacks, and linked charts, along with layers, facets, responsive layouts, themes, axes, legends, and tooltips. The repository describes XY as an alpha project and notes that its matplotlib compatibility does not yet cover all charts and functionality.
Searchable transcript of GitHub Trending Weekly #42: nimbus, xy, persona, drawesome, letsseal, camelAI, OpenSlides, numbat — Github Awesome (15:16). 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 weekly number 42. 35 trending open source projects on GitHub right now. Let's go. Nimbus builds Astro documentation sites where the visible system lives in your repository, not behind a theme package you can barely bend. It scaffolds editable layouts, components, styles, and routes, then generates agent readable markdown twins, llms.txt files, JSON-LD, and ordinary reader features like full-text search.
00:30 Static output can deploy anywhere with Cloudflare configured by default. Decimal optical transfer moves a file as light. One browser displays endless QR frames while another reconstructs them through its camera. Fountain coding makes the stream order independent, so missed or blurred frames add delay instead of forcing a full restart. Each frame carries session metadata, and the finished payload is checked by hash.
00:59 XY is a Python charting library with a Rust core that computes only what the screen can show. Under a couple hundred thousand rows, every point goes to the browser. Above that, it draws a density surface, and zooming drills back to exact rows with selections returning the original data. Their benchmarks show flat render time from 10,000 to 100 million points self-run on one machine.
01:25 Persona gives desktop voice conversations a visual character without recording your microphone. It listens only to a supported app's playback process, then switches imported VRM models between idle and speaking animations. You can add custom VRMA actions and expose them through a local MCP server for Codex control. Audio isn't saved, transcribed, or sent online.
01:48 Drawsome gives React apps a finished drawing toolbar without pulling in a canvas framework. Its seven tools model real pen behavior. Some change width with speed, while the fountain pen responds to direction. The eraser removes only the area you rub, instead of deleting an entire stroke. And drawings export as SVG or PNG. It has no runtime dependency beyond React, though it's opinionated toolbar is better for shipping quickly than building a fully custom editor.
02:19 Let's Seal signs any file, so anyone can prove it hasn't changed when it existed and which certificate sealed it. Each format keeps its native signature, so a sealed PDF is PAdES, an image is C2PA, anything else gets a detached CMS file. Time anchors to Bitcoin through OpenTimestamps, and every seal lands in an append-only Merkle log. One caveat, their root isn't in Adobe's trust list, so readers won't show a green check without pinning it.
02:47 Camel AI runs each coding conversation as its own Cloudflare durable object, giving the workspace persistent chat and project state without keeping a full VM alive. It stores files across SQLite and R2, runs builds, notebooks, and SQL jobs in short-lived sandbox containers, and can publish apps through Workers for Platforms. It supports several model providers and keeps credentials outside execution sandboxes.
03:14 TeliVia is self-hosted analytics built around revenue rather than page views. It ties visits to actual customers through Stripe, Lemon Squeezy, Polar, or Dodo with first and last touch attribution, subscriptions, refunds, and disputes. Session replay is included. It's the open-source subset of a commercial product with the search console and social integrations reserved for the paid cloud.
03:39 Open slides turns code into animated presentations while keeping every project on your machine. Instead of pasting screenshots into a slide deck, you create successive code states, and magic move animates the changes between them. Stepped highlights can reveal selected lines with custom dimming, scaling, and timing, while full-screen mode supports keyboard control or auto play.
04:03 World model optimizer turns the OpenTelemetry traces your agent already emits into a routing policy. It scores every registered model on held-out tasks pulled from those traces, fits a policy, then serves one endpoint that picks per request. It also ships world models that simulate your environment, so an optimizer can rewrite prompts, tools, and runtime code, and only a candidate passing the evaluation gates becomes the new champion.
04:32 Find Disk Killer shows which Mac app is hammering your storage, then keeps the investigation in one workspace. You get 5-second CPU, disk, and network rates, timelines up to an hour, and smart or NVMe health data when macOS exposes it. A time-limited trace can inspect requested file reads and writes, while monitoring stays local with no telemetry. It requires macOS 14, and deep tracing may need administrator approval.
04:59 Ratel tackles tool overload by hiding most capabilities until an agent actually needs them. It indexes tool schemas and skill metadata, searches them with in-process BM25, then injects only the matching tools or instructions for that turn. Semantic and hybrid ranking are optional, so a vector database isn't required. The maintainers advertise roughly 80% fewer tokens across their tests.
05:25 Numbat watches what coding agents do on an endpoint without requiring a cloud service. It collects supported hooks, OTLP logs, and saved session artifacts into one event format, then applies local CEL rules for live alerts or forensic timelines. Pre-action blocking is available only on supported hooks, and ships disabled. Secret redaction helps, but records can still contain sensitive workstation context, so storage and access need care.
05:54 H New Hair brings Hacker News discussion directly onto the article being discussed. The user script searches for matching submissions, opens their comments in a resizable sidebar, and can blend multiple threads when an article was submitted more than once. It's clever part is quote-linked annotations. Matched passages in comments connect back to highlighted article text and can filter the thread.
06:20 Mirador is a terminal dashboard designed to be glanced at, not constantly watched. Its 12 rearrangeable panels cover tasks with due dates, local notes, calendars, weather, markets, and live system charts, while unfocused panels dim to reduce visual noise. Tasks, notes, and watchlists stay in editable TOML files with no account or telemetry. Network panels fetch external data only when enabled.
06:45 Snowfloat is a WebGPU snow demo where every visible element is generated at load time with no textures, meshes, HDRIs, or animation assets in the repo. Handwritten WGSL drives a roughly 870-m radius terrain clip map, persistent snow deformation with raised berms, and a procedural character with planted foot IK and simulated cloth. The author measured a 3.22 ms GPU frame on an RTX 5070 Ti at 1440p.
07:15 Kimiko 3 MLX ports MoonShots 2.78 trillion parameter model to Apple Silicon, and the finding underneath it is the interesting part. Prune experts using a Chinese corpus, and the model's coding ability collapses. Prune with English and code, and its Chinese collapses. Calibrate on both, and both survive. Experts cluster by domain, and Chinese and Python overlap below random chance.
07:40 Quen Scribe turns an Apple Silicon into a private transcription and system-wide dictation machine. When 3 ASR runs on the metal GPU, so audio and transcript text stay local after the model downloads. You can drop in media for word timestamps and SRT export, or hold a push-to-talk key to paste speech into any text field. It is still beta and source only, and save transcript history is readable unencrypted JSON.
08:08 Code Jury splits software delivery into six terminal-driven stages, then reviews each change with a configurable panel instead of one model. A persistent code graph and local semantic search help the planner find relevant symbols, while dev works on an isolated branch and QA runs tests. Nothing edits code before human ticket approval, and demo mode dry runs delivery by default.
08:32 Netherite is a from-scratch C and CUDA re-implementation of Minecraft 1.11.2, checked against the real Java client bit for bit, with batched CUDA reinforcement learning on top. Recorded tapes replay through both, and the demo renders an Oracle versus re-implementation video side-by-side. No Mojang content ships, so you supply your own copy. Turbo VLA removes the large language model from the center of a robot control pipeline.
09:00 It encodes vision and instructions separately, exchanges information through a lightweight bidirectional module, then predicts continuous action chunks with a compact decoder. On Libero, the researchers report 97.7% average success from a 0.2 billion parameter model, with 31.2 ms latency and 0.9 GB of VRAM on an RTX 4090. SpinWin rotates any macOS window without private APIs or disabling system integrity protection.
09:32 It moves the original window offscreen through accessibility, captures its live buffer with screen capture kit, then draws a borderless overlay using a core animation transform. You can pick 90° steps, arbitrary angles, or continuous spinning. The Illusion needs screen recording and accessibility permissions. Geo Look turns generative engine optimization from a monitoring dashboard into an execution loop.
09:56 It samples brand visibility across AI engines, audits the site for citation gaps, then produces implementation tickets and deployable assets such as LLMs.txt and JSON-LD. Later runs recrawl the project, compare question level results, and reopen regressed work. Control B Decompose turns huge log files into recurring structural patterns instead of dumping every line into your context window.
10:24 In one streaming pass, it normalizes timestamps and variables, clusters similar messages, then adds typed fields, P50 and P99 statistics, anomaly flags, and severity scores. Results can be rendered for a terminal, JSON, or compact model input. Audio 8 TTS packs multilingual speech generation and zero-shot voice cloning into a 0.6 billion parameter model.
10:46 It covers 11 recommended languages, bundles its own 44.1 kHz neural codec, and includes both batch inference and supervised fine-tuning tools. The team's published benchmarks place it near much larger systems, though they warn that differing evaluators make those comparisons approximate. Perception bench tests whether multimodal models can actually see before asking them to reason.
11:14 It contains 3,000 verified open-ended questions across 10 atomic skills, including counting, depth, localization, OCR, and hallucination resistance. The included evaluator works with OpenAI compatible endpoints and scores answers through a separate judge model. True Deck puts several coding agents on the same repository without replacing their native terminal interfaces.
11:38 OpenGrok, CodeX, Claude, Gemini, or Cursor in split panes, switch between tabs, and restore the whole workspace later. Its true memory layer keeps per repository and global context synced while MCP setup happens behind the scenes. Signative investigates infrastructure questions across source control, cloud, and Kubernetes as one read-only system. It fans API queries out through an ephemeral sandbox, cross-checks findings against live evidence, and records every tool call in a fail-closed JSONL audit log.
12:12 Credentials are attached only after policy checks with AWS sessions scoped through STS when possible. TanStack Charts is a typed visualization grammar that keeps your chart definition independent from the front end rendering it. A single definition can target React, Vue, Svelte, Solid, static SVG, or Canvas while native D3 scales and layouts remain available instead of being hidden behind another math API.
12:37 Responsive sizing, accessibility, interaction, and server rendering live in the runtime. Optum plans gives Claude and CodeX a structured path from vague requests to reviewed code with a human gate in the middle. The planning loop records decisions in versioned markdown. Then a reviewer or critic can challenge the draft before execution. Approved work runs from an immutable manifest in a dedicated Git work tree with read-only validation and checkpoints before integration.
13:09 Verified 3D mesh intersection does something rare. It proves a geometry kernel against a 93-line lean specification instead of asking reviewers to trust more than a thousand implementation lines. The checker verifies exact solid intersection and mesh well-formedness across every accepted input while a browser demo handles STL files locally. Soccer Crossing turns a real 3JS world into something that reads like a hand-painted anime background.
13:39 Every building, sign, lantern, and texture is generated at runtime while quantized lighting, depth-based ink lines, and inverted hull outlines flatten the scene into cell animation. You can walk the tiny spherical town, ride an e-bike, and call a train through the crossing. Quatu is a Linux verification browser built for coding agents that need measurable front-end evidence, not another screenshot to eyeball.
14:05 Its warm WebKit daemon returns pixel diff scores and heat maps, exposes animation timing as numbers, and can turn a headless session into a live window for human handoff without losing state. The maintainers report roughly 35-ms checks. Grafana AI SDK gives Go backends the interaction model many teams already use with Vercel's TypeScript AI SDK. It can stream server-sent events straight into React hooks like use chat, execute typed Go tools across multiple steps, and validate structured model output against schemas.
14:39 Provider adapters cover OpenAI and Anthropic while middleware adds retries, metrics, logging, and fallback. Interior is a copy-paste collection of React micro-interactions focused on the awkward half second after a click. Each component comes as one file with a headless behavior hook and a styled example, so you can replace the appearance without rewriting its state logic.
15:03 The interactions reserve layout space, handle abandoned gestures, support complete keyboard paths, and preserve information under reduced motion settings.