New on BuildTube: 2026-39

Impeccable is a design skill for AI coding agents that pushes back against the handful of visual habits (the same font, the same gradient, cards nested in cards) that show up on nearly every AI-generated interface.

Describe what you want to build and have your AI coding agent ask questions, agree a written design with you, and only then write code from a step-by-step plan with tests.

A curated index of step-by-step guides for recreating the technologies you use every day.

Project NOMAD turns a small computer into a self-contained library and AI assistant that works with no internet connection: offline Wikipedia and reference books, Khan Academy-style courses with progress tracking, downloadable regional maps, and a local AI chatbot you can hand your own documents to.

Multica is a project-management workspace where AI coding agents get assigned issues the same way a human teammate would: they pick up the work, comment on progress, raise blockers, and hand it back for review, all visible on the same board as your human team.

json-render is a programming toolkit (framework) for "generative UI": instead of letting a model write arbitrary code for a screen, you give it a fixed catalog of components it is allowed to use, and it returns structured JSON describing which ones to show and how, which the framework then renders.

Run a 27-billion-parameter model that reads text, images and video and writes text back, on hardware you control.

A local-first desktop app that gives each of your AI agents a character at its own workstation in a pixel-art space station, with the station's rooms and activity reflecting what the agents are actually doing.

Jev Ultrafast is a browser-automation agent that completes a task like searching flights by reading the page, then asking Jev which button, field, or link to act on next — a decision that takes milliseconds instead of the seconds a normal chatbot needs to 'think' about it.

NanoJev is a small (0.6-billion-parameter) open reproduction of Jev: a model that reads a game or task state and a set of possible actions and returns a full probability distribution over them in one pass, with no text generation.

This project pairs two AI systems to play Minecraft: a planning model (the creator's own 'Astra') decides the overall strategy, while Jev picks the moment-to-moment actions — movement, camera angle, item use — needed to carry it out, sent through the normal player controls rather than by editing the game's code.

TRELLIS.2 is Microsoft's image-to-3D model: give it one picture of an object and it generates a full 3D asset — a mesh with physically based rendering materials, including transparent or translucent parts — using a 4-billion-parameter model and a new “O-Voxel” representation designed to handle shapes that older methods struggle with, such as open surfaces and thin structures.

Hum to Song takes a hummed melody (10 to 30 seconds), plus a style line and lyrics you provide, and turns it into a produced song that keeps your original tune and builds a full structure around it.

Qwen-Image-2.1 is Qwen's own official demo of its unified text-to-image generation and image-editing model.

Paint which neurons to stimulate on a real fruit-fly connectome and an animated 3D fly responds by walking, turning or flying, computed live in the browser via WebGPU.

Control already-trained reinforcement-learning policies for Pollen Robotics' Microduck robot entirely client-side: physics run through MuJoCo compiled to WebAssembly and the policy network runs via onnxruntime-web, with no server involved.

Live Speaker Diarization is Nvidia's browser demo for figuring out “who spoke when” in audio: start your microphone or upload a short recording, and it labels overlapping speaker turns in near-real time, showing speaker-activity lanes alongside a live transcript.

This Space runs System One, an independent reproduction of TypeSafe's Jev idea: Qwen3.5-4B-Base fine-tuned with a small scoring head so it answers a typed question with a calibrated probability across exactly the options you supply, instead of writing a free-text answer.

gods-eye-view opens a photorealistic 3D globe in your browser and plots public live feeds on it: aircraft transponders, ship beacons, satellite orbits, earthquakes, and public traffic and security cameras.

VoiceStudio is a desktop app for generating speech in a chosen voice without sending anything to a company: give it a clean reference recording, type your text, and it generates the audio, with separate workspaces for dubbing video and for designing a brand-new voice from a written description instead of a sample.

Supervision is a Python library of the parts that sit around a computer vision model rather than inside it: loading datasets, drawing boxes and masks, tracking an object across frames, counting what crosses a zone, and measuring how well a model did.

YuE2 is a music generation system that writes a song in two visible steps: given lyrics and a style description, it first produces a melody-and-chord plan in ordinary notation, then turns that plan into finished audio with vocals and backing.

This is a catalogue of video games whose source code is published, sorted into the genres a player would look under: action, adventure, city building, first-person, platformers, puzzles, racing, real-time strategy, roguelikes, role-playing, sandbox, shoot-em-ups, sport, tower defence and turn-based strategy.

Give it lyrics and a note about the style and it generates a song with vocals and backing.

llama.cpp is a C/C++ program with no external dependencies for running language models, including models that read images, on hardware such as laptop CPUs, GPUs and phones.

Nasiko is a self-hosted control-plane server that sits in front of every AI agent you run and handles the operational questions that pile up once you have more than a couple: which agent is allowed to call which other agent, what did that call cost, and why did it fail.

Treg is a marketplace and proxy for AI-agent tools such as SEO data, company lookups, scraping and image generation, with over 3,000 endpoints across 60-plus providers.

ai-memory is a self-hosted memory server that lets more than twenty different coding-agent tools — Claude Code, Codex, Cursor, Gemini CLI and others — share one memory instead of each keeping its own that vanishes when you switch tools.

Paperclip is an open-source dashboard for hiring and running a whole team of AI agents toward one goal, like "build the #1 AI note-taking app to $1M MRR." You define the goal, assign roles such as CEO, engineers and marketers, and set a budget; the agents come from tools you already use — Claude Code, Codex, Cursor or OpenClaw — while Paperclip tracks status, cost and progress from one screen.

CC Switch is a desktop app that manages provider settings for a dozen-plus AI coding tools — Claude Code, Codex, Gemini CLI, OpenCode and others — so switching from one model provider to another does not mean hand-editing JSON, TOML or YAML configuration files.

Hindsight is a memory system for AI agents, built to notice patterns across many conversations rather than only replaying past ones the way a typical chat history or a RAG lookup (search over stored documents) does.

Find Mac apps for a job, such as a screen recorder, a Markdown editor or an AI assistant, by browsing one long list sorted into categories.

CLI-Anything is a registry and toolkit for generating command-line interfaces that let AI agents operate software that was not originally built with agents in mind, from design tools to specialized desktop applications.

OpenCreator (formerly KrillinAI) is a desktop app that bundles AI creator tools into one workspace: write articles or short-video scripts, generate images and video from a prompt (through models like Seedance, Kling, or GPT Image), add AI voiceovers and dubbing, translate and re-subtitle existing videos, and turn a script into a narrated stick-figure animation.

Route your phone or computer’s traffic through proxy servers from your own provider, and manage those connections from one app on Android, Windows, macOS and Linux.

This is the original "Awesome list": a single hub page linking out to hundreds of curated lists of tools, libraries, books and resources, organized by topic — programming languages, security, databases, hardware, and far beyond software into things like business and health.

Octop is a self-hosted AI assistant that runs entirely on your own machine and gives each person in a household or small team their own set of specialized agents to switch between — one for weekly reports, one for scheduling, and so on — while keeping every conversation and credential local.

HydraDB is a graph database (a database built around connections between records, not just rows and tables) written in Rust that keeps its durable data in S3-compatible object storage instead of on a server's own disk.

Univer is an open-source software toolkit (SDK) for embedding spreadsheet, document and presentation editing into your own product, rather than a finished app you open.

A collection of open-source pieces for letting AI agents use a computer rather than an API: drivers that move the mouse and keyboard, isolated cloud desktops, local macOS virtual machines, small models trained to make computer-use decisions, and a benchmark for scoring how well an agent did.

A desktop interface for Clash Meta, a rules-based network proxy engine, letting you manage server profiles, edit routing rules visually and switch the whole system between direct and proxied traffic without editing configuration files.

This is a collection of reference agents, skills and data connectors built for financial-services work — investment banking, equity research, private equity and wealth management — that plug into Claude as Cowork plugins or run through Anthropic's Managed Agents API.

GitDiagram reads a public GitHub repository, or a private one with your token, and draws an interactive architecture diagram with each piece linking to its code.

codebase-memory-mcp is an MCP server (a tool that plugs into AI coding assistants) that reads a codebase and builds a searchable knowledge graph of its functions, classes, call chains and cross-file links, so an agent can answer structural questions about the code in under a millisecond instead of reading files one at a time.

Claude Code Templates is a catalog and installer for over 100 ready-made agents, custom commands, hooks and MCP integrations for Anthropic's Claude Code, browsable through an interactive web interface (aitmpl.com) or installed directly from a terminal command.

Agent Substrate is infrastructure software for running very large numbers of AI agent sandboxes efficiently, built on the observation that agent-like programs spend most of their time idle.

The Mobile Verification Toolkit (MVT) is a set of command-line tools for gathering forensic evidence from Android and iOS devices that may help identify a spyware infection.

This is a hands-on tutorial that walks through building a language model entirely from scratch in PyTorch, with no shortcuts through libraries like transformers, trl or peft that normally do the heavy lifting.

AutoClip is a desktop app that watches a long video — a podcast, stream or lecture — and automatically finds and cuts the moments worth sharing, using an AI model you connect (you choose the provider in settings).

Strands Agents is a software development kit (SDK) for building AI agents without writing your own agent loop (the code that lets a model plan, call tools and repeat) from scratch.

Bend is a new programming language built around the idea that AI-written code should come with a checkable proof, not just a plausible-looking diff.

quiche is a Rust software library that adds QUIC and HTTP/3 — the newer protocols behind faster, more reliable web connections — to another program's networking code.

Needle is a small foundation model built specifically for devices too limited for a normal chatbot model — phones, wearables, smart-home hubs, robots and microcontrollers — packaged as a single binary as small as 8 to 29 megabytes.

Higgsfield is an open-source framework for training very large neural networks — the kind with billions to trillions of parameters, such as large language models — across many computers at once, a process the project's own documentation describes as normally involving a lot of manual, error-prone setup ("multi node training without crying").

Bonsai Demo is the official runner for PrismML's Bonsai family of heavily compressed language models, letting you run a 27-billion-parameter class model on a Mac, a single GPU, or plain CPU.

WebCodex connects cloud-based AI chat tools like ChatGPT and Claude to a real development environment running on your own machine, so you can ask them to inspect a repository, modify code, run tests, use Git, or investigate a failure without moving your project into a hosted workspace first.

A self-hosted chat application that puts a single interface in front of many model providers at once — OpenAI, Anthropic, Google, AWS, DeepSeek and others — so switching models does not mean switching product.

A small model built to make decisions rather than write text.

Qwen-Image-2.1 is a 7-billion-parameter image model from Alibaba's Qwen team that both generates pictures from a text description and edits existing ones in the same model, using up to 10 reference images and letting you mark local edits with circles, painted annotations or masks while preserving the identity of people and products shown.

Run a 27-billion-parameter reasoning model from a roughly 6 GB file on a laptop or a single GPU.

Audio8 ASR Infinite is a speech-to-text model built for real-time, unlimited-length transcription rather than processing a finished audio file after the fact.

This repository repackages Alibaba's Qwen-Image-2.1 image generation and editing model as single-file weights formatted specifically for ComfyUI, a popular node-based visual tool for building AI image and video pipelines by connecting boxes on a canvas rather than writing code.

Hemmingway-1 is a 27-billion-parameter language model (built on Qwen3.5) fine-tuned specifically for the everyday writing people actually do — a text to a landlord, a work email, an awkward note to a colleague — rather than the longer, hedged answers general chat models tend to give.

MiMo-V2.6-Pro-RL is Xiaomi's flagship model in the MiMo-V2.6 family, reading and generating text, images, video and audio in one model with a 1-million-token context for handling large codebases or long multi-session agent work.

MiMo-V2.6-Distill-Qwen-9B is a 9-billion-parameter model made by fine-tuning Qwen3.5-9B on data generated by Xiaomi's larger MiMo-V2.6 models, a technique called distillation where a smaller model is trained to imitate a bigger one's outputs rather than being trained on raw internet text from scratch.

Give one model text, images, video or audio, with up to a million tokens of context such as a whole code repository or a long tool log, and get text back for coding and agent work.