Practical AI Guides & a Free Claude Code Course

Whistle is a small speech recognition AI model developed by Cactus-Compute. Speech recognition refers to converting spoken language into written text. This model can be used in mobile phones, wearables, robots, home automation systems, vehicles and microcontrollers (small computers with limited memory). This model is just a single file of size 16.9 megabytes. It uses the same processing core as another AI model developed by Cactus-Compute called Needle. This model doesn’t require any GPU (Graphics Processing Unit) or additional software. Whistle performs the following tasks on a device: Transcribes up to 30 seconds of audio in English, German, French, Spanish, Italian, Dutch and Polish. Provides time stamps for beginning and end of each word along with its confidence score. Generates audio fingerprints for searching and comparing audio files. For example, if your device plays back a recording of ‘turn off the kitchen lights’ then this model will output the exact same text. This model is available under Apache 2.0 License. This means that you can freely download and modify this model according to your needs including commercial purposes. This model can be easily installed as a python package (Python is a popular programming language). Also, there is a C interface for this model that can be used in embedded systems (tiny computers). Ask your agent to go to this link of Whistle on Hugging Face (a platform for sharing AI models) and then ask it to install the package, recognize some english audio file and print the recognized text along with time stamps for each word. You can run this model on any normal computer or even on your mobile phone. No GPU is required.

storytold/vectorcraft

VectorCraft is, in the authors' own words, a vector illustration program: an open-source, clean-room rebuild of Adobe Illustrator (Adobe's paid design app), written from scratch in Rust, a programming language. It runs natively on macOS, Windows, Linux and FreeBSD, and in the browser, and keeps Illustrator's familiar tool layout — pen, direct selection, Pathfinder, Smart Guides, and panels like Appearance, Swatches and Layers — so there's almost nothing to relearn. Rendering is multithreaded and uses the processor's SIMD instructions (a way to do many calculations at once) instead of the graphics card, so 20,000 shapes render in about 27 milliseconds at full retina resolution while the interface still holds 120 frames per second. Every tool, menu and dialog is also reachable through a JSON control channel and an MCP server (a standard way for AI agents to call outside tools), so an agent like Claude can draw, edit and export through the same command path a human uses, not through a layer of scripts bolted on top. The README states plainly that every illustration shown there, gradients and posters included, was made entirely through that same interface, which is the main proof that the agent route actually works and isn't just a demo. The authors are upfront about where the project stands: by their own estimate, roughly 69-75% of Illustrator's features exist and work, and only 40-55% reach the point where a professional couldn't tell the difference. What's missing is listed too: 3D and Photoshop-style raster effects, scripting, interaction-detail polish, and proper installers for Windows and Linux, so right now it's a solid foundation rather than a finished replacement.

storytold/pdfcraft

PdfCraft is a free, open-source program for reading and editing PDF files (a common document format). It's built in Rust (a programming language) by ArtCraft, a community of people who build creative apps together. The authors call it a PDF workbench — an open clone of Adobe Acrobat. It's built from scratch, without using any of Acrobat's own code. It runs on macOS (including Apple's own chips), Windows, Linux, FreeBSD, and even right in a browser. Your document never leaves your machine — no account, no tracking, no cloud step. Inside it there's an agent layer: an MCP server, the standard way AI agents talk to tools. It stays off by default and never opens a network port at all. It only talks through stdin/stdout, a local text pipe to the program on your own machine. So it never makes a network connection at all. Once an agent starts it, the agent gets the same tools a person has in the app. Search and pull out text, flip through and reorder pages, reply to comments. Fill form fields, password-protect, merge and split documents, undo and save. To try it, open the repository on GitHub. Ask your agent to set up PdfCraft, or build its pdfcraft-cli (command-line tool), following the README (the project's instructions file). Then turn on its MCP server, pointed at a folder with a few of your PDFs. Have the agent find a piece of text and list the form fields in one of them.

storytold/lightcraft

LightCraft is a free, open-source app for organizing and developing photos. A team called ArtCraft is building it from scratch in Rust, a programming language. It contains none of Adobe's original code. The authors call it a clean-room rebuild of Adobe Lightroom, the well-known photo-editing program. LightCraft runs natively on macOS, Windows and Linux. It also runs in the browser, using WebAssembly, a way to run native-speed code on web pages. There's no account, no cloud, and no telemetry, meaning no usage data gets sent back to the makers. It needs no subscription either. Your photo catalog is stored as a plain, readable log of edits, so you can compare or roll back changes line by line. On a 24-megapixel RAW photo — a camera's unprocessed image file — tested on an Apple M4 Pro chip, moving a slider redraws the image in about 4 milliseconds. A quick preview takes about 30 milliseconds, and a full export takes about 0.3 seconds. Without a graphics card, the same work runs on the regular processor instead, just slower. Every slider, brush stroke and menu click in LightCraft is its own named command. LightCraft also ships an MCP server — a connector that lets an AI agent send it those same commands. You start it with lightcraft-cli mcp. Through this server, an agent can pick out a photo. It can develop the RAW file, adjusting exposure and light. It can paint a mask over the sky. Then it can export the finished photo. The agent can do all this in the background. It can also do it inside the open app, clicking buttons and taking screenshots as it goes. The screenshots in LightCraft's own documentation were made exactly this way. By the authors' own count, LightCraft now covers about 79% of Lightroom's total features. For the core features only, that number is 98%. As an everyday replacement for daily editing work, the authors put it closer to 60-70%. It works best with JPEG and DNG files, and with most Nikon, Sony and older Canon RAW files. One honest limit: sky and subject masking look like an AI feature. In fact, it runs on classic computer-vision heuristics — fixed, hand-coded rules, not a trained AI model. LightCraft also has no AI-based noise reduction yet. It has no HDR either, which stands for high dynamic range, a way to show both bright and dark detail at once. The authors' roadmap lists a future option to plug in Meta's SAM 3 model for some mask types. That model would be a separate, user-approved download. It is not part of the app by default. To try it yourself: install LightCraft from github.com/storytold/lightcraft, following its README, a file that explains setup. The simplest route is the installer at getartcraft.com/apps/lightcraft. Import a test folder of RAW or JPEG photos, start the MCP server, and ask your agent to pick one photo, develop it, mask the sky, and export it.

storytold/deckcraft

DeckCraft is a project by ArtCraft, a small studio that makes open creative tools. It rewrites PowerPoint, Microsoft's presentation software, from scratch in Rust, a programming language, with no Microsoft code inside — the team calls it a clean-room clone, meaning it was built without looking at Microsoft's own code. Today it already has slide masters, eleven layout templates, eight themes, over 150 ready shapes, tables, charts, basic SmartArt (ready-made diagram shapes), transitions with a Morph effect, and animation by paragraph. You work in it with your mouse and menus, the same way you work in PowerPoint. A finished deck opens and saves as a real .pptx file, PowerPoint's own file format, or exports to PDF, PNG or JPEG. By the team's own estimate, DeckCraft covers about 79% of PowerPoint's features by breadth and 93% of the core ones, and is roughly 80% of the way to its first alpha version — a working draft, not a finished product. Every action in the program is also a command with undo, and there are over 200 of them. That means an AI agent can call these same commands too, through MCP (a protocol that lets AI agents control apps) or a JSON control channel, building slides, tables and charts on its own. The project appeared only a few days ago, so don't expect stability yet, but ready installers for Windows, macOS, Linux and FreeBSD are already in the Releases section on GitHub, so you can install it without touching any source code.

storytold/effectcraft

EffectCraft rebuilds Adobe After Effects from scratch. Adobe makes After Effects, the standard program for video effects and animation. EffectCraft is written in Rust, a programming language built for speed and safety. It comes from ArtCraft, a team with similar free, open programs: PhotoCraft for photos and FilmCraft for video. The first commit landed on October 1. Within a week the project had over 1,600 stars on GitHub, a way people bookmark and vote for projects. Inside it has layers, keyframes, 3D cameras and lights, and 306 effects. A keyframe sets a property's value at one exact moment; the program fills in the motion between keyframes on its own. The 306 effects cover all 298 effects from After Effects, plus more. You can export animations as Lottie, a format that drops straight onto a website with no plugin. Rendering — turning your project into a finished video file — skips FFmpeg, the usual open tool for this. EffectCraft wrote its own video encoders in pure Rust instead, for H.264, ProRes, HEVC and AV1. Project files are plain JSON, a readable text format. You can open one as text and see its edit history in Git, the tool that tracks file changes. Two optional features borrow from outside: cutting out objects (Roto Brush) can use MobileSAM, a small recognition model built on Meta AI's Segment Anything and Microsoft's TinyViT. Face tracking can use Google's Face Landmarker. The makers themselves say it isn't ready to replace After Effects on paid client work yet. Nobody has checked frame-by-frame if it matches the original, and Linux and Windows still have small interface bugs. To install: grab the installer for your system — Mac, Windows or Linux — from the project's Releases page on GitHub, or just open it in your browser with no install needed.