Updates

Taking a screenshot, saving it, then finding it again in a file picker is three steps to do nothing. Paste Image to PNG reads the image already on your clipboard and hands back a PNG. The clipboard read happens in the page; the image is never sent anywhere.

The device mockup generator moved off the shared assets module onto its own rendering module and a dedicated editor. The visible change is control: frames, shadows and backgrounds are composited rather than picked from a fixed set of templates.

The same change added a React error boundary around tool rendering and a route-level error page. Before it, a crash inside one tool took the whole route down to a blank screen.

The room design visualizer takes a photograph of a room and lets you change wall colour, materials, patterns, lighting and furniture placement. It is built from separate modules for masking, inpainting, relighting, geometry and scene state rather than one model call, because each of those steps needs to be inspectable when a result looks wrong.

It also introduced the shared transformers loader that the AI tools now use, so a model downloaded for one tool is not downloaded again for the next.

The Markdown editor renders as you type, with the parsing and rendering both done locally. It joins the existing conversions in and out of Markdown rather than replacing them.

The header changed at the same time. Navigation is now generated from the category configuration instead of being listed by hand, which is what made a catalogue this size browsable without a search box.

The PDF side gained editing, form filling and form building, signing, redaction, password protection and removal, page reordering, scanning to PDF, and OCR.

The document chat tools (PDF, Word, Excel and a generic document reader) answer questions about a file you open, with the model running in your browser.

Speech to text and speech to subtitles arrived for both audio and video, and four media conversions landed alongside them: MP3 to M4A, MP4 to MOV, video to image frames, and video to animated WebP.

This change also fixed a header bug worth naming. The Permissions-Policy header carried microphone=(), an empty allowlist, which denies the feature to this site's own pages and not only to embedded frames. Every recording tool showed a start button that could never succeed, and reported the reason to the browser console rather than to the person pressing it. The value is now microphone=(self), which still prompts for permission on every use.

Two kinds of tool landed together here. The first kind is arithmetic anyone building on a model needs and nobody wants to redo: token counting, VRAM estimation, context window budgeting, API cost, GGUF file size, and embedding similarity.

The second kind runs a model in your browser: summarisation, transcription, text to speech, background removal and zero-shot classification. The weights download to your device on first use and the work happens there.

Also in this group: prompt template and injection tooling, a RAG chunker, a fine-tuning dataset converter, an invisible character detector, and an llms.txt generator.

Most browser-based PDF to Word conversion extracts a stream of text and drops it into a document, which loses every column, table and heading position the original had. PDF to Word writes DOCX with the text positioned as the PDF placed it.

The same change added a search alias table, so a query for a name a tool is commonly known by finds the tool even when that name is not in its title.

A browser cannot query a DNS record or read another site's response headers, so these are the tools that genuinely need a server. They run as thin, cached edge routes with no stored state: what is my IP, DNS lookup, WHOIS, HTTP headers and a social preview inspector.

The QR and barcode studios went the other way and are entirely local, because there is no reason for the contents of a QR code to leave the page.

The JSON formatter reports the location of a syntax error instead of failing silently, and the parsing module underneath it is shared with the validator, the diff and the JSON conversion tools, so all of them agree about what is valid.

This is the change that made the catalogue size manageable. Tool metadata, the code-split loader map and the route list are now generated from the descriptors by a script, so a tool cannot exist in the registry and be missing from the sitemap, or appear in search and have no route.

The image resizer and the interactive image workbench landed on top of it. The distribution layer was rebuilt in the same group of commits: RSS, JSON Feed and llms.txt had shipped in the first commit as one route each, and they were replaced by a shared renderer that also produced Atom, a feed per category, and a split sitemap index in place of the single flat file.

The repository starts with the descriptor contract rather than with tools. A tool declares its inputs, its output kind and a pure run function; the kernel renders it, tests it from its own examples, and generates its route, its metadata and its feed entry. Everything since has been built inside that shape.

The cross-origin isolation headers for the media routes date from this commit too, which is what made ffmpeg.wasm possible later without a site-wide policy that would have broken remote images everywhere else.