What's New in GenMB: June to September 2026

We stopped writing these for a while. Here is everything that shipped between June and September 2026, what each thing does, who it is for, and what plan it needs. Five releases, no roadmap talk.

1.4.5, June: a bigger free tier and drop-anywhere uploads

The free upload limits went up a long way. A free app now gets 1GB of storage, 50MB per file, and 200 files. Paid plans get 500MB per file, 250GB per app, and 100,000 files. There is a daily burst cap on top of that so the free tier cannot be used as a CDN: 200MB a day on free, 50GB a day on Pro.

Uploads over 8MB now use GCS resumable sessions. The client asks for an upload URL, pushes 8MB chunks, and finalizes. A dropped connection resumes instead of starting over, which is the difference between a 400MB video uploading from a train and not.

The home page became a full-page dropzone. Drop a screenshot, a PDF, or a video anywhere on it and that file becomes the starting context for a new app. The chat input takes paste-from-clipboard and drag-drop too.

Who it is for: anyone whose app has real files in it. Plan: free, with the limits above.

Details: /docs/media-platform.

1.5.0, July: video transcoding and `<genmb-video>`

Upload an MP4, WebM, or MOV and a background worker transcodes it into an HLS multi-bitrate ladder, extracts a thumbnail, and makes a 3-second animated GIF preview. None of that happens on the request path, so the upload returns immediately.

The playback side is a Web Component:

```html

<genmb-video src="/assets/intro.mp4" controls captions chapters></genmb-video>

`

It plays the original MP4 the moment the upload lands, then swaps to HLS the instant transcoding finishes, over a server-sent event. Mobile Safari uses native HLS. Everywhere else gets bundled hls.js, and if that fails to load it falls back to the MP4 rather than showing a dead player. captions loads the generated caption track; chapters renders a chapter rail.

Audio works the same way, with a <genmb-audio> widget that gives you a waveform strip and a click-to-seek transcript.

Who it is for: course sites, product demos, anything with a hero video. Plan: free. (Not to be confused with AI video *generation* from a text prompt, which is a Pro add-on at /docs/video.)

Details: /docs/media-platform.

1.6.0, August: knowledge files

Open the Knowledge tab in the app editor and upload a PDF, DOCX, Markdown, or plain text file. A worker extracts the text, chunks it with overlap, embeds the chunks with OpenAI, and stores them in a pgvector table scoped to that one app.

From then on, both code generation and chat refinement retrieve the most relevant chunks and put them in the model's context. So "make the pricing page match our spec" pulls from your actual spec instead of the model's guess at what a pricing page looks like.

The home page learned the same trick from the other end. Drop a PDF product brief on it and the new app is scaffolded from that brief. Loom and YouTube URLs work too, though those only contribute the title and description, not the transcript.

Who it is for: anyone building against a document that already exists, which is most client work. Plan: free, up to 10 documents per app; Pro raises it to 100 and Business to 10,000.

Details: /docs/media-platform and /docs/vector-db.

1.7.0, September: image pipeline, Asset Library, social cards

Three separate things landed together.

Every uploaded image goes through a pipeline. EXIF metadata is stripped (so your users' photos do not ship their GPS coordinates), HEIC is converted to JPEG, and the image is re-encoded at 600, 1200, and 2400 pixels wide in both AVIF and WebP. You also get a blurhash placeholder, the dominant color, and an AI-written alt text. The <genmb-image> widget picks the right variant; a plain <img> tag keeps working untouched.

The Asset Library is a cross-app collection. Star a file in one app and it shows up in the library, so the next app can reuse it without a re-upload. Starring records a reference rather than copying, so if the source app deletes the file the star becomes a dead link. It is per-user, not per-workspace.

Shared app links get a social card. When someone posts your app's URL, the preview image is generated from the app itself: its title over its first image asset, tinted with the dominant color the image pipeline already extracted. No setup, no uploading an OG image.

Who it is for: anyone who shares a link to what they built. Plan: free.

Details: /docs/asset-library.

1.7.1, September: every app is an MCP server, and a quality pipeline

The biggest of the five.

Your app's functions are MCP tools. Every app serves its own MCP endpoint at /api/mcp/app/{appId}/mcp. Flip mcpExposed on a backend function and Claude, Cursor, or any other MCP client can call it, with the tool description and parameters read from the @description and @param annotations in your function source. Exposing your own app is free on every plan. Listing it in the public MCP registry is the paid part, and it needs a public app.

A quality pipeline runs on every generation. Before the build, generated code goes through a shared validate-and-heal stage: static checks, design checks, an OWASP Top 10 security scan, and an SEO scan, with an automatic heal loop that fixes what it finds. This matters more than it sounds, because roughly 97% of new apps are generated over the WebSocket path, which used to skip most of it. Image-to-app and codebase import run through the same stage. Refinement saves run static checks and automatic fixes too, and surface anything they could not fix rather than reporting success.

Required function secrets show up in the Functions panel. If a function reads an environment variable you never set, the panel now tells you which one, instead of the function returning a 500 forever.

Importing an existing codebase landed here as well, via zip or a GitHub repo. It is deliberately narrow: a Vite + React SPA or a pure static HTML/CSS/JS site imports cleanly. Next.js, Vue, Svelte, Astro, Nuxt, and CRA are detected and rejected with a specific message rather than being quietly mis-bundled into something broken.

Who it is for: everyone, mostly invisibly. Plan: free, including the per-app MCP endpoint.

Details: /docs/mcp, /docs/functions, /docs/security, /docs/seo.

What this cost you

Nothing, if you are on free: the upload tier, video playback, knowledge files (up to 10 documents per app), the image pipeline, social cards, MCP, and the quality pipeline are all on every plan. The free plan is 8 credits a day and unlimited apps.

Pro raises the knowledge cap to 100 documents per app and Business to 10,000. The full breakdown of what each plan includes is at /pricing, and the running list of releases is at /changelog.