Product Updates & Releases

What’s new in Anvil

New features, engine improvements, model rollouts, and fixes shipped directly to the Anvil platform and CLI ecosystem.

Current Version

v1.32.0

Last Updated

October 27, 2026

Total Releases

65 Updates

System Status

Operational

4 weeks ago
v1.25.0MINOR
Platform Update
Added
  • anvil-studio now makes images for your own website or app screens: a landing-page hero, section art, an onboarding or empty-state illustration, a background or an OG image. Ask for "an image for the landing page" or "an empty-state illustration for the inbox screen". It reads the slot's size from your code (web, React Native and Expo, Flutter, Capacitor), keeps words out of the picture, and leaves the screen's code to you. App icons and splash screens are not covered yet.
  • studio.mjs publish copies finished images out of the studio folder, which git ignores, into your repo. It detects the kind of project and writes what it loads: a WebP for a website or Capacitor app, 1x, 2x and 3x files named for React Native or Flutter, the original file for a native iOS or Android app. It asks once where the files belong and remembers. A final keeps a stable name your code can point at, drafts keep their version number, and an OG image publishes as a 1200x630 JPG next to the page image.
Changed
  • anvil-studio picks the image model for you. It tries Nano Banana 2.1 first and moves to GPT Image 2.5 Sunburst when the first one fails, noting the switch in the project log; 4K images go straight to Sunburst. Name a model with --model to use only that one. Model names and prices live in one file, scripts/models.json, so adding or swapping a model is a one-file edit. Drafts are made at 1K or 2K and 4K is kept for finals shown large, and a shape both models cannot make (Instagram's 4:5) is refused before anything is charged.
  • The landing page tours now play on phones and tablets too. After the headline, the hero's pictures pin under the header and step through the Anvil workbench, the Prompter terminal, the top skills and getting started, each with its caption above. In "Skills in action" the chapters scroll normally under a small card that stays at the top of the screen. Its Spec, Studio and Davinci tabs light up as you reach each chapter, and a bar under each tab fills as you read through it. Short screens and reduced motion still get the plain stacked layout.
Fixed
  • Reference images sent to Nano Banana through anvil-studio now reach the model; they were sent in a field it ignores.
4 weeks ago
v1.24.0MINOR
Platform Update
Added
  • anvil-studio now auditions voices before it records a voiceover. It plays you up to five samples of your script's first line, with at least two female and two male voices, in the gallery, and records the full voiceover only with the voice you pick (--yolo takes its recommendation, and you can still swap at review). Every voice speaks your copy language. Free voices (HeyGen, then local Kokoro) come first; kie.ai's Gemini 3.8 Flash TTS voices are used only when you ask for them, with the language set on every request and a shortlist of voices by use and gender.
  • anvil-studio now plans every video around one object your product owns. Before the directions it notes what the name means, what your product does that a simpler one would not, the things on its screen and the moment it hurts, then picks one object that appears in every beat and becomes your mark or the finished result at the end. Each plan states why the idea could not belong to a competitor or a simpler product, keeps the words inside a budget for its length, and ends on a held frame with your one practical fact (date, offer or link) that works as the poster.
  • anvil-studio now watches a video before it shows it to you. It lays the film out as a sheet of frames every quarter second, finds any stretch where nothing moves, and measures loudness, then judges the hook, the thread through the film and the ending from them. --audit on a video does the same, so its report no longer says motion and sound were not reviewed.
  • studio.mjs lint checks a studio project for banned fonts in its canvases and banned words in its image prompts. A hit warns when a render is made and stops a piece from being marked final; fonts that come from your own brand are kept.
  • kie.mjs image takes --resolution 1K, 2K or 4K for kie.ai models that accept an output size, so anvil-studio can draft at one size and render a print final at a larger one.
Changed
  • The landing page now introduces Anvil as the home of Prompter. Scrolling the hero plays a short tour that opens on an Anvil workbench, with a proposal, a finished poster and a three-screen design laid around the mark, and hands off to the Prompter terminal and its top skills. Further down, a second tour walks through three skills step by step: the spec-driven workflow (proposal, apply, archive, each started by its prompt), Anvil Studio (a real run from a one-line idea to a finished poster) and Davinci (preview, three screen sizes, self-critique, implement). A grid presents the eight top skills and an access section explains how to get started, in English or Indonesian. Pricing, the waitlist, the video and the FAQ left the homepage; the dashboard and docs take their place, and the page's metadata now describes Anvil and Prompter.
  • anvil-studio renders films frame by frame and encodes them once, so flat colour no longer smears, and checks every film's size, frame rate, colour tags and audio before it reaches your gallery. A vertical version for Reels, TikTok or Shorts is now its own layout built for the tall frame instead of a crop of the wide one, and the mp4-1080 export no longer re-encodes a film that is already ready to post.
  • anvil-studio critiques more like an art director. It asks a second, fresh reviewer to challenge its findings before a piece becomes final, weighs problems on the opening, the payoff and the ending more heavily, checks any shot of an app for dates, counts and labels that disagree, and gives every pattern that reads as AI-made a named fix.
  • anvil-studio no longer opens the HyperFrames editor on localhost when it makes a video. Once the film passes its checks, the studio renders it, adds it to your studio, and opens the gallery in your browser. The gallery now opens by itself for every finished piece, once per run; later revisions ask you to refresh it.
4 weeks ago
v1.23.0MINOR
Platform Update
Added
  • cerebro-wiki can now mirror your project-wiki into Cerebro memory, with project-wiki as the source of truth. cerebro-wiki init shows the plan once, then copies every project-wiki page except its log into a reserved wiki/project-wiki/ folder of your Cerebro wiki, splitting very large pages by section and adding a pointer from the wiki's overview. cerebro-wiki update then syncs without asking: it compares project-wiki with the last synced commit on your machine, rewrites only the pages that changed, removes pages whose source is gone, and rolls back pages mirrored from uncommitted edits you later discarded. Pages you wrote by hand are never touched, and finding changes never reads pages back from Cerebro.
  • cerebro-wiki lint checks a Cerebro memory wiki for broken and one-way links, orphan pages and missing frontmatter, fixes the mechanical problems itself and lists the rest for you. cerebro-wiki status reports, without changing anything, whether the wiki is behind project-wiki or behind recent project work, and what to run next. Every operation can now be called by name: init, update, answer, load, ingest, file-back, lint and status.
4 weeks ago
v1.22.0MINOR
Platform Update
Added
  • project-wiki now keeps cross-repo bindings current after bind. update refreshes cross-repo.md when the code that talks to a bound repo changes and checks those lines against the other repo's wiki, warning you when you change a route the other repo calls in a way that could break it (path, method, auth, request or response shape). lint checks every binding: it corrects a moved wiki path, confirms each endpoint is still called or served, and checks the contract against the other repo. status shows whether each bound repo's contract changed since your last check. The other repo is only ever read, never written, and a repo that isn't on this machine is skipped with a note.
Fixed
  • project-wiki can now be uploaded to claude.ai and the API: its description is under the 1,024-character limit and no longer contains angle-bracket placeholders.