What changed, and when.

Twelve new screens, seven components, four effects

More to start from: maps, camera, photos, video, travel, and the first iPad and foldable screens.

• New screens for Pro: maps with a place sheet, ride booking, a travel stay, a music library with a mini player on the tab bar, a video feed, a camera viewfinder, a photo library, a paywall sheet, Control Center, a web landing page, an iPad product page and a foldable mail app. Each one in light and dark, live in the gallery, with its prompt and React source.
• New components, free: the tab bar accessory (a mini player that rides on the tab bar), the map callout, the map control stack, the page control, the shutter, the Control Center module and the vertical rail. The toolbar gains its grouped look: two or three capsules with one prominent action.
• New effects, free: materialize (glass appears by bending in, not by fading), touch illumination, background extension and light spill, with the web code and the SwiftUI calls.
• Ask for a whole screen (a map explorer, a camera) and the brief carries that screen's own pieces, says which parts are flat, and names what is under the glass; get_template lists the colour tokens its source uses.
• SwiftUI: a map with a sheet that leaves the map usable, MapKit's own controls, and when to build your own. Web: a class for white labels over a camera or video.
• White labels on accent and red buttons now pass 4.5:1 in dark mode too.

Copy a prompt: liquid glass without installing anything

Two ways to use Meniscus now: copy a prompt into any AI, or connect your AI once.

• Every component and effect has a prompt for web, SwiftUI, Flutter, React Native and Figma, written for Claude Code, Codex, Cursor or a chat app: the rules, the numbers, the glass code and the checks in one text. Free, no sign-in.
• Pro adds prompts for the whole screens and an app builder that puts two or three of them in one prompt.
• A new site: the gallery opens any piece with its prompt, the setup connects an assistant in four steps and shows when it is connected, and pricing shows the plan you have.

Seven tools, made for coding agents

• Your agent starts with get_brief: the rules, tokens for both themes, the specs the screen needs, the build code for your stack, two reference images and how to check it: check_code first, then screenshots against the checklist.
• New tool, check_code: checks the code your agent wrote against the rules, without a browser, and lists the errors, the warnings and what only a screenshot can confirm. The review_glass prompt starts with it.
• Fewer, plainer names: get_spec takes any component or effect (add a medium for its build code), get_rules gives the rules as one block to keep in CLAUDE.md or AGENTS.md, or one section at a time, and get_references both searches and loads. get_recipe adds mounting in plain JavaScript and a QA script for every medium.
• create_design_brief, get_component_spec, get_effect_spec, get_guidelines, search_references and get_platform_support are gone. If your CLAUDE.md, AGENTS.md or Cursor rule names create_design_brief, change it to get_brief; reconnect your client if it still lists the old tools.

A plugin for Claude Code and Codex

• Meniscus installs as a plugin in Claude Code (/plugin marketplace add meniscus-site/meniscus-mcp) and in Codex (codex plugin marketplace add meniscus-site/meniscus-mcp): the server plus the liquid-glass skill, which sends the agent to the tools whenever you ask for glass.
• Guides for Claude Code, Cursor and Codex: setup, the one file that makes each agent use Meniscus unasked (CLAUDE.md, a Cursor rule, AGENTS.md), prompts to start from and what to expect.

Whole screens as code, and 14 days of Pro

• New tool, get_template: ten whole screens as React source — smart home, wallet, AI chat, AI workspace, fitness, weather, lock screen, music player, SaaS dashboard and a desktop library. Briefs name the closest one.
• The React kit: a Vite and React 19 project with the 34 components, the ten screens and the WebGL effects, as source you own. Pro, Team and Lifetime download it from the account page.
• Every new account gets Pro for its first 14 days, then the free plan, with nothing to cancel. The source downloads need a plan.

The free plan builds on every platform

• Free accounts now get every recipe with its code: web, Figma and design tools, SwiftUI, Flutter and React Native. Before, a SwiftUI, Flutter or React Native question got a medium-neutral answer and a note; now it gets the same code the paid plans get. Free is 50 calls a day, up from 30.
• The paid plans add the whole reference library (any render by id, at full frame, up to 4 per spec), the React Native starter and fair-use volume.
• Your assistant now knows when to reach for Meniscus without being told: the server describes itself to clients that read server instructions, and offers three prompts (build a screen, review glass, add a component) that Claude Code lists as /mcp__meniscus__ commands.
• New on the site: a free liquid glass CSS generator, build guides for CSS, SwiftUI, Flutter, React Native and Figma with the full recipe for each, and sign-in with a code you type in the same tab.

The code now reaches your agent

• Assistants write any answer over about 50,000 characters to a file and hand the model only its path; Claude Code and Cursor both do. The design language plus the web recipe was 61,000, so an agent read the code in pieces or not at all, and often built blur-only glass. Every answer now stays under 45,000 characters, and a test keeps it there.
• New tool: get_recipe returns a medium's build recipe with its code, whole or one section at a time — the SVG displacement map, the markup and CSS, the React component, the rules that make or break it, the WebGL shader, liquid metal, border beam. get_guidelines keeps the design language and lists the sections.
• A brief now carries the code its screen needs: for the web the displacement map, markup, CSS and rules, the WebGL shader when the intent names one, the React component for React stacks, liquid metal and border beam when named. SwiftUI, Flutter, React Native and Figma briefs carry their whole recipe. One call, working glass.
• An effect spec asked for a medium attaches the section that implements the effect there: refraction brings the map and the rules, dispersion, caustics, iridescence and morph bring the shader.
• Every tool tells assistants that read the hint how large its answers may be, so none is written to a file on the way to the model.

Glass that visibly bends

• The edge profile is now a rounded bevel. The old one packed the bend into the outer one or two pixels, so glass looked merely blurred at small settings and mirrored at large ones. Now the backdrop is pulled in and magnified across the whole edge band, and circles magnify like a lens.
• Displacement is a fixed share of the bezel: 0.4 × on panels, bars and buttons, 0.5 × the radius on circles. The same numbers hold over photos, text and grids, and none of them folds.
• The profile once called squircle is now called bevel; corners are still squircles. Every spec, the web, WebGL and Flutter recipes and every reference image were updated and re-rendered to match.

Setup for every assistant

• The docs now walk through 32 clients, each as its own docs describe it: Codex, Gemini CLI, Copilot CLI, OpenCode, Amp, Zed, JetBrains, Visual Studio, Kiro, Antigravity, ChatGPT, Perplexity and more, next to Claude Code, Cursor and VS Code.
• One-click install for Cursor, VS Code and LM Studio.
• Gemini CLI 0.60 and later can sign in: the sign-in now tells your client which server answered, as newer clients require.
• Clients that run only local servers connect through the mcp-remote bridge; the docs give the config.

An interactive viewer in your assistant

• In Claude, ChatGPT, VS Code and other assistants that show MCP Apps, specs, briefs and reference searches now open as a viewer under the answer: the reference image at full size with what to look for, the brief's colour tokens as swatches over a striped backdrop, and a checklist you can tick.
• From the viewer you can load more reference images, open the spec of any component a reference shows, and hand a reference to your assistant as the look to aim for.
• Assistants without MCP Apps get the same text and images as before, and the viewer adds nothing to what your agent reads.
• Reference images now carry a thin Meniscus strip along their bottom edge. On crops it sits below the glass, never over it, and your agent is told it is not part of the design.

Web recipe: a silent Chrome failure, closed

• A test build mounted the refraction filter through an HTML template. Chrome dropped the whole glass effect without an error, leaving the bars fully clear. The web recipe now says where the filter goes and how to check it took.
• The recipe now says why its browser check must stay as written: shortened, it misses headless Chrome, and screenshot tests see only the frosted fallback.

Ten screen templates and a hosted server

• Ten screen templates your agent can start from, up from two: smart home, wallet, assistant, AI workspace, analytics, fitness, weather, lock screen, player and a desktop library.
• 109 reference renders in light and dark, on redrawn backdrops: satin ribbons, soft clouds, lit windows. The scenes kept their shapes, so every note on what to look for still holds.
• A hosted server at https://meniscus.site/mcp with OAuth sign-in: add one URL to your assistant, nothing to install or update.

Updated for iOS 27

12 days after Apple’s release

The design language moves to the iOS 27 generation of the material:

• Stronger diffusion over busy content: regular and frost glass sit at the top of their blur range over text, photos and feeds.
• A darker look at the edge: a 1 px dark edge line now runs just outside a brighter specular rim.
• A clear-to-tinted setting lets the viewer make all glass clearer or more opaque; designs must hold at both ends.
• Sidebars on desktop and tablet run edge to edge instead of floating as an inset card.
• Desktop windows use continuous corners of about 20 px (26 px on macOS 26).

Refraction without folds; every image our own

• Displacement limits re-derived on the pixel grid — at most 0.7 × √bezel for squircle edges and 0.9 × √bezel for convex ones — so lines bend once at the rim instead of looping back.
• The refraction references were re-rendered to those limits.
• Every reference image is now rendered by us from the Meniscus showcase over procedural backdrops. No third-party screenshots.

Flutter and React Native recipes, built by agents

• A fresh agent built a Flutter app from the recipe alone: flutter analyze and an Android build pass, and the Impeller refraction shader runs on Android and iOS.
• A fresh agent built a React Native screen on the Expo starter from the recipe alone; TypeScript passes.

References cropped to the glass

• Each reference marks where its glass is, so an agent sees the image cropped to the edge that matters, with notes on what to look for.
• Specs attach the one image that shows them best, which keeps briefs small.

Specs for design agents too

• Every rule is stated as what you see plus the numbers, so design agents — in Figma, on mockups, in critique — read the same specs as coding agents.
• The design-tool recipe: Figma's Glass effect settings, a manual layer stack for tools without one, and wording for image-generation prompts.