Seroter's Daily Reading: Your app’s frontend UI is now optional · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/66e68fc866cf9c9a04be442a020f2de11ab852a489629dff821a621fe6bf16d8.mp3

Source: https://seroter.com/2026/09/30/your-apps-frontend-ui-is-now-optional/

Today's post is Your app’s frontend UI is now optional, from Richard Seroter's Architecture Musings. Seroter opens with a blunt observation: he has more than three hundred apps on his phone and a browser stuffed with sites he rarely uses. His argument is that the era of learning a bespoke interface for every task is ending quickly, because AI agents still need the data and function behind an app, but they do not need us to see the app itself. Drop the user interface, go headless, and let an agent, harness, or super-app such as Gemini, Grok Bot, Meta Muse, or Claude Cowork do the work.

Seroter starts with the link he labels acknowledging, Ben Thompson's Stratechery essay Apps, Agents, and Aggregation. Thompson argues that messaging, AI operating computers, and the death of write once, run everywhere UI are not separate predictions but one current reality. He describes Meta Muse giving users a real virtual machine so an agent can build a personal recipe app in minutes while the user walks the dog. The wider point is that once agents abstract services into implementation details, discovery stops being the scarce problem and inspiration becomes the scarce one. Seroter's take is simpler: many are acknowledging that agents have made many apps unnecessary, and he still wants the app's data, but never wants to see it again.

Seroter then lays out the first category of developer work: create CLIs, APIs, skills, and MCP servers for anywhere access. First on that list is WebMCP tools. Chrome's proposed WebMCP standard annotates HTML form elements and exposes structured tools so agents can interact with pages reliably instead of scraping the DOM. Seroter says good models can navigate a frontend with computer use if they must, but spelunking the DOM wastes tokens and time, so adding WebMCP tools is the minimum.

Next is ClaudeForce. Salesforce and Anthropic are partnering so Salesforce data works inside Claude as a plugin with thirty-seven seller skills, governed by existing permissions. Seroter calls this a headless experience. Then Box too, from the Box blog. Box and Salesforce are positioning Box as the filesystem for AI, with governed content flowing through MCP into Claude, Slack, and Agentforce. Seroter says Box, HubSpot, and most SaaS products will follow this pattern, exposing data and commands without forcing users into their UI.

On the platform side, he cites Android CLI. Android's command line makes environment setup, project scaffolding, device management, and agent access to Android skills available from any harness. Google also published 150 agent skills. The repository covers Google Cloud, GKE, BigQuery, security, developer tools, and more. Seroter groups these with the catalog of remote managed MCP servers, where Google Cloud now offers managed HTTP MCP endpoints for services from BigQuery and Cloud SQL to Compute Engine and Kubernetes Engine. His point is that platform vendors are making UI optional.

Next, creating MCP servers points to FastMCP. FastMCP is a Python and TypeScript framework for building MCP servers, clients, and apps, with schemas, validation, transports, and authentication handled for the developer. Seroter says toolkits like this make it much easier to build these tools now. Generating skills points to Anthropic's Skill Creator, an interactive tool that guides developers through writing skill files and running evaluations. Cranking out a CLI points to Cloudflare's Forge, an open source, pluggable generation pipeline that turns an OpenAPI spec into SDKs, CLIs, docs, and even things like Cap'n Web or MCP servers. Seroter says developers can now generate the surfaces agents need without forcing users through a custom UI.

Seroter then moves to the second category: create A2UI or MCP Apps components for dynamic rendering. A2UI is his first link here. A2UI is a protocol from Google and the community that lets agents generate rich, interactive user interfaces by sending declarative component descriptions, not arbitrary HTML or JavaScript, so clients render native widgets safely across web, mobile, and desktop. Renderers details A2UI support for React, Lit, Angular, Flutter, SwiftUI, and Jetpack Compose. Seroter says developers should consider building components that agents can compose into beautiful, personalized sites.

The next item, building components, could not be retrieved. From Seroter's context, it points to launching MCP Apps in Google Cloud Toolbox. His advice stays the same: build portable UI components that users can render on whatever surface they are already using. Then he turns to MCP Apps. MCP Apps let servers return interactive HTML interfaces that render directly inside an agentic chat experience, with sandboxing, bidirectional data flow, and host integration. Build these covers the mechanics: Node.js projects, AI coding agents with a create-mcp-app skill, or manual setup with server tools and UI resources. Seroter says this is how chat users get more than boring text results.

Next he demonstrates retrieval and action from where he is. The remote MCP Server for Cloud Run documents how to connect agents directly to Cloud Run for listing, deploying, and managing services. Seroter says he has a reference to it in his Google Antigravity CLI and could also use a local CLI. Three screenshots illustrate the Antigravity setup: one shows the remote Cloud Run MCP reference in the CLI, another shows running Cloud Run services returned in the harness, and a third shows a hotel concierge agent used from inside Gemini Enterprise instead of a dedicated website. Two more screenshots cover the MCP App dashboard: one configures an MCP App running in Cloud Run as a data source, and another shows the full dashboard returned inside Gemini Enterprise chat.

That leads to Seroter's next activity: build on-the-fly visualizers, apps, and pages personalized to each user. The animated GIF here could not be fetched, but from the post it shows a hotel website built from pre-built A2UI components that swaps the experience based on whether the visitor is new, checked in, or a frequent guest. Seroter says instead of dozens of static pages, one page applies relevant components on the fly.

For the final activity, build wherever-you-want-it experiences, Seroter points to the new MCP server for the Google Cloud CLI. Google's remote server executes gcloud and bq commands on behalf of an agent, with restricted command support and IAM controls. Seroter says this is handy when an agentic client cannot execute CLIs, trusted or otherwise. Two final screenshots show the result: a Chrome plugin uses the MCP server through an agent to build a full Google Cloud management experience, and Seroter spins up a Pub/Sub topic while browsing his daily reading list.

Seroter finishes with a warning: static frontends will not become truly optional for a while, but developers should not wait too long. Personal agents are growing fast, and soon many websites and mobile apps will have more agent visitors than human ones. Build for that.

Your signing session ended