AI Security & Safety

MCP Architecture and Future

Last updated 2026-09-25

What's new

2026-09-25

Meta has opened up Muse (a personal AI assistant) to developers, allowing them to create "connectors" (tools that let Muse use their services to help users with tasks).
This is like the App Store moment for AI, where businesses can build tools that help Muse complete user requests, potentially creating a new market with billions of dollars at stake.
Muse connectors can become useful mid-task, like when someone organizing an event realizes they need to rent equipment, and a connector for a rental service could step in.
Connectors work by using APIs (a way for software to talk to each other) to connect Muse with a business's services, making it easier for users to access and pay for those services.

2026-09-10

Always keep passwords and secret codes (called "credentials") out of your code, emails, or shared documents to reduce security risks and make it easier to update them later.
Use special encrypted storage (like "AKMS" in cloud computing) to protect these credentials when your application is running or being updated.
Regularly update and revoke old credentials to minimize damage if they are ever exposed, and use separate credentials for different stages (like development, testing, and deployment) to prevent a single leak from affecting everything.
Monitor your tools and set spending limits to catch any unusual activity or exposure early, before it becomes a bigger problem.

2026-08-31

Figma (a design tool) built an MCP (Machine Communication Protocol) server to let AI tools share info between code and design without needing special integrations.
They started with a local MCP server for developers (people who code) to use AI workflows, like pulling design data into their coding tools.
Figma tested different ways to represent design data for AI, like React Tailwind (a way to describe web page designs) and plain images, finding that a mix worked best.
They learned that manually evaluating AI outputs is time-consuming and should be avoided when possible.

2026-08-28

WebMCP (a new tool from Google and Microsoft) lets websites communicate directly with AI agents (computer programs that can perform tasks), making it easier for them to browse, search, and buy.
Websites using WebMCP can offer specific actions (like search, book, or buy) to AI agents, making the process more efficient than the old way of scanning and clicking.
WebMCP could be a big opportunity for businesses, as it can help AI agents make purchases and could lead to billions of dollars changing hands.
One potential use for WebMCP is in online shops, where AI agents could help customers find products that fit their needs, like espresso machines with specific features.

2026-08-22

Uber has integrated AI agents (software that automates tasks) into their engineering workflow, leading to a 70% increase in automated code reviews (PRs) and doubling the amount of code written per engineer.
They've developed a "model gateway" to ensure sensitive data (PII) never leaves their network, with built-in safety checks and fast processing (under 100 milliseconds).
To simplify tool access, Uber created an "MCP gateway" (a central hub for software tools) that handles authentication and optimizes data usage, saving 40% fleetwide.
They've enhanced their cloud-based coding environments (devpods) to better support AI agents, with pre-configured pods for quick, isolated, and global access.

2026-08-04

MCP apps (a way for apps to send their own user interface into chats) are now supported by major platforms like ChatGPT, VS Code, and Slack, making interactions more visual and interactive.
MCP apps allow apps to send their UI to chats, so users can interact with services like Shopify or Hugging Face directly within the chat, rather than just receiving text responses.
MCP apps standardize how apps communicate with the chat host, ensuring the host controls the flow of information and actions, like favoriting a song on Spotify.
MCP apps are gaining widespread adoption, with early supporters like 11 Labs, Shopify, and Postman, and are now recommended by OpenAI for building ChatGPT apps.

2026-08-01

AI can help create a virtual executive officer (a digital assistant for business tasks) using tools like Claude Code (a coding assistant) and frameworks like Seed (a planning tool) and Skill Smith (a skill-building tool).
To build this officer, you need to know what you want it to do, what data it can use, and how to connect it to your other software tools using MCPs (command-line tools that act as bridges).
The focus is on AI augmentation (using AI to improve decisions) rather than full automation (replacing all human tasks), especially if your business processes aren't clearly defined yet.
You can use tools like Appify (a data scraper) to gather data from platforms like Instagram and YouTube, and integrate it with your officer for tasks like competitor analysis.

2026-07-25

Learn to create unique web designs by first building a personal library of design inspiration from sites like Dribble (a website for designers to share their work), Pinterest (a visual bookmarking tool), and Twitter (a social media platform), to help cultivate your own taste.
Use tools like Impeccable (an open-source design improvement tool with 23 different commands) to enhance and refine your web designs, making them more impactful and easier to understand.
Develop a flexible roadmap for AI design work, including how to prototype, iterate, and tweak your designs until you achieve a result you like.

2026-07-22

AI tools (computer programs that use artificial intelligence) often struggle with large amounts of information, as performance can degrade after using just 25% of their context window (the amount of data they can process at once).
Documentation (instruction manuals) for AI tools is often written for humans, not AI, making it difficult for AI to understand and use effectively without additional context engines (tools that help AI understand information better).
In the future, AI tools may focus on creating a "fluency mode" (a smooth, easy-to-use experience) rather than a "friction mode" (making it hard for users to switch to a different tool), with skills (specialized abilities) playing a key role in this shift.
When evaluating AI platforms, users often consider factors like security, data governance (how data is managed and protected), and integration (how well it works with other systems) capabilities.

Key points

What it is

MCP (Model Context Protocol) is a standard that lets AI agents talk to any business tool through one connection, like a bridge between AI and software.
It acts as a manual for AI, showing what actions it can take in each application, like drafting emails or creating contacts.
MCP is vendor-neutral, meaning no single company owns it, and it's donated to the Linux Foundation.
There are two complementary protocols: MCP for backend (server-side) tools and WebMCP for frontend (browser-native) tools.

How to use it

Start by understanding the three core parts: the host (program accessing data), the client (connects host to MCP servers), and the server (bridge between AI and software).
Create an MCP server using a command like `npx create MCP app` and define tools using primitives (basic building blocks).
Set up the connection in AI platforms like Claude Code, configuring MCP so the AI knows how to access tools like Gmail or Slack.

Watch out for

Beginners often create MCP servers without planning for distribution or user experience, ignoring MCP apps (frontend interfaces).
Just building an MCP server isn't enough; users need to find it, so check third-party platforms like Zapier or Compasio for integrations.
Keep your MCP server focused; too many tools can make the agent less efficient than a raw API.
Master basic coding before adding MCP complexity, and publish your MCP app while platforms like GPT and Claude are accepting them.

Tools named

npx (a command-line tool for executing Node.js packages), VS Code (a code editor), Claude Code (an AI coding assistant), GPT (a language model by OpenAI), Zapier (a tool for automating workflows), Compasio (a platform for AI integrations)

Lesson 1: What is MCP Architecture and Future and why it matters

MCP (Model Context Protocol) is a universal standard that lets AI agents talk to any tool in your business through one connection. Think of it as a bridge between your AI agent and other software you use, like a CRM. Instead of each AI needing separate API keys and custom code for every tool, MCP provides a manual the AI reads to understand exactly what buttons it can press in each application—like drafting emails or creating contacts. Anthropic donated MCP to the Linux Foundation under the Aentic AI foundation to make it vendor neutral, meaning no single company owns it.

Two complementary protocols exist. MCP is the backend protocol connecting AI to server-side tools. WebMCP is Google’s front-end protocol, a browser-native API connecting agents directly to client-side web pages. Together, they form the full stack of AI tool integration—server and browser, backend and front end.

Why does this matter? MCP is more token efficient (uses less processing cost) than raw API calls because it explains exactly how a tool works, so the AI doesn’t waste resources figuring things out. However, an MCP server can be more expensive than direct CLI (command line interface) usage because it carries all tool descriptions.

The future is domain-specific agents. MCP sits as a layer on top of tools and skills, meaning all software may eventually be used through this standard. Understanding MCP now positions you to build AI systems that reliably connect to any business tool without custom coding each connection.

Sources

Lesson 2: How to use MCP Architecture and Future: step-by-step

To use MCP (Model Context Protocol) step by step, start by understanding its three core parts: the host, the client, and the server. The host is a program like VS Code or Claude Code that wants to access data. The client connects the host to MCP servers, which are bridges (connections) between your AI agent and other software like a CRM or Gmail. The agent figures out when to use each tool and what parameters to fill in, so you do not have to think about that.

First, create your MCP server. Run a command like `npx create MCP app` to get a template. Design your server as you normally would, defining tools (actions the AI can take) using primitives (basic building blocks). For example, a button in a UI can trigger a tool call to gather data from the server. Primitives also let you check if the client supports MCP apps.

Next, set up the connection. In Claude Code, configure the MCP so your AI knows who you are and how to access tools like Gmail or Slack. The AI digests your request, determines which tools it needs, and uses the MCP bridge to call them automatically.

For distribution and discovery (finding and sharing MCP servers), there is a store for MCP apps where you can publish yours. Platforms like GPT and Claude are now accepting more apps, so it is a good time to submit your own. The goal is not to pick just one MCP server but to streamline all your tools into one agentic coding system.

Sources

Lesson 3: Best practices and pitfalls

When building with MCP (a standard that lets AI talk to your tools), beginners often jump into creating MCP servers (backends that provide tools to AI) without planning for distribution or user experience. A key future pitfall is ignoring MCP apps—frontend interfaces that sit on top of MCP servers to make them interactive and user-friendly. For example, an MCP app can add a button in VS Code that triggers a tool call to gather data, turning raw functionality into something non-developers can use.

Discovery is another common mistake. Just building an MCP server is not enough; users need to find it. Relying only on native integration inside tools like Claude is limiting. Check third-party platforms like Zapier or Compasio, which offer many ready-made MCP integrations that are harder to reach otherwise. Do not waste time wiring custom connections when a standard integration already exists.

Best practices start with understanding the primitives: MCP hosts (programs like VS Code that want to access data), clients (sometimes the same as hosts), and tools. Keep your MCP server focused—shove in too many tools and the agent struggles to look through descriptions, making it less token-efficient than a raw API. Sequence your work: master basic coding before layering on MCP complexity. The protocol has been around since 2024, so leverage existing patterns rather than reinventing the wheel. Finally, publish your MCP app now while platforms like GPT and Claude are rapidly accepting them—this is the moment to establish presence.

Sources