Hey folks,

we are creating a series of short FAQ like explainers on topics that seem too technical but keep popping up everywhere. We’re starting with MCP (read the post here). What other topics would you like me to explain?

Let’s get into it

 🔎 What’s Trending

In just the first day of Cloud Next, Google has bamboozled my mind with the number of features launched. Haven’t got access or the time to try them all, but here’s the dispatch:

First up, Google Workspace - Docs got audio overviews, Sheets got AI analysis, and there’s a new automation creator—Workspace Flows. Flows connect different Google apps like Forms, Sheets, Chat, along with Gemini & Gems. Yep, you can add an AI step between your flows. (more workspace launches)

Gemini API has two new entries: 1. Live API for real-time talking applications—audio in, audio out. 2. Veo 2 for video generation is now available to all paying developers. Also, Google AI Studio has a new look. Gemini 2.5 Flash was teased but we have zero info about it.

Let’s talk Agents now. Agentspace, the enterprise platform for building agents, now has integrations like search, external data and more. It also has pre-built agents like Deep Research and Idea Generation agents. Looks like after some initial setup, you can configure these agents with simple prompts.

There’s new developer tooling too: Agent Development Kit (ADK), another agent framework, and A2A, a new protocol for Agent-to-Agent communication. You can also deploy/run your agents built with any framework or tooling on Agent Engine without hunting for other tools.

But here’s the most important news for us: Google now has its own text-to-app tool. Firebase Studio helps you make an app like Bolt or Lovable can, but it feels richer in features. It gives you a preview of what it’ll build, allows you to easily edit code manually, select elements on screen and easy deployment. I need to play more with it, but it looks really insane (and fast).

Claude’s got the OpenAI bug. You can now get Claude MAX by paying $100 (5x higher limits) or $200 (20x higher limits) to Anthropic for letting you use Claude more. No extra features for now, that’s it.

Midjourney is not dead yet—after a year, they have released Midjourney V7. It’s a smarter model and creates beautiful images (that 4o fails at). This launch also comes with a new draft mode for lower cost and faster generations.

Building AI apps but stuck tweaking prompts? Move beyond guesswork with AI Evals For Engineers by Hamel & Shreya (Google, GitHub). Learn to measure what matters and create systems that improve over time. Ben's Bites readers get an exclusive $250 off — first 25 only!*
*sponsored

 💬 Dear Ben… what is MCP?

LLMs are getting more capable, but are still siloed by default. Each integration is bespoke, every model has its own quirks, and context still resets between apps.

Retrieval Augmented Generation (RAG) has been a popular attempt to solve this problem. While RAG works, there’s a new term creating buzz – MCP.

Model Context Protocol (MCP) is a proposed standard that aims to make connecting AI applications to external data sources easier.

I have been struggling to understand how MCP stands out amongst dozens of existing solutions and the reason why we need a common standard. Every post online is about “how MCP works”, but I want to understand why we should care about it.

So we did some research and spoke to our community to identify some key questions that we answer.

Q1: What is MCP?

Q2: How is it different from “tool calling”?

Q3: Does a normal user need to learn MCP?

Q4: Is it like HTTPS? Why not?

Q5: How will companies adopt MCP?

Q6: What’s the best way to look at MCP—now and in future.

The questions in the post are more nuanced and answer my doubts around MCP. Hopefully, it helps you get a better look at MCP, keeping all the hype aside.

Read the full post for free →

⚙️ Top tools

AssemblyAI: The most accurate Speech AI for conversation intelligence.*

Airtable Assistant - They peg it as a app builder, data analyst, and web researcher all within Airtable.

Sculptor - Launch multiple agents in parallel to fix your code, synced with your IDE.

WordPress has a new AI website builder. I did a mini benchmark for it and it’s…..pretty bad.

Voicenotes Pages - Create a public voice notes page to share your stories and ideas. Just press record and tap publish. This is from one of my favourite entrepreneurs who has the best taste. highly recommended.

ProductRank - Understand how the top AI models promote products and brands.

More tools →
*sponsored

 🌐 News flash

Gemini Deep Research is now powered by 2.5 Pro. This is the best thinking model you can use right now and I find its reports similar to/better than ChatGPT Deep Research.

Deep Cogito released 5 models (from 3B to 70B) with open licenses and better performance than other models in their category. These models are improved versions of base Llama/Qwen models by Deep Cogito’s IDA method.

Notion has open-sourced its MCP server that roughly follows its official API. It can read and write to existing pages in your workspace, plus create any new pages.

OpenAI is launching a new Pioneers Program for companies building in high-impact industries to create new evals and specialized models.

Grok 3 API is now out. It’s priced exactly the same as Claude’s API.

 📜 You should read

12-factor agents: Patterns of great LLM applications.

The 2025 edition of Stanford’s annual AI index report.

Anthropic loves peeking into our chats (don’t worry, they are anonymised). This time, it’s college students. So, what are university students using Claude for?

That’s it for today. Feel free to hit reply and share your thoughts. 👋