The App Store for Agents: Why Anthropic's Model Context Protocol Is Such a Big Deal
Plus: Google Dropped a Bag Of Goodies
TL;DR: Anthropic's new Model Context Protocol (MCP) fundamentally changes how AI agents interact with the digital world, eliminating custom tool development and enabling seamless connections to virtually any data source or service. For businesses, this means dramatically faster AI deployment, more capable AI assistants, and significantly lower development costs.
Have you ever wondered why your AI assistant can write poetry but can't check your calendar? Or why it can explain quantum physics but struggles to pull a simple customer record from your CRM?
These past couple of years our AI models have gotten better and better at reasoning.
You’ve probably heard terms like “PhD-level intelligence” and “better at math than your high-school science teacher” (admittedly a low bar) being thrown around.
So why haven’t our AI agents taken over the world by now?
Well, for one, they’re stuck on their server racks.
This is partly by design — AI models make too many mistakes to be used in most real-world contexts — but perhaps even more so due to something much more mundane:
Their pet humans can’t code quick enough.
The Integration Bottleneck
Before MCP, if you wanted Claude or ChatGPT to access your company database, Slack channels, or email system, human developers would needed to:
Build custom API integrations for each service.
Create specialised code for handling authentication.
Develop bespoke parsing systems for each data type.
Maintain these integrations whenever APIs changed.
And there’s just not enough money or engineers around to do this for each of the many interfaces for each of the many agentic AI systems currently in development.
Enter Anthropic’s Model Context Protocol (MCP): a standard interface AI agents can use to talk to any system that hosts an MCP server.
Some notable early adopters of MCP outside Anthropic include Perplexity AI, Apollo.io, Cursor, and Block (formerly Square).
The App Store for AI Agents
MCP provides a standardised protocol that works like a "USB-C port for AI", as Norah Sakal puts it in her excellent write-up.
Instead of building custom integrations, developers can create a single MCP server that exposes their tool or data source to any compatible AI agent system.
It’s like giving AI agents access to an app store full of tools and databases they can use to make sense of the world around them.
So instead of cumbersome manual integrations, with MCP your AI agents get:
Plug-and-Play Integrations
Because MCP creates a standardised way for AI agents to discover available tools and data sources automatically, your AI assistant can dynamically explore what's available and use what it needs—no hard coding required.
Two-Way, Real-Time Communication
Unlike traditional APIs that follow a request-response pattern, MCP enables continuous, bidirectional communication between AI models and external systems.
With MCP integrated, AI systems will be able to do things like:
Monitor Slack channels in real-time
Track inventory changes as they happen
Follow document edits as users make them
Execute multi-step workflows across multiple systems
… and much much more.
Robust Security and Controls Out of The Box
MCP gives organisations granular control over what their AI agents can access, with built-in permission systems that mirror how you secure your existing tools. This addresses one of the biggest concerns businesses have had about AI integrations.
How MCP Works
MCP uses a client-server architecture where:
Clients are AI applications (like Claude Desktop or custom AI tools)
Servers expose functionality from tools/data sources
Communication happens via JSON-RPC, enabling rich, structured interactions
The protocol defines core primitives like:
Tools: Functions AI can call to perform actions
Resources: Data sources AI can query or manipulate
Prompts: Instructions for how AI should interact with users
Anthropic has provided open-source SDKs in Python, TypeScript, and Java to simplify implementation, along with pre-built connectors for common services like Slack, GitHub, and PostgreSQL databases.
The Future: An Ecosystem of Connected AI
The long-term implications of MCP could be revolutionary:
AI Agent App Stores: Third-party developers creating specialised AI agents that can plug into any MCP-compliant system.
Cross-Model Compatibility: Use different AI models interchangeably against the same tools and data sources.
Composable AI Workflows: Chain together multiple AI agents that each specialise in different tasks but share context through MCP
If you're interested in exploring MCP for your organisation:
Check out Anthropic's official MCP documentation.
Explore the Python and TypeScript SDK.
Start with simple integrations to prove the concept.
MCP represents a fundamental shift in how AI agents interact with the digital world. By standardising these interactions, Anthropic has eliminated one of the biggest barriers to practical AI adoption in businesses.
Text-to-Image-And-Text Updates
It looks like we’re finally closing in on multimodal models that can generate images with the correct text… but we’re not quite there yet.
The promise is of course amazing — picture that you're a marketing director facing a tight deadline for a new campaign. You need 20 product variations across different backgrounds, each matching your brand aesthetic perfectly.
Traditionally, this means days of design work and thousands in agency fees.
Now, imagine completing this entire project in under an hour—with just your AI.
Gemini 2.0's Native Image Generation Capabilities
Google's Gemini 2.0 Flash introduces native image generation and editing capabilities that fundamentally change what's possible with AI-assisted visual content.
A lot previous image generation systems consisted of image generation models bolted onto existing language models.
With Gemini 2.0, the Google team tried to build a model designed for multimodal understanding—meaning text, images, audio, and code are all processed natively within the same architecture.
What sets Gemini 2.0 from competitors is that these multimodal capabilities come with an extraordinary 1-million token context window that allows Gemini to maintain consistent character features, styling, and brand elements across multiple images.
Image Generation For Humans
And because you can talk to your images, you no longer need to know the ins-and-outs of Midjourney prompting just to get some useful images for your day-to-day.
Soon you’ll be able to do stuff like
Conversational Image Editing: Upload any image and refine it through natural conversation—"make the sky more dramatic," "change the character's hair to blonde," or "add a mountains backdrop."
Character Consistency for Entertainment and Gaming: Generate characters with uniform designs across hundreds of scenes and poses—a task that previously required extensive style guides and quality control.
Marketing Asset Generation: Quickly create on-brand visuals at scale without starting from scratch each time. A product can be visualised in different environments, with different demographic groups, or in seasonal settings—all while maintaining perfect brand consistency.
Intelligent OCR and Reproduction: Gemini can analyse existing images, extract text, and regenerate visuals with specified modifications—ideal for updating legacy marketing materials or transforming competitors' concepts into your brand's style.
Getting Started with Gemini 2.0's Image Generation
Ready to explore how this technology could transform your visual content workflow?
You can try out their latest gemini-2.0-flash-exp model for free in Google AI Studio.
I’ve played around with it a bit, and the promise is there but there’s still a lot of room for improvement:
This was generated from a single prompt that read:
generate four images that together illustrate the origin story of AI in four panels:
Silicon in the earths crust and how it is formed
Edison figures out how to tame electricity
Transistors enable programmabel computers
Deep learning leverages massive amounts of data to learn patterns
the images in the origin story are background for the newsletter article below showcasing your capabilities, so give it your best shot! :)
Check out the new Gemini 2.0 yourself and let me know what you think (or generate :))!
This week in AI
Manus AI from China made waves this week by opening up its multitasking agents to the public on invite-only basis. Their agents are capable of handling complex research tasks autonomously at scales and speeds that make the Deep Research capabilities of ChatGPT, Perplexity and Gemini look like child’s play.
Google released Gemma 3, a family of open-source multimodal AI models that support over 140 languages and an expanded 128k token context window. There are four model sizes: 1B, 4B, 12B, and 27B parameters. Because these models are open source, you can download and run them locally without sending data back to Google servers.
OpenAI unveiled a set of new tools designed to simplify the development of AI agents. The release includes the Responses API (which combines the simplicity of Chat Completions with tool-use capabilities), built-in tools for web search, file search, and computer use, an Agents SDK for orchestrating workflows. They’ve also integrated new observability tools for debugging.