On the podcast, [Tom] and I were talking about the new generation of smartphones which are, at least in terms of RAM and CPU speed, on par with a decent laptop computer. If so, why not just add on a screen, keyboard, and mouse and use it as your daily driver? That was the question posed by [ETA Prime] in a video essay and attempt to do so.

Our consensus was that it’s the Android operating system holding it back. Some of the applications you might want to run just aren’t there, and on the open side of the world, even more are missing. Is the platform usable if you can’t get the software you need to get your work done?

But that’s just the computer-as-a-tool side of the equation. The other thing a computer is, at least to many of our kind of folk, is a playground. It’s a machine for experimenting with, and for having fun just messing around. Android has become way too polished to have fun, and recent changes on the Google side of things actively prevent you from installing arbitrary software. The hardware is similarly too slimmed-down to allow for experimentation.

Looking back, these have been the same stumbling blocks for the last decade. In 2018, I was wondering aloud why we as a community don’t hack on cell phones, and the answer then was the same as it is now – the software is not friendly to our kind. You can write phone apps, and I have tried to do so, but it’s just not fun.

The polar opposites of the smartphone-as-computer are no strangers in our community. I’m thinking of the Linux single-board computers, or even something like a Steam Deck, all of which are significantly less powerful spec-wise than a flagship cell phone, but which are in many ways much more suitable for hacking. Why? Because they make it easy to do the things that we like to do. They’re designed to be fun computers, and so we use them.

So for me, a smartphone isn’t a computer, but oddly enough it’s not because of the hardware. It’s because what I want out of a computer is more than Turing completeness. What I want is the fun and the freedom of computering.

Good morning! Let’s play Connections, the NYT’s clever word game that challenges you to group answers in various categories. It can be tough, so read on if you need Connections hints.

What should you do once you’ve finished? Why, play some more word games of course. I’ve also got daily Strands hints and answers and Quordle hints and answers articles if you need help for those too, while Marc’s Wordle today page covers the original viral word game.

We have all heard about model context protocol (MCP) in the context of artificial intelligence. In this article, we will dive into what MCP is and why it is becoming more important by the day. When APIs are already available then why do we need MCP? Although we have seen a large rise in popularity of MCP, is there staying power in this new protocol? In the first section, we will look at the parallels between APIs and MCP and then start to explore what sets it apart.

From APIs to model context protocol

A single isolated computer is limited in the amount of data that it can access and that has a direct impact on its usability. APIs were created to enable data transfer between systems. Just like APIs, Model Context Protocol (MCP) is the protocol for communication between AI agents that are using large language models (LLMs). APIs are primarily written for developers while MCP servers are created for AI agents (Johnson, 2025). 

What is MCP?

MCP was introduced by Anthropic on November 25, 2024 as an open source standard to enable communication between AI assistants and external data sources. AI agents are constrained by the fragmentation of data in isolated systems (Anthropic, 2024). The protocol defines how agents can interact with external systems, elicit user input and enable automated agents.

At its core MCP utilizes the client server model and there are three main features for clients and servers.

• MCP servers: tools, resources, and prompts
• MCP clients: elicitation, roots, and sampling

To keep this article concise, focus will be kept on the most important feature of both client and server. For MCP servers, tools are the primary way to perform complex tasks and clients utilize elicitation to enable a two way communication between the agent and the user.

Instead of explicitly calling APIs, agents select and use the appropriate tools (functions) based on the input they receive from the user. If a tool requires certain parameters the agent will use elicitation to get the data from the user. This allows for a more responsive workflow where two way communication between LLM and the user is possible.

Why do we need MCP now?

A very valid question to ask is if APIs are already present then why is there a need for MCP? APIs are designed to connect fragmented data systems and SaaS applications already enable a two way communication with a user. So, why do we need MCP now?

The main need for MCP is that the user of external data has changed from developers to AI agents. A developer will usually program an application using APIs that behaves in a deterministic fashion. Whereas, AI agents will use the user prompt and make autonomous decisions to execute on the user request. By nature, the execution of a workflow by an AI agent is not deterministic.

APIs are a machine-executable contract which acts in a deterministic fashion. APIs work if the users of APIs know what action needs to be taken next (Posta, 2025). AI agents run on top of probabilistic LLMs which do not consistently deliver repeatable results across all tasks (Atil, 2024). Variance in a LLM’s response is expected and this poses a problem for autonomous execution.

MCP to the rescue

MCP solves the problem of variance in agent execution by providing high level abstraction that wraps functionality rather than API endpoints. Tools enable LLM models to perform actions like searching for a flight, booking a calendar and more (Understanding MCP Servers, 2026). 

One common misconception for tools is that they are just an abstraction over existing API calls. Tools are not designed to be an abstraction over API calls but rather abstraction over functionality. If a lot of APIs are just exposed as tools it will increase the cost and context size for the agent which is not ideal (Johnson, 2025).

A tool may include multiple API calls in its implementation to achieve the desired outcome. An agent will review the list of available tools to automatically select the most appropriate tools and determine the appropriate order of execution.

MCP adoption boom

Since its release in 2024 MCP has seen a steady rise in popularity. The following chart from Google Trends showcases the relative interest in MCP since its launch.

A lot of companies have launched their own MCP servers to facilitate building autonomous agents. As of February 2026, the official MCP registry has over 6400 MCP servers already registered. This number of MCP servers is only expected to grow in the near future. The official registry for MCP servers is still in preview and the ecosystem has grown massively in less than a year.

Other major players in the market have adopted MCP and added support to their clients. OpenAI added MCP support to ChatGPT in March and Google added support a few weeks later in April 2025. This showcases the staying power of the protocol and the fast pace of adoption.

What lies ahead?

MCP is still in the early stages of widespread adoption where a lot of applications need to mature and start hitting production. Leonardo Pineryo from Pento AI summarized it the best “MCP’s first year transformed how AI systems connect to the world. Its second year will transform what they can accomplish” (2025). 

Guardrails around tools is an area that will see further development as trust is one of the biggest concerns with AI agents. With better guardrails in the tools, an AI agent can be allowed to perform with more autonomy. Over the next year, MCP is certain to see continued growth, both in the sophistication of its capabilities and the volume of its application.

Strands is the NYT’s latest word game after the likes of Wordle, Spelling Bee and Connections – and it’s great fun. It can be difficult, though, so read on for my Strands hints.

Want more word-based fun? Then check out my NYT Connections today and Quordle today pages for hints and answers for those games, and Marc’s Wordle today page for the original viral word game.

The feature, announced at SXSW by co-CEO Gustav Söderström, lets Premium listeners see and shape the data model powering their recommendations, starting with a beta rollout in New Zealand

For a decade, Spotify’s recommendation engine has worked largely in silence. It watched what you played, noted what you skipped, inferred meaning from the time of day and the tempo of your commute, and it never told you what it had concluded. On Friday, at SXSW in Austin, the company decided to change that.

Gustav Söderström, Spotify’s co-CEO, announced Taste Profile: a new feature that surfaces the algorithmic model the platform has been building about each listener, and crucially lets users modify it directly. The beta will begin rolling out to Premium subscribers in New Zealand in the coming weeks.

The premise is straightforward enough. Taste Profile aggregates a listener’s behaviour across music, podcasts, and audiobooks into a single view: the genres explored recently, the artists listened to most, the patterns that define a listening day.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

Where a user notices the profile is wrong, too heavy on music they played years ago, or missing a phase they have been quietly working through, they can flag it. They can ask for more of a particular vibe, or less. They can describe a current context, training for an event, commuting on weekdays, and the system will factor that in when deciding what to surface on the Spotify homepage.

“This is the next step in our vision to make personalization more transparent, responsive, and truly yours,” Söderström told the SXSW audience.

Spotify cited an internal figure that more than 80% of its listeners name personalisation as what they value most about the service. The claim, which the company has referenced in various forms since at least 2023, positions algorithmic curation not just as a feature but as the primary reason people stay.

The competitive logic behind Taste Profile follows directly from that: if personalisation is the product, giving users more control over it is a way to deepen their investment in it.

The announcement comes roughly two months after Spotify expanded Prompted Playlist, a separate but related feature that lets users generate playlists by describing what they want in natural language, from its initial New Zealand testing to Premium users in the US and Canada in late January 2026, and subsequently to subscribers in Australia, Ireland, Sweden, and the UK in February. The sequencing is deliberate.

Both features push the same underlying argument: that the future of streaming personalisation is collaborative, not passive.

Where Prompted Playlist is generative,  it creates something new from a description, Taste Profile is corrective. It works with the model that already exists, giving users a chance to audit and adjust what years of listening have written about them.

Whether someone has been an accidental customer of the algorithm (playing whatever appeared on the homepage, not particularly caring) or has strong views about the direction their recommendations have taken, the feature is designed to accommodate both. “You can shape your Taste Profile as much as you’d like,” the company said in its announcement, “or leave it and enjoy Spotify as usual.”

The beta will start in New Zealand, a market Spotify has used repeatedly for early-stage testing of AI-adjacent features, including the initial Prompted Playlist launch. No timeline was given for a broader global rollout. Taste Profile will be available to Premium subscribers only; there was no indication of when, or whether, it might reach free-tier accounts.

Spotify is marking 2026 as its 20th anniversary year, and its SXSW presence this week has been calibrated accordingly, concerts, a headline session with Söderström, country artist Lainey Wilson, and podcast host David Friedberg.

The Taste Profile announcement landed on the last day of the company’s main SXSW programming, providing a product note to accompany the celebration.

What the feature represents, beyond its functionality, is a shift in how Spotify frames its relationship with listeners. The algorithm has always existed; the company is now making the case that knowing it is there, and having some say in what it does, is a feature in itself.

For years, third-party cybersecurity relationships between vendors and customers have relied on contracts and trust. That model is now showing its age. In the past year alone, 51% of UK organizations have reported a third party-related breach, while vendors have become ideal attack vectors for hostile actors.

At the same time, cyber attacks grow in their frequency and damage, even when they rely on relatively unsophisticated techniques. It is a clear shift: the way vendors reassure their customers of the strength of their security can no longer rely on words alone.

Trust based compliance to evidence-based security