For two years, AI could talk brilliantly but do almost nothing. Then I connected mine to my actual tools. Here’s why the Model Context Protocol might be the most important piece of infrastructure you’ve never heard of.

By Ben Anchor — Wednesday, 10 June 2026 · Listen to the podcast episode

I’ve spent a fortnight doing something that still feels faintly magical: I’ve given the assistant on my screen a proper set of hands.

This is the subject of the latest episode of Reinventing Broadcast, and it’s worth pulling the thread properly because the implications are enormous. For two years, these AI models could draft a script beautifully, summarise a meeting with precision, and answer questions with alarming fluency. But they couldn’t open your files. They couldn’t post your notes. They couldn’t do anything.

That’s changed. I’ve connected mine to my voice studio, my audio editor, my notes, my automation platform. I asked it to get on with things. And it did.

The Thing That Makes It Possible

The connective tissue underneath all of this is a quiet little standard called the Model Context Protocol — MCP for short. It was published by Anthropic in November 2024, and the idea is almost embarrassingly simple.

Before MCP, every time you wanted an AI to talk to a tool — your scheduling system, your asset store, your inbox — somebody had to hand-build a custom connection. Every pairing was bespoke, brittle, and expensive. MCP is one shared standard for all of those connections. People call it the USB-C of artificial intelligence, and that’s a fair picture.

On one side, you have the AI. On the other, you have a small program called a server that knows how to talk to one particular tool. Plug them together, and the AI can suddenly see and use that tool. Write the server once, and any AI that speaks the protocol can use it.

And it caught on. OpenAI adopted it in March 2025. Google built it into Gemini. Microsoft put it across its developer tools. Then in December 2025, the whole protocol was handed to the Linux Foundation, with those same rivals as co-sponsors — which is the moment it stopped being one company’s idea and became neutral, shared infrastructure.

By March this year, the software libraries behind it were being downloaded 97 million times a month, up from about 100,000 at launch. There’s now a sprawling open library of these servers. More than 10,000 public ones at the last industry count, and one directory alone lists over 22,000.

What This Actually Means for Broadcast

Let me be concrete, because the mapping to our industry is direct.

Think about every system a broadcaster runs. Your media asset management. Your content system. Your scheduling and playout. Your audience data warehouse, where the audience figures and the delivery logs live. Your creative and design tools — a Canva or a Figma. Your messaging. Every one of those is a candidate for a connector.

Once they all speak the same protocol and an intelligent agent can reach across the lot of them through one doorway, instead of you wiring 15 private tunnels by hand. The tools were always there. What was missing was a common way for the AI to reach them.

For this podcast, I connected an agent to 11 Labs for the voices, to Descript for the audio editing, and to N8n for the workflow automation. I can hand it a script and it reaches the voice service to produce the audio, then passes it down the chain to be processed and quality checked while I supervise, rather than press every button.

A year ago that was a software project with a budget. This fortnight it was an afternoon of connecting things that already knew how to talk to each other.

The Shift from Assistant to Agent

This is where language matters. An agent workflow is when you describe an outcome in plain English and the AI works out the steps, calls the tools it needs, checks its own results, and corrects itself when something breaks. MCP is what gives it the tools to call.

Picture one instruction the morning after you publish: Take the finished episode, transcribe it, draft the show notes and three social posts in our voice, load it into the content system, schedule the posts, and tomorrow bring me the early numbers.

Before this, that’s five tools and a person copying between them for an hour. With it, that’s one agent reaching five connectors and reporting back.

Or on the analysis side, an agent that pulls viewing figures, delivery logs, and social chatter together to explain why a show over-performed or under-performed last week — because each of those is simply another connector.

That’s the whole shift in one sentence. The work that used to be human glue is becoming the agent’s job, and the human moves up to judgement.

Where This Bites First

Four places, and none of them are exotic.

First, the archive. Decades of programmes sitting behind thin metadata. An agent that can reach your asset store and your transcription service can watch, listen, tag, and describe that content at a scale no human team could ever match — which turns a dead archive into something searchable and sellable.

Second, compliance and access services. Captions, audio description, profanity and rights checks. All jobs where an agent reaching the right tools can do a first pass and flag only what a human must sign off.

Third, the newsroom, where a research agent can pull from the wires, their own past coverage and the open web through connectors, and assemble a briefing in minutes rather than hours.

Fourth, advertising operations, reconciling schedules, playout logs and billing across systems that have never once spoken to each other.

In every case, the pattern is identical. The intelligence was already available. The connections were not. This is the thing that supplies the connections.

The Honest Rough Edges

I don’t trust frictionless, and you shouldn’t either. So here’s the honest list.

Prompt injection, where hidden instructions buried in content can trick an agent into doing something it shouldn’t. Supply chain risk, because anyone can publish a server, so an unvetted one is code you’re trusting blind. Identity and permissions, because agents can slip past the access controls we built for humans. Data exposure, because sensitive material can end up sitting in the model’s context.

Researchers catalogued around 30 serious vulnerabilities in popular servers early this year, and the security agencies have begun publishing guidance. The blunt summary is that the connecting is easy and the governing hasn’t fully caught up.

But notice the shape of it. Not one of those is a reason the technology doesn’t work. They’re reasons to adopt it with a grown-up’s discipline. And the market is already answering, with gateways that put permissions, audit trails and access control in front of every connector.

For broadcast the rule is simple. Treat a connector like any other supplier. Know what you’ve plugged in, give it only the access it needs and keep a record of what it did. Do that and the downside list becomes a checklist rather than a wall.

My Honest Verdict

Care, and start small.

This is the most important piece of AI infrastructure most people have never heard of, and it’s crossed from experiment to standard in record time. It doesn’t replace the scheduler, the producer or the editor. It hands them an assistant that can finally reach their tools and do the dull connective work — which makes their judgement matter more, not less.

The data and the craft were always the moat. This is just the bridge that lets the AI walk up to it. Build the bridge carefully, and then walk across it.

And if you’re wondering whether the expertise needed to build these is a software team — increasingly, it’s not. Because the agent assembles the steps from your plain description and the tools already exist as connectors, the person directing it needs to understand the broadcast problem, not the programming language. The expertise that matters shifts back towards domain knowledge and editorial judgement.

For an industry full of deep specialists and rather short on engineers, that’s genuinely good news, not a threat.