MCP, or how Claude plugs into anything (and what it changes for creators)

I understood what MCP really was one February evening, inside Claude Code, while asking it to pull view stats from my eight channels for comparison. I was expecting a "I cannot access that data". Instead, Claude called the list_publications tool, aggregated the numbers over a month, identified the episode that stood out, opened the script for that one, and asked me whether I wanted it to propose an editorial variant based on the pattern it had just detected.

All of it, in a single conversation. No tab switching. No copy-pasting. Not even a Python bridge script I had to write.

That's the moment I stopped thinking of MCP as a Claude feature and started seeing it for what it actually is: plumbing. A silent convention that lets any language model plug into any system — a database, a third-party API, a local file, a business tool — without anyone having to code the integration from scratch.

For anyone running an AI video production chain solo, that's exactly the kind of shift that changes the calculus.

What MCP actually is

The acronym holds three letters: Model Context Protocol. The full name is Model Context Protocol, published by Anthropic in November 2024 under an open license. The proposition is almost trivial in its formulation, and powerful in its implication: standardize how a language model talks to an external system.

Before MCP, every integration between an AI and a tool was bespoke work. Want to connect Claude to your database? Someone had to code a proprietary adapter. Want to plug it into your CRM? Another adapter. Onto your GitHub repo? Yet another. Multiply the number of models by the number of systems, and you got an N×M matrix of integrations to maintain — painful to design, unmanageable to update.

MCP does what USB did twenty-five years ago for computer peripherals. Rather than a cable per mouse brand, a cable per scanner brand, a cable per printer brand, USB imposed one convention: if your device speaks USB, it plugs into any USB-speaking computer. MCP does the same for the bridges between AI and external systems: if your tool exposes an MCP interface, any compatible model can use it, with no bespoke integration.

The operational result for someone who doesn't write the protocol but uses it: you install an MCP server (Shortflow's, for instance), you give its address to Claude Code or Cursor, and the AI instantly gains access to every operation that server exposes — no further plumbing.

What changes concretely for creators

The shift is less impressive in theory than in practice. Here are three concrete moments in my routine that weren't possible before MCP, and are now.

First moment: I finish a script for a new WhyFactory episode. I want to know whether the subject has been covered on one of my other channels, to avoid topical overlap. Before: I open eight dashboards, run eight manual searches, compile mentally. Now: I ask Claude "does any of my episodes across the eight channels already cover radioactive bananas?". Claude calls list_publications on each channel via the Shortflow MCP, cross-references the titles, replies "Cocorico Histoire EP15 briefly mentions natural radioactivity but doesn't focus on bananas. Edge case, not a duplicate." Thirty seconds instead of twelve minutes.

Second moment: an episode just went live, and a viewer comments on YouTube "do you have a source for the 'three times more radioactive than average soil' figure?". Before: I find the local script, locate the source in my Notion notes, copy the link into my YouTube reply. Now: "Claude, pull the factual source for segment 4 of the radioactive bananas episode." Claude reads the episode's JSON config via MCP, locates the segment, returns the cited source (an article from the French nuclear safety institute, IRSN), generates a polite reply that includes the link. I paste into YouTube.

Third moment: I want to understand why a given episode underperforms. "Compare the first thirty seconds of episode 04 — the one that scored -22% — against the channel median in terms of structure." Claude retrieves the scripts, measures the number of twists in the first thirty seconds (two for the median, zero for the -22% outlier), and concludes that the episode misses the reversal before the eight-second wall. Diagnosis filed without my needing to write a script parser.

None of those three moments is miraculous. Each could have been hand-coded. The difference is that none of them was. MCP turned each use case into "I tell Claude what I want, it calls the right tools", and the cumulative gap across eighty monthly episodes becomes significant.

The underlying paradigm shift

Before MCP, the AI promise for creators looked like this: "use our proprietary tool, built around our LLM, optimized for our use case". One platform per function. One UI per task. The creator became a tool operator, juggling a dozen SaaS to get a day's work done.

MCP proposes the inverse. You pick a language model (Claude, GPT, Gemini), a runtime environment (Claude Code, Cursor, Claude Desktop, custom agent), and you plug in the MCP tools you need for your value chain. Business logic is no longer encapsulated in proprietary UIs; it lives in MCP servers you compose like Lego bricks.

Concretely, the operations layer of a multi-channel creator can fit inside a single conversation. The eight channels I run talk to Claude via:

• The Shortflow MCP for multi-platform publishing (list connected accounts, upload media, trigger publication, fetch publish status, analytics).
• A filesystem MCP for reading scripts, JSON configs, renders.
• A GitHub MCP for committing new episodes into each channel's repo.
• A Notion MCP for syncing the editorial backlog.

Four servers, one conversation, zero tab switching. The difference with the pre-MCP world isn't the quality of each individual tool; it's that they now communicate in a shared dialect, and Claude can orchestrate them without human intervention.

Why now

MCP has been public since November 2024. The adoption curve was fast: by mid-2025, most developer-targeted SaaS had their MCP server. By autumn 2025, creator and marketing tools followed. As I write this (spring 2026), a solo production chain without MCP looks to me like an aesthetic choice — "I want to operate the way I did in 2023" — not a technical reality.

The network effect runs at full strength. The more a system supports MCP, the more usable it becomes from any AI agent, the more AI agents are valued, the more systems have reason to support MCP. The loop isn't threatening to stop in the next twelve months.

For creators specifically, the strategic window is open right now for one simple reason: most direct competitors haven't yet integrated MCP into their daily routine. Those who do this year realize a substantial productivity step-up. Those who wait until 2027 will find the terrain hardened.

Where to start

If you're reading this without having ever touched an MCP server, here's the reasonable itinerary.

First step — install Claude Code or Claude Desktop. Both expose an MCP configuration. The first servers you'll plug in will likely be filesystem (read your local files) and github (interact with a repo). Both are public, documented, take ten minutes to wire. The goal of this first step isn't to operate through MCP; it's to feel the shift between "I tell Claude to do X" and "Claude calls the necessary tools to do X".

Second step — identify your first repeated business operation. For me it was multi-platform publishing. For you, it might be research (sourcing news on a topic), translating a script into multiple languages, or syncing a Notion database with an editorial calendar. Check whether an MCP server already exists — the registry at modelcontextprotocol.io/registry lists the most popular ones. If yes, wire it. If not, you'll know where to point your attention when the next wave of servers ships.

Third step — not before completing the first two — study how to write a custom MCP server. It's a few-hours project for a developer, much more for a creator without technical background. But it's the only way to expose your unique business logic to any future AI agent. MCP is a long-term investment: a server you write today will tomorrow speak to GPT-5, to Gemini 4, to any model that implements the protocol — and all of them do.

The observation underneath the rest

The bet behind MCP, for creators as much as for tool vendors, is the same: the production unit of tomorrow won't be the isolated SaaS, but the AI agent that orchestrates N specialized servers. If that bet holds, then the highest-leverage move for someone producing solo content isn't learning a new tool — it's structuring their operations to be accessible to an agent.

That's exactly the reading grid that guides what Shortflow builds. The platform isn't optimized to be used inside its UI; it's optimized so an AI agent can use it on your behalf. The UI exists, it's honest, but it's no longer the primary entry point to the product in 2026.

---

Shortflow exposes a public-facing MCP server that covers video publishing on YouTube, TikTok, Instagram, Facebook, Threads and LinkedIn. Connecting from Claude Code takes two minutes; the setup guide sits in the MCP documentation. If you want to test without setup, creating an account opens a seven-day free trial with MCP enabled by default.