Five things people build with the MCP server

The shape they all share

Before the examples, the shape they have in common — because it is what makes them safe enough to leave running.

An assistant can do everything up to and including starting a production. It cannot approve the price for you, it cannot approve the storyboard for you, and it cannot publish anything anywhere. In practice that means every one of these workflows has exactly two moments where you are needed, and they are the two moments where a mistake would cost real money or real embarrassment. Everything either side of them is unattended.

Drafting, validating, quoting and polling are all free. The single charge lands when a production starts, for the figure you were shown. If you want the mechanics, the tool reference has them; this article is about the shapes.

1. The back catalogue that becomes a Shorts channel

The most common first use, and the one that pays for itself fastest. You have eighty blog posts. Nine of them still get traffic. None of them are videos.

"Look at these nine posts. Turn each into a 45-second vertical video, our usual brand, and give me the dry run first."

What the assistant does: reads the accepted design vocabulary from the platform so it briefs in terms StudioCut.Video actually understands, writes nine briefs, and runs them through the dry run. The dry run creates nothing and charges nothing — it comes back with each item normalised exactly as it would be created, plus warnings for anything it had to fix.

You read nine one-line summaries instead of filling in nine forms. Then it creates nine drafts, which is free, and quotes the first one.

The pacing is deliberate: one production at a time per key. That feels slow until you realise what it prevents — nine simultaneous starts, one of which was wrong, all charged. Instead you see the first video, decide whether the treatment works, and only then tell it to keep going.

2. The agency key that can only see one client

A key is bound at creation to either your personal account or one specific agency, and the binding cannot be changed afterwards. There is no argument that widens it and no way to switch context mid-conversation.

That single constraint is what makes assistant-driven work viable in an agency. A key issued against a client's workspace can list, read, create and spend there and nowhere else. Point an assistant at it and the blast radius of a badly-worded instruction is one client, not your whole book of business.

"Using the Northwind key: three explainer videos from this brief document, their brand kit, 16:9. Dry run first."

Practically, agencies end up with one key per client workspace, each labelled for the client, each with its own spend ceiling. When a contract ends you revoke one key. When a producer leaves you revoke theirs. Neither action touches anything else.

The one thing to design around: an assistant working a client key cannot answer "what videos do we have across all clients?" — it genuinely cannot see them. That is the feature working, not a gap.

3. Release notes that turn into a video without anyone remembering to do it

A pattern for product teams. Every release, someone writes notes. Almost nobody makes the video, because the notes are done at 6pm on a Thursday and the video is a fresh half-hour of work.

"Here are this week's release notes. Make a 90-second product update video, professional tone, existing customers. Show me the price."

What is interesting here is not the automation — it is the latency. The video gets briefed while the person who wrote the notes is still holding the context of what changed and why it matters. That is a better brief than the same person would write a week later, and it is the reason this one tends to survive as a habit when other automations get abandoned.

Two things worth knowing before you build it:

• The premium. A production started through an assistant carries a 12.5% premium, and only when the quality tier is 2 or higher and the target duration is over 30 seconds. A 90-second release video at a mid tier crosses both thresholds, so budget for it — the quote shows it as its own line before you agree to anything.
• Your plan doesn't gate it. Assistant-driven work isn't held back by plan feature tiers and doesn't consume the monthly app allowance you bought for hand-made videos. The two budgets run in parallel.

4. Language fan-out, asked for in a sentence

One approved master, several languages, each a complete separate video with its own visuals and its own voiceover in a locale-appropriate voice — not a subtitle track.

"Same video in French, German and Spanish too."

This one deserves care, because it is the fastest way to be surprised by a bill. Three extra languages is four charged videos, not one. The pricing tool prices the whole batch, so the number you are shown is the real one — but an assistant that says "adding three languages" without saying "four videos" has told you the truth badly.

The pattern that works: get the master right first, approve its storyboard, look at the finished video, then ask for the languages. Fan-out multiplies whatever you have — including a treatment you would not have approved if you had seen it rendered.

5. The overnight batch with a hard ceiling

The one that sounds most dangerous and is actually the most controlled, provided you use the brake that exists for it.

Each key can carry a credit ceiling — a number of credits per day, per week or per month. A start whose own quote would breach that ceiling is refused before any money moves, and the refusal names the cap, what has been spent so far, what this call would have cost, and when the window resets. It is not a warning after the fact; it is a door that does not open.

"Work through the content calendar. Stop when you hit the daily ceiling and tell me where you got to."

An assistant can also read what a key is allowed to spend before it designs anything, which is the difference between "you can't afford this" arriving at brief three and arriving after twenty briefs have been written.

Two more brakes sit behind it. One production at a time per key, so a batch is a queue rather than a stampede — and a storyboard waiting on your approval holds that slot, which is deliberate pressure. And separately, if your balance drops below a threshold you set, the platform raises a top-up invoice and emails you the link. No card is ever charged automatically; there is no saved payment method on this platform. A runaway integration produces an email and a halt, not a bill.

What people try that doesn't work

Worth naming, so you plan around them rather than discovering them at 11pm:

• "Publish it to YouTube when it's done." There is no publish tool, on purpose. Pushing a video to a platform is irreversible and outward-facing, so it stays in the app where a person does it deliberately.
• "Save this as our new brand preset." An assistant can brand a batch, but creating a reusable saved preset is a job for the Brand Library.
• "Approve the storyboard yourself, you know what I want." It cannot. That is the one gate, and it does not have an override — not a setting you have not found, not a flag. A person says yes once per video.
• Polling every two seconds. Each tool has its own hourly budget and a tight poll loop burns it. Twenty to thirty seconds between checks is plenty; the error tells you exactly how long to wait if you overrun.

Where to start

If you want to try this today: generate a key on the MCP and API Keys page in your account, paste it into your client's config, and ask your assistant what video styles StudioCut.Video supports. That single read-only question proves the whole path works, and costs nothing.

Then pick the smallest of the five above — one blog post, one 45-second video — and go all the way through to a download link before you automate anything. The two decision points are the whole design; you want to have felt them once before you hand an agent a content calendar.

The setup guide takes about five minutes. The tool reference is there when you want to build your own client, and there is an open-source test console that exercises the whole surface in a browser before you write a line of code.