Let Claude or ChatGPT make your videos

Most people who use an AI assistant every day have had the same thought about a video tool: could you just do this for me? Now it can. Hiroska is an MCP server as well as an app, which means the assistant you already talk to, whether that is Claude, ChatGPT or the coding agent in your terminal, can work in your studio as you: open your projects, write a film, generate the shots, build the timeline and export it.

This post says what that looks like in practice, how to connect in two minutes, and what keeps it from spending your tokens while you are at lunch.

What "connected" means

MCP (Model Context Protocol) is the open standard AI assistants use to reach other apps, the way they reach a calendar or a drive. Hiroska publishes its tools over it: list and read projects, search the library, talk to the Director, quote and start a film run, lay out a timeline from a shot list, generate images, video and audio, import a file from a public address, look at an asset, and export. Every tool goes through exactly the same code the buttons do, so the Director writes the same film, the wallet charges the same tokens, and a project the assistant makes is a project you can open and edit by hand.

Nothing is special about any one assistant. claude.ai on the web and on mobile, Claude Desktop and ChatGPT sign in through a connector; Claude Code, Cursor, Windsurf and anything with a config file take a key.

What a request looks like

You: "Make a 20-second spot for the coast path in October. Four shots, a calm narrator, music under it."

The assistant creates a project, asks the Director to write the moments, and quotes the run: four shots, one narrator, music, about 180 tokens, go ahead? You say yes. It starts the run, checks on it every so often, and when the shots are placed it exports and gives you the download link. The project stays in your account, so the next message can be "slower narrator" or "make a vertical cut for Reels" rather than a restart.

Or the other direction, where you have the footage: "Find the three harbour clips from last week, put the wide one first, add the voiceover I recorded on Tuesday under it, and export." It searches the library by name, looks at each clip before placing it (it gets a small still, and will tell you what it sees), writes the plan, checks it against the library without changing anything, applies it, and exports.

Connecting in two minutes

If your assistant signs in (claude.ai, Claude Desktop, ChatGPT): in its settings, add a connector with the address hiroska.com/mcp and no credentials. A browser tab opens on Hiroska, a screen names the app and lists what it may do, you press Allow access, and you are back in the chat. The app then shows under Settings → Connected AI → Connected apps, next to a Disconnect button.

If your assistant takes a key (Claude Code, Cursor, Windsurf): Settings → Connected AI → New API key. Name it, choose Full access or Read only, pick an expiry, and the key is shown once with the exact line for your client. For Claude Code it is one command:

claude mcp add --transport http hiroska https://hiroska.com/mcp --header "Authorization: Bearer <your key>"

The guide walks through both, with the spending limit: Connect Claude, ChatGPT or Cursor to Hiroska.

What stops it spending everything

An assistant in a loop spends faster than a person notices, so the walls are built into the wallet rather than left to good manners.

• A quote before every spend. The assistant is told to quote and tell you before anything that costs tokens, and starting a film run without an explicit yes only returns the quote.
• A daily limit. All your connected assistants together may spend 1,000 tokens per 24 hours by default. Change it or remove it at the top of Connected AI, which also shows the last 24 hours and each key's 30-day total. The limit is enforced where the tokens move, so a film run that would cross it is refused before it starts rather than left half-made.
• Read only. A key can be made read only: it can look at projects, the library and prices, and cannot create, change or spend. Right for an assistant that answers questions about your work.
• Your workspace only. It works in the workspace you have active and cannot see another one, your password, your billing or your team.
• Off in one click. Disconnect the app or revoke the key and its next request fails.

Reading, planning, checking a plan and exporting are free. Generation costs what it costs you in the app, on every plan including Free.

What it cannot do yet

It cannot upload a file from your computer, because MCP has no file transfer. Footage on your disk goes in through the app, and the assistant then finds it by name; it can also import from a public web address. It cannot watch a finished cut for you either: it gets a still when it asks to look at an asset, not the video, so the export is yours to judge. And if you are editing the same project in a tab at the same moment, the app asks you to reload rather than let either side overwrite the other.

One thing worth knowing: what the assistant reads to do the job (project names, scripts, that small still) goes to the assistant's provider the way your own messages do. Connect the assistants you already trust with your work.

Why this way round

We could have built a chatbot into the app. We built the other half instead: the assistant you already have, with its memory of your week and the way you phrase things, reaching into your studio. The studio gains nothing by pretending to be an assistant, and an assistant gains a great deal by being able to make the video rather than describe one.