How to make a video with an AI agent
Give Claude, ChatGPT, Cursor, or any MCP client one URL, and it drafts a real project on a real timeline — not a black-box clip.
Most AI video tools give an agent a prompt box and return a finished file you cannot open. That is a dead end the moment the result is nearly right, which it usually is. The Model Context Protocol offers a better shape: the agent calls actual tools, builds an actual project with timed captions, paced scenes, and a coherent visual treatment, and hands back a link. The link opens the full editor with the agent's work already on the timeline.
This guide covers the whole workflow rather than any one client: connecting the server, what the agent can and cannot do, how to prompt it so the first draft is close, and where to take over. If you already know which client you are using, each one has its own setup page with the exact config, and they are linked at the bottom.
Add the MCP server to your agent
Motion Text runs a hosted MCP server over Streamable HTTP, so there is nothing to install and no key to generate. Anonymous drafting uses https://motiontext.video/mcp. Your client either takes that URL in a settings panel (Claude, ChatGPT) or in a small JSON or TOML config file (Cursor, VS Code, Windsurf, Cline, Codex); Claude Code takes a single command.
Confirm the tools are visible
Every client has somewhere that lists a connected server's tools. You are looking for create_story_draft, list_visual_presets, and get_spec on the anonymous endpoint. If the server connects but no tools appear, the agent is usually in a chat-only mode — tool calls fire in agent mode, not ask mode.
Describe the video, not the software
Give it the words and the feel: the lyrics or script with one beat per line, a mood, a length, and a format. The agent maps that onto scenes, caption timing, and a visual preset. Asking it to "set scene three to 4.2 seconds" works too, but it is rarely how you want to spend the first pass.
Open the link and take over
The agent replies with an edit link. It opens the project in the browser editor with everything adjustable: drag any caption, retime any scene, swap the treatment, add your own music and images. Anonymous drafts last 30 days; sign in and save one to keep it permanently.
Sign in for your own workspace
Swap the anonymous URL for https://motiontext.video/mcp/studio and your client walks you through a one-click Google sign-in. There is no token to copy. The agent then works on your saved projects with the full toolset — scenes, timed text, presets, Director cuts, validation — and on Pro it can queue and poll real 1080p or 4K server renders.
Export
Export the MP4 from the editor in your browser, or on Pro ask the agent for the render and it queues the job and polls it to a download link. One more click creates 9:16, 1:1, and 16:9 variants as independent, editable projects.
What an agent can and cannot make
The division is deliberate and worth knowing before you prompt, because it explains most of the surprises. Agents handle structure: timing, words, scene order, pacing, styling, format. Those are things a language model is genuinely good at and that you can verify at a glance from a link.
Agents do not handle media. Uploading your song, importing your photos, and generating AI imagery all happen in the editor, where you can see the result and own it. An agent that silently pulled in images on your behalf would be producing material you had not looked at and could not account for. So the honest shape of the workflow is: the agent builds the structure, you bring the media.
- Can: draft a full project, time captions to a script, pace scenes, apply a visual preset, set format and duration, edit and delete existing text, validate layout.
- Can, signed in on Pro: queue a real 1080p or 4K server render and poll it to a download link.
- Cannot: upload your audio or images, generate AI imagery, or export from your browser — those stay in the editor with you.
- Cannot: see the video. It builds from structure, which is exactly why the edit link matters.
Prompts that produce a usable first draft
The difference between a draft you refine and a draft you discard is almost entirely in how the words are given. Four things carry most of the weight: the text itself with deliberate line breaks, a mood, a length, and a format.
Line breaks matter more than people expect. The Director paces scenes and captions from them, so one beat per line gives it a rhythm to work with, while a wall of prose gives it nothing and it will guess. Naming a mood — cinematic, dreamy, energetic, noir — makes it pick a matching visual preset instead of defaulting. Naming a length and an aspect ratio up front avoids a rebuild later.
A prompt with everything the Director needs:
Make a 30-second portrait lyric video called "Night Drive"
from these lyrics. Dark, cinematic mood, fast pace,
karaoke word highlights:
I remember the night
The city was still awake
And nothing had a nameAgents that do not speak MCP
MCP is the convenient path, not the only one. The same drafting capability is a plain HTTP endpoint: POST to /api/drafts with the body create_story_draft takes and you get a draft link back, and GET /api/spec returns the machine-readable schema, capabilities, and limits. That is enough for a shell script, a cron job, or a framework with no MCP support.
There is also /llms.txt, a short machine-readable brief covering what the product does, what it costs, and where the endpoints are. It is generated from the same source as the pricing page, so an assistant reading it gets current numbers rather than whatever was true when someone last edited a static file.
Questions
What is an MCP server, in one sentence?
A standard way for an AI assistant to call real tools on a real service — so instead of describing a video, the agent builds one and hands you the link.
Do I need an account or an API key?
Not for drafting. The anonymous endpoint needs nothing at all. For work on your own saved projects and for server renders, use the studio endpoint and sign in with Google when your client prompts; a free account is created on the spot.
Which AI apps can connect?
Anything speaking MCP over Streamable HTTP: Claude Code, Claude desktop and web, ChatGPT, OpenAI Codex, Cursor, VS Code with Copilot, Windsurf, Cline, and most agent frameworks. Each has a setup guide with its exact config shape.
Can the agent make the whole video by itself?
It makes the whole structure by itself — scenes, timing, captions, styling, format. It cannot add your music or your photos, and it cannot see what it built. Treat the draft as a strong first cut, not a finished film.
Is my work visible to the AI company?
Your client gets a scoped access token for the tools you approve, not your password, and everything the agent touches is limited to your own account. Revoke it any time from your client's connector settings or your Motion Text account page.
What happens to drafts I do not save?
Anonymous drafts stay available for 30 days and then expire. Signing in and saving one moves it into your workspace permanently.