How to generate AI video in Claude and ChatGPT with MCP

MCP (the Model Context Protocol) is an open standard that lets an AI assistant use outside tools. A service publishes an MCP server that describes its tools and their inputs; an app such as Claude or ChatGPT connects to it, and from then on the assistant can call those tools when your request needs them. Claude and ChatGPT do not render video themselves, so a video model has to be connected as a tool.

The KreatorFlow MCP server is that tool. KreatorFlow is an official partner of BytePlus, ByteDance's enterprise AI platform, and video runs on ByteDance's Seedance models while images run on Seedream. When you ask for a clip, the assistant turns your words into a prompt and settings, calls generate_video on your own KreatorFlow account, and gives you the finished link. You keep chatting to adjust it, the same way you would brief an editor.

Add it as a custom connector with the server URL https://kreatorflow.ai/api/mcp, then sign in. This works on Claude on the web and in Claude Desktop. Our steps are in the list below.

Claude moves its menus from time to time. As of 17 October 2026, Anthropic's help center puts them under Customize, then Connectors, then + Add and Add custom connector. It lists custom connectors on Claude's Free, Pro, Max, Team and Enterprise plans (Free is limited to one), and on Team and Enterprise an owner adds the connector before each person connects to it.

If Claude shows authentication options, keep sign-in on: KreatorFlow uses OAuth and needs no request headers or API key.

Run one command in your terminal, then sign in from inside Claude Code. Handy when you want clips for something you are building without leaving the terminal.

Run the command below in your terminal.

Start Claude Code, type /mcp, pick kreatorflow and sign in when the browser opens.

Check the connection with claude mcp list.

The --scope user flag makes the server available in every project. Claude Code's MCP docs (checked 17 October 2026) describe the same browser sign-in from /mcp, and claude mcp list shows each server's status.

Add KreatorFlow as a custom connector with the server URL https://kreatorflow.ai/api/mcp and OAuth sign-in. Our steps are in the list below.

OpenAI renames this area from time to time (connectors became apps, next to plugins). As of 17 October 2026, OpenAI's developer docs describe adding a server from ChatGPT's Plugins page with Add custom MCP server, its public HTTPS URL and authentication. Which plans and workspaces can add custom MCP servers is set by OpenAI and your admin, so check OpenAI's help center if you don't see the option.

If ChatGPT asks you to confirm a tool call, confirm it: generating a clip is an action on your account that spends credits.

Yes. Every app connects to the same server URL and signs in with your KreatorFlow account. Setup steps for Cursor, VS Code (GitHub Copilot), Windsurf and Other MCP clients are on the KreatorFlow MCP page, with one-click install links for Cursor and VS Code.

Any client that supports remote MCP servers over Streamable HTTP with OAuth can connect directly; local-only clients can use the mcp-remote bridge: npx -y mcp-remote https://kreatorflow.ai/api/mcp.

The assistant starts a job and checks it until the clip is ready. generate_video takes your prompt plus any duration, aspect ratio, resolution or model you gave, and returns a job id with the status QUEUED or PROCESSING. The assistant calls get_job until the status is COMPLETED and reads the result URL. A FAILED job gives its credits back.

Defaults, when you don't say otherwise: Seedance 2.5, 5 seconds, 720p and 16:9, with the clip's sound and spoken voices generated (ask for a silent clip if you don't want them). In the studio, Seedance 2.5 offers 4 to 30 second clips at 480p, 720p, 1080p and 4K. Images default to Seedream 5.0 Pro at 2K; an image comes back as a link in the same reply when it finishes quickly, or as a job when it takes longer.

Every result lands in your KreatorFlow History, like anything made in the studio, so you can download it or ask the assistant to edit it later.

Illustration only, not a recorded transcript: this is the shape of a typical exchange. The assistant's wording and the timings vary.

You: "Make a 5-second vertical video of a glass perfume bottle on wet black stone, slow orbit, soft rim light, raindrops in the air."

Assistant: calls generate_video with your description as the prompt, duration 5 and aspect ratio 9:16, and tells you the clip is rendering. It then calls get_job, sees PROCESSING, and checks again later. When the status is COMPLETED it replies with the video link. At the default 720p, that clip costs 24 credits.

You: "Nice. Now a 16:9 version at 1080p." The assistant starts a second job. If your plan doesn't include 1080p, the tool returns the studio's plan message and the assistant can offer 720p.

Illustration only, not a recorded transcript.

You: "Here's a link to my clip: https://example.com/walk.mp4. Restyle it as a watercolour painting."

Assistant: calls upload_media with the link, which copies the file into your workspace and returns an uploadId, then calls edit_video with that uploadId, the edit mode restyle and your prompt, and checks the job with get_job. Something already in your library can be passed by its kreatorflow.ai link or id, with no new upload.

The same pattern drives edit_image, face_swap and kitsune: upload or pick the source, run the tool, check the job. Ask "use the moves from my last dance video on my saved character" and the assistant finds both with list_history and list_characters, then runs kitsune in motion-transfer mode.

Illustration only, not a recorded transcript.

You: "Write a 15-second voiceover for the perfume clip and read it in a warm, low voice."

Assistant: writes the script in the chat, picks a built-in voice (or one of your cloned voices from list_voices) and calls generate_speech. Audio tools answer in the same reply, so the link to the voiceover comes straight back. Text to speech costs 2 credits per 1,000 characters, charged in whole 1,000-character blocks, and one request can be up to 5,000 characters.

Describe the shot like a director. The assistant writes the prompt for generate_video, so the more concrete you are, the better it is.

The same as in the KreatorFlow studio: each generation spends your plan credits, and longer or higher-resolution video costs more. With the defaults, a 5-second 720p Seedance 2.5 clip is 24 credits and a Seedream 5.0 Pro image at 2K is 4 credits. The table above lists the rest.

KreatorFlow adds nothing for using MCP. Your Claude or ChatGPT subscription is separate and pays for the assistant itself.

Plan credits and Developer API tokens are separate balances: the connector spends credits, the developer MCP server (below) spends tokens.

The connector follows the same rules as your account in the studio. These are the ones people run into:

Remove the connector, add it again with the exact server URL, and sign in with the account that has the paid plan. That fixes most problems. The common messages:

Yes. For your own agents, scripts and backends there is the developer MCP server at https://kreatorflow.ai/v1/mcp. It takes a kf_ API key (as Authorization: Bearer kf_... or an x-api-key header) instead of a sign-in, and spends prepaid Developer API tokens instead of plan credits: one token is $0.05, and a 5-second 720p Seedance 2.0 clip is 11 tokens.

Its tools are list_models, generate_image, generate_video, edit_video, get_job and get_balance, calling the same /v1 endpoints, scopes and prices as the Seedance API. It has no audio tools.