You can ask Claude for a video and actually get one back. Claude doesn't render video itself. What it can do is call a video generation service through a connector built on Model Context Protocol (MCP): it writes the prompt, picks the settings, starts the render and hands you the file. Connecting one takes a few minutes. The harder part is that every render costs money, and the thing deciding when to call it is a language model reading your chat.
This guide covers how it works, the three things a connector should do before it spends anything, setup in Claude and ChatGPT, and where ClipSpeed stands: one connector, https://api.clipspeed.ai/mcp, for both generation and clipping. Model prices are published API rates as of October 2026.
The language model never draws a frame. It picks a tool the connector offers, fills in the inputs and waits while another service renders. If MCP is new to you, our plain-English MCP explainer covers the basics. For video generation, the loop usually runs like this:
Step 2 is where the assistant helps most, expanding a one-line idea into camera movement, lighting and sound cues. It's also where mistakes start: it picks a wrong duration or aspect ratio as confidently as the right one.
A note for ChatGPT users: OpenAI's own video model, Sora 2, has been discontinued. The Sora app went offline on April 26, 2026, and the API followed on September 24, 2026 (OpenAI Help Center, OpenAI API deprecations). From inside ChatGPT, video now means a connector to another service.
Video models are priced by the second, so costs build quickly. On ClipSpeed, a 10-second Seedance 2.5 clip is 37 credits at 480p and 83 at 720p, and an 8-second Veo 3.1 clip with audio is 114, before any retries. Every generation shows its exact credits before it runs. Our AI video cost breakdown covers the other major models.
Put an assistant in charge of calling that service and three things can go wrong:
MCP clients can ask you to confirm before a tool runs, which helps. But it's easy to click "always allow" by the third prompt, and a confirmation inside the chat is still part of the chat. If the assistant can type "the user approved" and the tool believes it, your safeguard is only as reliable as the model. Our MCP security notes go deeper on this.
Start with the connector that's live
ClipSpeedAI's clipping connector works in Claude today. Give it a video or live-stream link and get back vertical clips with burned-in captions and a viral score on each one.
Try ClipSpeedAI →Before you connect anything that can spend money, check for three things. Our roundup of AI image and video generation MCP servers compares the main ones on cost before you run and spend caps.
The first call should return a price, not a render: a maximum price for the exact settings (model, mode, duration, resolution, audio), with no provider work started and no money held. That's your chance to catch the 15-seconds-instead-of-5 mistake before it costs anything. If a connector's only tool is "generate", you see the price for the first time on your bill.
This is the check that matters most. The assistant should show you the price and wait for your yes before every paid run, and the service should also enforce spending limits that live in your own account, on a page where you're signed in: a maximum per generation and a cap per day. Nothing typed in the chat can raise them. Then it doesn't matter whether the assistant misread you, got stuck in a loop or picked up an injected instruction: the worst case is bounded by limits you chose.
You want the video file itself, not a preview that only plays in someone else's viewer. Download links are often time-limited, so save the file to your own storage as soon as it's ready. Inputs matter too: a connector that only accepts references you've uploaded to your own account is safer than one that fetches any URL the assistant pastes in.
The full checklist we'd run on any connector:
| Check | Why it matters | What to ask |
|---|---|---|
| Quote first | Catches a wrong duration or resolution before money moves | Does the first call return a price without starting a render? |
| Limits outside the chat | A loop or an injected instruction can't spend past what you allowed | Where are spending limits set, and can the AI change them? |
| Safe retries | A dropped connection shouldn't mean a second charge | Does the generate call take a request key so a repeat is recognized? |
| Refunds on failure | Renders do fail, and you shouldn't pay for those | Are failed generations refunded, and in full? |
| Revocable sign-in | You can cut off access without changing your password | Is it OAuth or an API key you can revoke? |
| Downloadable output | The video ends up in your storage, not only theirs | Do you get the video file, and how long is the link valid? |
A connector is a server URL plus a sign-in. Apps differ in where you paste the URL, not in how the connector behaves.
On claude.ai, go to Customize > Connectors, click Add custom connector, paste the server URL and approve the sign-in prompt. Per Anthropic's docs, custom connectors work on Free, Pro, Max, Team and Enterprise plans. The Free plan allows one, and on Team and Enterprise an Owner adds the connector under Organization settings > Connectors before members connect. In Claude Code, one command adds a connector with an API key:
claude mcp add --transport http clipspeed https://api.clipspeed.ai/mcp --header "Authorization: Bearer <API_KEY>"The same URL carries generation and clipping. Claude is the client we've verified end to end, and our ClipSpeedAI MCP guide walks through it step by step. Setup for Claude Desktop, Cursor, Windsurf and other clients is on our connect page.
Whether your ChatGPT account can add a custom MCP connector by URL is controlled by OpenAI, and the option isn't present in every account. What you need is a setting that accepts a server URL; picking from a catalog of approved apps isn't the same thing. On our side, the connector uses the same standard MCP OAuth sign-in for ChatGPT as for Claude, but we haven't verified the ChatGPT flow end to end, so we list it as rolling out.
The guardrails above don't depend on the client: if the price check and the limits live on the service's side, they work whichever assistant is asking.
Here's the flow in ClipSpeed for a product b-roll shot.
estimate_cost returns the exact credits. It's free, and no provider work starts.generate_video checks the per-generation maximum and the daily cap you set in your ClipSpeed account, and refuses anything over them. Retries with the same request can't start a second render.wait_for_generation reports status and, once the render is done, a download link. Save the file straight away.If a generation fails, the credits are refunded in full. Image-to-video and reference-to-video follow the same steps, except that references must be files you've uploaded to your ClipSpeed account, not arbitrary URLs.
The ClipSpeed MCP brings generation from ClipSpeed's AI Creator workspace into the chat, on the same URL as clipping. Every generation shows its credits first, is paid from your ClipSpeed credits, and runs only after you say yes, within the per-generation maximum and daily cap you set in Settings. Its list_models tool shows which models your account can use from chat.
AI Creator itself puts eight models side by side in one workspace: Seedance 2.5, Seedance 1 Pro Fast, Kling 3.0, MiniMax H3 and Veo 3.1 for video, plus Nano Banana Pro, GPT Image 2 and Grok Imagine for images.
Pick the model that fits the job. Seedance 2.5 suits ads and multi-shot stories and works from reference images; ByteDance's own API runs it for up to 30 seconds in one pass (ByteDance Seed), and hosts can set shorter limits. Veo 3.1 makes native 9:16 for vertical hero shots (Google). Kling 3.0 suits talking-head clips with optional audio, Seedance 1 Pro Fast is quick and low-credit for variants, and Nano Banana Pro makes thumbnails and start frames with legible text. Our Seedance 2.5 guide goes deeper.
| Clipping | Creation | |
|---|---|---|
| Server | https://api.clipspeed.ai/mcp (one URL for both) | |
| Status | Live, verified in Claude | On paid plans |
| What it does | Finds the best moments in your videos and live streams, cuts vertical or landscape clips, burns in captions, scores each clip | Brings AI Creator generation into the chat |
| Spending | Uses your ClipSpeed clipping plan | Credits shown first, runs after your yes, within your per-generation and daily limits |
| Example skills | /podcast-to-shorts, /livestream-to-shorts, /post-ready-pack, /caption-style-match | /ugc-product-review, /product-broll, /hook-variants, /faceless-hand-drawn |
Skills are ready-made workflows; the full list is on the MCP page.
Assistants write good video prompts, but they fill gaps with guesses. A few habits keep those guesses off your bill:
Generation is for shots you can't film: a product on a set you don't have, a three-second hook, an illustrated explainer for a faceless channel. Clipping is for footage you already have. If you record podcasts, streams or tutorials, the clipping connector turns them into captioned vertical shorts. The two fit together: clip the long recording, then generate the opener or b-roll the footage is missing. Our generation vs clipping comparison goes further.
Video from Claude or ChatGPT comes through an MCP connector: the assistant writes the prompt and runs the steps, and a separate service renders. Setup is easy; spending is what needs care. Judge any connector on three things: a price before anything is spent, limits the AI can't change, and a file you download and keep.
The ClipSpeed MCP is built that way, for AI Creator's eight models and for clipping, on one URL. Our MCP page has 32 presets to start from, each with a real example and the exact credits.
Claude doesn't render video. It writes prompts and calls tools, so the video comes from a service connected through MCP. OpenAI has discontinued Sora 2: the app went offline on April 26, 2026 and the API on September 24, 2026. If you want video from inside ChatGPT, a connector to another service is the route.
It's a server that tells your assistant which video tools it can call, such as getting a price, starting a generation, checking status and fetching the file. You add it once by URL, sign in, and then ask for video in plain language. Our MCP explainer covers how the protocol works.
Use a connector that shows the price before every generation, waits for your yes, and enforces spending limits you set in your own account, where no chat can change them. Don't set a client to "always allow" a tool that spends money. The ClipSpeed MCP shows the credits first, runs only after you say yes in the chat, and checks the per-generation maximum and daily cap you set in Settings on every request.
Our connector uses the same standard MCP OAuth sign-in for ChatGPT as for Claude, but we haven't verified the ChatGPT flow end to end, so we list it as rolling out. Whether your ChatGPT account can add a custom connector by URL is controlled by OpenAI, and the option isn't in every account. Claude is the client we've verified with the clipping connector.
Yes, on any paid plan. The ClipSpeed MCP (https://api.clipspeed.ai/mcp) handles clipping and creation on one URL. Every generation shows its credits first and runs only after you say yes in the chat, within the per-generation maximum and daily cap you set. AI Creator on the web has eight models: Seedance 2.5, Seedance 1 Pro Fast, Kling 3.0, MiniMax H3 and Veo 3.1 for video, plus Nano Banana Pro, GPT Image 2 and Grok Imagine for images.
That depends on the terms of the service and the model provider. In the US, the Copyright Office says purely AI-generated material isn't copyrightable, while human-authored elements and arrangement can be. Our commercial use guide covers the rest. This is not legal advice; check current terms before commercial use.
Published by ClipSpeedAI · AI video generation and AI clipping in one place — create with Seedance, Veo, Kling and Nano Banana, then cut it into captioned shorts.
AI video, AI images, AI influencers and AI clipping on one credits wallet, in the web app and inside Claude and ChatGPT.
Start creating