How to Generate AI Video from Claude (MCP Guide)
Connect 8frame to Claude in about two minutes and generate images and video without leaving the chat. Full setup, three worked prompts, and where to fall back to the canvas.
By the end of this guide you'll be typing "make me a 5-second product video of a matte-black water bottle on a wet stone" into Claude and getting a real clip back, generated by the same models you'd use on the 8frame canvas. No browser tab switching, no copy-pasting prompts between tools. Claude writes and refines the brief, 8frame runs the generation, and the result comes back into your chat with a link to the canvas.
The mechanism is the 8frame MCP server. MCP (Model Context Protocol) is the open standard that lets an AI assistant call external tools, and 8frame ships a live MCP server that exposes image generation, image editing, upscaling, and video generation across 16 models. Once it's connected, Claude can drive all of it. This guide walks the full setup, three prompts that work, and the limits worth knowing before you rely on it.
What you get, and what it costs
The output is identical in quality to generating on the canvas directly, because it is the canvas. The MCP server routes your request to the same Veo 3.1, Kling 3.0, Seedance 2.0, Nano Banana Pro, and other engines you'd pick manually. What you save is the interface friction: instead of opening the app, choosing a model, and typing a prompt, you describe the shot in plain language and Claude handles the routing.
Cost is pay-per-generation compute, the same per-clip pricing as the canvas. A 5-second Kling 3.0 clip runs $0.28-0.40; a Veo 3.1 clip at 4K runs $0.85-1.20; a Nano Banana Pro still is $0.04-0.08. You aren't paying a separate MCP fee. The only prerequisite on the Claude side is a paid plan, which we'll cover in the setup.
Step 1: Open Connectors in Claude
Go to claude.ai, open Settings, and find the Connectors section. Connectors are how claude.ai attaches custom MCP integrations. You'll see any connectors you've already added; to add 8frame, click Add custom integration.
This step requires a paid Claude plan. Custom connectors are available on Claude Pro, Team, and Enterprise, not the free tier. If you're on free, this is the one wall you'll hit, and it's an Anthropic account setting, not an 8frame one.
Step 2: Add the 8frame server
In the custom integration dialog, give it a name (8frame) and paste the server URL:
Click Connect. Claude will register the server and pull in its tool list: generate images, edit images, upscale, and generate video across the model roster. You authenticate to your 8frame workspace during this step, so generations draw from your account and land in your canvas.
Once connected, the 8frame tools show up as available to Claude in your chats. There's a full walkthrough and the server details at 8frame.co/mcp if you want the reference page open while you set up.
Step 3: Write the brief in plain language
You don't need to name a model or format the prompt yourself. Describe the shot the way you'd describe it to a director, and let Claude translate. Claude is good at expanding a loose idea into the kind of structured prompt these models want: subject, action, camera, lighting, mood. If you already know your Veo 3.1 prompt structure, you can hand Claude a fully specified prompt and it will pass it straight through.
A useful pattern is to iterate in text before you spend a single generation. Ask Claude to draft three prompt variations, read them, pick one, then tell it to generate. You've done your creative direction in a medium that costs nothing, and only the final generation touches compute.
Step 4: Generate, review, refine
Tell Claude to generate. It calls the 8frame tool, the model runs (roughly 8 seconds for a Nano Banana Pro still, 60 seconds for a Kling clip, 90 for Veo), and the result comes back with a link to the canvas. From there you can ask for a variation, an edit, or an upscale, and Claude chains the next call.
For anything client-facing, open the canvas link and review at full resolution. The chat gives you a fast preview and the generation history, but frame-level judgment (a warped label, an off hand, a lighting mismatch) is easier to catch on the canvas than in a chat thumbnail.
Three prompts that work
1. Product hero still, then motion.
"Generate a photorealistic product still: a matte-black insulated water bottle standing on wet dark stone, soft overcast light, shallow depth of field, condensation on the metal. Then animate it into a 5-second clip with a slow push-in and drifting steam."
Claude routes the still to Nano Banana Pro (fast, cheap, clean product renders) and the motion to Kling 3.0. Total: roughly $0.35-0.50 for a usable hero clip you can drop into a Shopify product page.
2. UGC-style hook shot.
"Make a 5-second vertical clip, handheld phone look, a person's hand picking up a skincare jar off a bathroom counter in morning light, natural and slightly imperfect framing. 9:16."
This is a UGC opener. Claude specifies the aspect ratio and the handheld feel; the model delivers the shot. Ask for four variations and you've got a hook-testing set for the cost of four Kling clips.
3. Edit an existing frame.
"Take the last still and change the background to a sunlit kitchen counter, keep the product identical, then upscale the result to 4K."
Here Claude uses the edit capability to swap the background while holding the subject, then the upscaler to finish. This is the everyday loop: generate, edit, upscale, all from the chat.
Cost math: a five-clip test set
Say you want to test five hook variations for a paid social ad before committing to a hero production. Five Kling 3.0 clips at $0.35 each is $1.75. Add a Nano Banana Pro still to anchor the product at $0.06, and an upscale on the winner. You're under $2.50 of compute for a full variant set, generated entirely from a Claude conversation, versus a traditional shoot that starts in the low thousands before you've booked talent.
The MCP route doesn't change that math; it changes where you do the thinking. Briefing in a chat window means you can reason about the concept and generate in the same place, which is why this workflow tends to produce more variations per session than tab-switching does.
Limits worth knowing
Fine control lives on the canvas. The chat is excellent for briefing, generating, and quick iteration. For precise composition, layering multiple reference images, or building a reusable multi-step workflow, open the canvas. The MCP server is a fast front door, not a replacement for the full surface.
Model behavior still applies. Every model's strengths and boundaries carry over. Kling is cheap and fast but tops out below Veo on complex physics; Veo is the premium option at a premium price. Claude will route sensibly, but you can always name the model in your prompt if you have a preference.
Review before you ship. Nothing auto-publishes. Generations land in your workspace for you to check. Treat the chat output as a draft until you've seen it at full size on the canvas.
Related workflows
If you want the same power inside your code editor instead of claude.ai, AI video in Cursor and Claude Code via MCP covers the developer setup. And if you're building an autonomous system rather than prompting by hand, how to add video generation to your AI agent shows how the same MCP server becomes an agent's generation tool.
Connect 8frame to Claude and generate from the chat, or open the 8frame canvas directly for the full multi-model surface. Setup details are at 8frame.co/mcp.