How to Use MiniMax H3 in Claude

MiniMax H3 is a 2K multimodal video model built for controlled generation from text, images, video, and audio. It generates native stereo sound, supports first and last frames, and can assign different reference assets to identity, environment, motion, camera language, dialogue, or pacing. CreativeClaw now makes H3 available directly inside Claude through the model ID video/minimax-h3.

What makes MiniMax H3 special?

• 2K output - The hosted H3 endpoints currently expose one high-resolution 2K output setting
• Native stereo audio - Every generation includes audio, so visible actions, dialogue, ambience, and effects can be directed together
• Multimodal references - Supply up to 9 images, 3 videos, and 3 audio files, with a maximum of 12 total references
• First and last frame control - Define where an image-to-video shot begins and where it should land
• Reference-based editing - A source video can control motion and camera rhythm, or become the video to transform while other assets preserve identity and style
• Useful duration range - Generate 5 to 15 seconds at 24 fps in widescreen, square, portrait, or cinematic aspect ratios

H3 is especially useful when a prompt alone is not enough. Instead of asking a model to guess the actor, setting, movement, and sound, you can provide each part and tell H3 exactly what role it plays.

Why use CreativeClaw for MiniMax H3?

CreativeClaw is the fastest and simplest way to use MiniMax H3 in Claude. Here's why:

• No API keys needed - No accounts, no configuration files. Connect one URL and every model is available instantly.
• No subscriptions - Pay only for what you generate. $10 = 1,000 credits. No monthly fees, credits never expire.
• MCP Apps - Preview generated media directly in Claude's UI. See results inline without opening files or navigating to external URLs.
• Expert skills built in - CreativeClaw knows how to get the best results from MiniMax H3. You don't need to be a prompt engineering expert - Claude handles the optimization.
• Let Claude iterate - This is the real power. Claude generates, evaluates the result, refines the prompt, and regenerates - all in one conversation. Your AI agent becomes your creative director.
• Run from anywhere - CreativeClaw is a remote MCP server. Use it from Claude Code, Claude Desktop, Claude Web, or OpenClaw - same results, same account, wherever you work.

How to use MiniMax H3 in Claude with CreativeClaw

Step 1: Connect CreativeClaw. Add the plugin using the connection guide. One account gives your agent access to H3 and the rest of the CreativeClaw model library.

Step 2: Pick the right mode. Use text-to-video for a scene from scratch, image-to-video to animate an opening frame, or reference-to-video when images, clips, and sound need separate jobs.

Step 3: Assign every reference. H3 uses natural-language, one-based labels: Image 1, Video 1, and Audio 1. Do not use Seedance-style @Image1 tokens or Grok’s <IMAGE_0> syntax.

Step 4: Direct the timeline. A strong H3 prompt names reference roles first, then gives 2 to 4 filmable beats, camera direction, sound, and continuity constraints.

Setup by client

Claude Code - Install the CreativeClaw plugin for the full experience with skills and optimized prompts. See setup guide.

Claude Desktop (Cowork) - Add the CreativeClaw MCP URL in your MCP server settings.

Claude Web (claude.ai) - Add CreativeClaw as a remote MCP server in your MCP settings. The plugin with advanced skills is coming soon, but the MCP tools work today.

OpenClaw - Add CreativeClaw as an MCP server in your configuration.

H3 prompt examples

Text-to-video

0-3s: A cyclist enters a rain-soaked city intersection at night. 3-7s: The
camera tracks beside her as reflections streak across the pavement. 7-10s:
She brakes beneath a warm shop light and looks back toward camera. One smooth
tracking move, realistic wheel spray and cloth motion. Stereo rain, tire hiss,
and distant traffic, no music. 10 seconds, 16:9.

Image-to-video with a final frame

Begin from the supplied first frame and end on the supplied last frame. The
presenter lifts the closed package, turns it once, and places it in the final
position. Slow push-in with natural foreground parallax. Preserve her identity,
wardrobe, package geometry, logo, table layout, and warm amber lighting. Soft
room ambience and one clean placement sound, no music.

Multimodal reference-to-video

Use Image 1 as the locked character identity and costume reference. Use Image 2
as the cafe environment and color reference. Follow the body movement and
camera rhythm of Video 1 without changing the character. Use Audio 1 as the
exact dialogue and timing source.

0-4s: She crosses the cafe and sits at the table. 4-8s: The camera arcs into a
medium close-up as she delivers the line from Audio 1. 8-10s: She smiles and
glances toward the window. Preserve her face, hair, costume, cafe layout, and
warm color grade. Connect the beats as continuous motion, not a slideshow.

H3 prompts are limited to 2,000 characters on the currently hosted endpoints. Be specific, but keep the direction compact.

Reference limits and syntax

Input	Limit	Prompt label
Images	Up to 9	Image 1, Image 2
Videos	Up to 3	Video 1, Video 2
Audio files	Up to 3	Audio 1, Audio 2
All references combined	Up to 12	Use each media type’s label

Reference video and audio files must each be 2 to 15 seconds, and each media type can total up to 15 seconds. An audio reference must be paired with at least one image or video reference.

Pricing in CreativeClaw

MiniMax H3 costs 52 credits per output second at 2K.

Duration	Cost
5 seconds	260 credits
10 seconds	520 credits
15 seconds	780 credits

The first 5 reference images are included in the base generation cost. Each additional reference image costs 16 credits. Reference video can add input-duration charges, so a tightly trimmed motion reference is usually both clearer and more economical.

When to use H3 vs alternatives

Model	Best for	References	Audio	Output
MiniMax H3	Detailed multimodal direction, continuity, edits	Image, video, audio	Native stereo	2K
Grok Imagine 1.5	Prompt adherence and image identity	Up to 7 images	Native	Up to 1080p
Seedance 2.0	Director-level multimodal storytelling	Image, video, audio	Native	Model-dependent
Kling 3	Cinematic human motion and multi-shot work	Frames and images	Tier-dependent	Up to 1080p

Use H3 when several source assets need distinct jobs and the final clip must include intentional sound. Use Grok for a simpler identity-preserving workflow, Seedance for its reference-driven directing style, or Kling when facial animation and cinematic motion are the priority.

Explore the full AI video generation guide to compare the complete model lineup.

FAQ

Does MiniMax H3 always generate audio?

Yes. H3 generates native stereo audio with every clip. Include clear audio direction such as dialogue, room tone, effects, or no music so the sound design is intentional.

Can H3 edit an existing video?

Yes. Use the source clip as a video reference and describe what should change and what must stay fixed. Reference images can preserve a character, product, or visual style while the video controls motion and timing.

Why did my H3 reference prompt fail?

Check the reference labels first. H3 requires Image 1, Video 1, and Audio 1 with no @ prefix. Also keep the prompt under 2,000 characters and stay within the reference count and duration limits.

Where can I verify the model details?

See fal’s MiniMax H3 overview and the reference-to-video API.