How to Generate AI Videos From Claude Using the imageat MCP Server

Claude can turn a rough campaign idea into a shot plan, tighten a motion prompt, and review whether the instructions contradict each other. Connected to the imageat MCP server, it can also call video-generation tools from the same conversation instead of stopping at a prompt you must copy into another tab.

This guide is specifically about making videos from Claude. It covers the two current connection paths, a safe first test, text-to-video and image-to-video prompt patterns, review checkpoints, credit-aware habits, and fixes for common MCP and generation problems.

The workflow does not remove the need for creative judgment. Claude still needs a clear brief, and every output still needs review. MCP improves the handoff: planning, tool use, and revision can share one working context.

What Claude, MCP, and imageat each do

It helps to separate the three layers before connecting anything.

• Claude interprets the goal, asks or reasons through missing constraints, structures the prompt, and decides when an available tool should be called.
• MCP is the connection standard that lets a compatible AI client discover and use external tools.
• imageat runs the image, video, and editing task and charges the connected imageat account according to the selected workflow.

The live imageat MCP page currently lists text-to-video and image-to-video capabilities with video model families including Veo, Kling, and Seedance. It also lists image generation, more than 20 editing tools, and a credit-balance check. The exact tools and models visible to Claude can change, so ask Claude to inspect the connected tool list rather than relying on an old tool name from a tutorial.

If you want a protocol-level overview across several clients, read how to use MCP for AI image generation in Claude, ChatGPT, and Cursor. The article you are reading has a narrower intent: producing and refining AI video from Claude.

Choose the right Claude connection path

The current imageat MCP connection guide distinguishes between claude.ai and Claude Desktop.

claude.ai: remote connector and browser sign-in

For claude.ai, imageat documents a remote connector flow. The endpoint shown on the live page is:

https://mcp.imageat.com/mcp

Add a custom connector in the connector controls available to your Claude account, paste the endpoint, and complete the imageat sign-in and approval flow. The live imageat instructions classify claude.ai as a remote OAuth connection, so you do not copy an imageat API key into the remote connector form.

Claude interface labels and connector availability can vary by plan, workspace policy, region, and product release. Use the current imageat setup panel as the connection source of truth, and confirm that custom connectors are available in the Claude environment you intend to use.

Claude Desktop: local stdio connection

For Claude Desktop, the live imageat page documents a local connection through npx and an imageat API key. Create or manage the key from imageat Projects, then use the configuration displayed in the current MCP setup panel.

Do not paste a secret key into a chat message, screenshot, shared prompt library, or public configuration repository. Store it only where the local MCP configuration expects it. If a key is exposed or a device is no longer trusted, revoke that connection from imageat Projects and create a replacement.

Because package commands and configuration fields can change, this guide does not freeze a copied desktop JSON snippet that may become stale. Open the Claude Desktop tab on the live imageat MCP page, copy the current configuration, and verify that the package and environment-variable names match what the page shows today.

Run a small connection test first

Do not begin with a multi-shot commercial. First prove that Claude can see imageat, check the intended action, and return a result.

Start by asking Claude to inspect rather than generate:

Inspect the connected imageat tools. Confirm whether you can access a video
creation tool and a credit-balance check. Do not generate anything yet. Summarize
the available video inputs and controls without guessing unsupported options.

Then request one restrained clip:

Use imageat to create one short 16:9 text-to-video test.

Scene: a ceramic cup on a wooden table beside a rain-streaked window.
Action: steam rises gently while one raindrop moves down the glass.
Camera: locked tripod shot, no pan, no zoom, no cut.
Lighting: soft overcast window light, realistic shadows.
Avoid: text, logos, extra objects, camera shake, sudden motion.

Before using the tool, restate the requested format, motion, and exclusions.
Generate one result only.

This test is intentionally simple. One subject action, one environmental action, and a fixed camera make it easier to diagnose whether the connection works and whether Claude translated the brief correctly.

How to generate a video from text in Claude

Text-to-video is useful when you do not need to preserve an existing person, product, room, or composition. The imageat MCP video generator is the dedicated landing page for agent-driven video generation, while the AI video generator provides the direct web workflow.

Step 1: state the destination

Tell Claude where the clip will be used before describing its visual style. A vertical social hook, a wide website loop, and a product-page demonstration need different framing.

Include:

• destination or placement;
• orientation or aspect ratio;
• whether the clip must loop or end on a clean hold;
• whether you need one result or several reviewed variations;
• any safe-area requirement for copy that will be added later.

Step 2: separate subject, camera, and environment

Many weak prompts mix every type of movement into one sentence. Give Claude three separate motion channels:

• Subject motion: what the person, product, animal, or object does.
• Camera motion: what the virtual camera does.
• Environmental motion: wind, particles, water, fabric, traffic, reflections, or background activity.

A reliable prompt asks for one primary motion in each channel. More movement is not automatically more cinematic.

Step 3: describe a believable timeline

Write the clip as an opening state, development, and end state. This prevents a prompt from becoming a list of disconnected events.

Use imageat to generate one 9:16 product teaser.

Purpose: an organic social reveal for a matte black desk lamp.
Opening: the lamp is off in a dim, tidy workspace; the product is fully visible.
Development: the lamp turns on gradually while the camera makes a short, smooth
sideways slide. A notebook page moves slightly from a nearby fan.
End: hold on the illuminated desk with the lamp still fully in frame.

Continuity: preserve the lamp's shape, base, arm joints, finish, and position.
Avoid: cuts, hands entering frame, added labels, changing product geometry,
flickering light, fast zooms, and unreadable text.

First check the connected imageat video options. Choose only controls that are
actually exposed. Explain the planned call, then generate one version.

The instruction to inspect exposed controls matters. It keeps Claude from pretending that every model offers the same resolution, duration, audio, reference, or camera settings.

How to generate video from an image in Claude

Image-to-video gives the generator an approved visual starting point. It is usually the better choice when identity, product geometry, wardrobe, packaging, or composition must remain recognizable.

A strong source image does most of the continuity work. Use a sharp frame with a readable subject, complete edges, plausible anatomy, simple depth layers, and enough space for the intended movement. The detailed image-to-video guide explains the broader browser workflow.

Tell Claude what the image controls

Do not say only “animate this.” Define which attributes are locked and which may move.

Use the attached image as a strict reference for the bottle's silhouette, cap,
label placement, material, color, and position. It also controls the table layout
and the opening camera angle.

It does not control the final crop or environmental motion. You may add a subtle
camera push-in and natural condensation movement. Do not redesign the bottle,
rewrite the label, rotate the product, add hands, or replace the background.

Exact typography remains a high-risk area in generative video. If label fidelity is essential, keep movement modest, review the label frame by frame, and plan to composite approved text or packaging art in post-production rather than asking the model to recreate it.

Use restrained motion for the first pass

A source image already establishes the world. The first video request should animate that world, not replace it.

Create one image-to-video clip with imageat from the approved reference.

Subject motion: the model makes a small natural breath and turns her eyes toward
camera; keep facial identity, hairline, clothing, and body proportions stable.
Camera motion: a very slow push-in only.
Environment: curtains move slightly from a soft breeze.
Continuity: preserve face, hands, jewelry, background geometry, and lighting.
Avoid: speech, lip movement, a head turn, new objects, cuts, zoom bursts, beauty
filtering, extra fingers, and a frozen final frame.

Return the first result for review before creating any variation.

For image-to-video prompt ideas organized by motion type, use the cinematic image-to-video prompt library.

A reusable Claude video brief

Save this template in a Claude Project or team prompt library and fill only the fields that matter.

Use the connected imageat tools to create [ONE / NUMBER] video result(s).
Do not generate until you have checked the available tool inputs.

Goal: [BUSINESS OR CREATIVE PURPOSE]
Placement: [CHANNEL, ASPECT RATIO, SAFE AREA]
Input: [TEXT-TO-VIDEO / ATTACHED IMAGE AS REFERENCE]
Subject: [WHO OR WHAT MUST APPEAR]
Opening state: [FIRST FRAME]
Subject motion: [ONE PRIMARY ACTION]
Camera motion: [ONE CAMERA MOVE OR LOCKED CAMERA]
Environmental motion: [ONE OR TWO SUBTLE EFFECTS]
Lighting: [SOURCE, DIRECTION, QUALITY, COLOR]
Style: [REALISTIC, EDITORIAL, DOCUMENTARY, ANIMATED, ETC.]
Continuity locks: [IDENTITY, PRODUCT, WARDROBE, LAYOUT, PALETTE]
End state: [FINAL COMPOSITION OR LOOP CONDITION]
Avoid: [CUTS, TEXT, LOGOS, DISTORTION, EXTRA OBJECTS, UNWANTED MOTION]

Before calling imageat:
1. restate the brief in five lines;
2. flag any physical or visual contradiction;
3. list only the tool settings you can actually control;
4. check credits if the workflow exposes that capability;
5. create one draft and stop for review.

This template makes Claude act as a preflight reviewer, not merely a prompt relay.

Build a multi-shot workflow without wasting generations

For a sequence, Claude should plan first and generate serially.

1. Convert the idea into beats

Ask for three to five story beats, each with one purpose. A product ad might use problem, discovery, demonstration, proof, and end frame. Do not generate during this stage.

2. Turn each beat into a shot card

Every shot card should state subject, framing, action, camera motion, lighting, continuity locks, and transition. Reject shots that require too many events in one generation.

3. Choose text-to-video or image-to-video per shot

Use text-to-video for establishing shots where exact identity is unimportant. Use image-to-video when a person, product, wardrobe, room, or composition must carry across the cut. You can create a controlled reference still first with the AI image generator, approve it, and then animate it.

4. Generate one shot at a time

Have Claude call imageat for shot one, return the result, and wait for evaluation before continuing. If the product or character drifts, fix the reference and continuity language before producing later shots.

5. Edit outside the generation loop

Use a video editor for precise timing, cuts, color matching, captions, music, sound effects, logos, and final typography. Generating the full edit as one clip gives you less control over continuity and pacing.

For a complete planning-to-editing method, see the storyboard-to-final-cut AI video workflow.

How to review the result with Claude

Claude can help structure the review, but it should not declare a clip usable without specific checks. Ask it to evaluate a provided frame set or your written observations against a checklist.

Review in this order:

1. Brief compliance: subject, action, framing, and end state.
2. Continuity: face, hands, product shape, clothing, props, and background layout.
3. Physics: weight, contact, shadows, reflections, liquid, fabric, and acceleration.
4. Camera: consistent perspective, believable lens behavior, and intentional motion.
5. Artifacts: morphing, duplicate objects, unstable edges, flicker, unwanted text, or frozen frames.
6. Delivery fit: crop, safe area, pacing, and whether the final frame supports the intended edit.

A useful revision request changes one class of problem:

Keep the approved subject, scene, framing, lighting, and camera path unchanged.
Revise only the hand action: the fingers should remain wrapped naturally around
the cup without changing shape or separating from it. Do not add any new motion.
Generate one revision only.

If several fundamentals are wrong, do not keep stacking correction language. Return to the source frame or rewrite the shot with less motion. The guide to why AI videos look fake provides a detailed failure checklist.

Credit-aware production habits

The live MCP page says imageat MCP calls use imageat credits and offers a credit-check capability. It does not mean every video task has one permanent cost. Required credits can depend on the model and workflow, so inspect current controls before running a batch.

Use an explicit gate:

Before any generation, check my available imageat credits if that tool is
available. Tell me how many outputs you plan to request. Do not create more than
one output without my confirmation, and do not automatically upscale or rerun a
failed concept.

Other practical controls:

• solve the creative direction in text first;
• generate one shot before a sequence;
• lock orientation before generation;
• use the same approved reference for continuity;
• change one variable per revision;
• stop when identity or product geometry drifts;
• reserve high-cost finishing choices for selected clips;
• record the prompt and reference used for every approved shot.

For a planning framework that avoids fixed price claims, read AI video credits explained.

Troubleshooting Claude and imageat MCP

Claude cannot see imageat tools

Confirm that the connector is enabled in the current conversation or workspace. For claude.ai, check the remote connector authorization. For Claude Desktop, confirm the local MCP configuration, API key, runtime, and application restart requirements shown in the current setup instructions. Ask Claude to list connected tools before testing a generation.

The remote connection does not authorize

Verify that the endpoint is exactly the HTTPS remote MCP address displayed on the live imageat page. Make sure you are signing into the intended imageat account and that browser redirects are not blocked. Remove a broken connector and repeat the current authorization flow rather than inventing OAuth credentials.

Claude writes a prompt but does not generate

Make the action explicit: “Use the connected imageat tool to generate one video now.” If you asked for approval before generation, provide that approval only after checking the summary. Also confirm that a compatible video tool is visible to Claude.

Claude chooses unsupported settings

Ask it to inspect the tool schema and distinguish requested preferences from actual controls. Do not assume every available video model supports the same input mode, aspect ratio, duration, resolution, references, or audio behavior.

Image-to-video changes the subject

Use a cleaner source image, reduce motion, identify the exact attributes controlled by the reference, and remove unnecessary camera movement. A stable product shot with one small action is a better diagnostic than a complex scene.

Motion looks chaotic or artificial

Separate subject, camera, and environment. Keep one primary action in each category, describe the opening and end state, and remove conflicting instructions such as a locked camera combined with a sweeping orbit.

Text or logos deform

Avoid asking the generator to invent exact typography. Preserve simple packaging where possible, keep motion restrained, and add approved text or logos during editing. Review every frame in which the mark appears.

Claude starts generating too many variations

State the limit at the top and bottom of the request: one result, no automatic retries, return for review. Ask for concept alternatives in text before any tool call.

FAQ

Can Claude generate AI videos directly?

Claude can call an external video tool when a compatible MCP connection is enabled. With imageat MCP connected, Claude can structure the request and use available imageat text-to-video or image-to-video capabilities from the conversation.

Do I need an imageat API key for claude.ai?

The current imageat instructions classify claude.ai as a remote sign-in flow with no API key copied into the connector. Claude Desktop uses the local npx and API-key path documented on the same live setup page.

Which video models can Claude use through imageat?

The live imageat MCP page currently names Veo, Kling, and Seedance among its video model families. Availability and exposed controls can change. Ask Claude to inspect the connected tools before choosing a model.

Can I upload an image and animate it?

Yes, when the connected imageat video workflow exposes image-to-video input. Tell Claude what the reference must preserve and describe motion separately from identity, product, and composition constraints.

Does MCP make video generation free?

No. The live imageat MCP page says calls are billed against your imageat credits. Check the current balance and workflow requirements before generating.

Is Claude better than the imageat web interface?

They serve different working styles. Claude with MCP is useful when the creative brief, shot plan, and revision history already live in a conversation. The web interface is useful for direct visual operation and manual model browsing. A production workflow can use both.

Can Claude create a complete multi-shot ad automatically?

It can help plan and generate individual assets, but a reliable ad still needs human review, continuity control, editing, typography, sound, and delivery checks. Generate serially instead of treating one instruction as permission for an unchecked batch.

How do I disconnect imageat MCP?

Remove or disable the connector in Claude, then revoke the related authorization or key from imageat Projects. The exact control depends on whether you used the remote claude.ai flow or the local Claude Desktop configuration.

Start with one controlled shot

The most useful first Claude and imageat workflow is not “make an entire campaign.” It is one clear shot with one subject action, one camera decision, and explicit continuity locks.

Open the imageat MCP setup page, choose the instructions for claude.ai or Claude Desktop, and ask Claude to inspect the connected video tools before it generates. Once the connection is confirmed, use the reusable brief above to produce one result, review it against the six-part checklist, and make only one focused revision.

That small loop—plan, inspect, generate, review, revise—is the foundation for larger agent-assisted video production.