How to Generate Product Photos in Claude with MCP

To generate product photos in Claude, Claude needs to direct an external image-generation service. Claude can read a brief, inspect uploaded pack shots, turn requirements into a structured request, and call approved MCP tools. It does not render the final product photo itself; the connected service creates the image.

That division is useful. Claude can act like a producer—clarify the shot list, protect invariants, check accepted inputs, request a price, and track the job—while a person decides what is accurate enough to buy and publish.

| Input | Why it matters | Automatic no-go | |---|---|---| | Clean product reference | Establishes shape, materials, color, and label placement | Product is obscured, blurry, or the wrong SKU | | Shot specification | Defines crop, angle, setting, light, and placement | “Make it premium” is the whole direction | | Invariant list | Protects details that must remain true | No owner can say which details are factual | | Rights note | Records permission to use references | Image origin or talent permission is unclear | | Review rubric | Makes approval consistent | “Looks good” is the only standard |

If you need product scenes today, use Oakgen's current AI image generator with a clean product reference. Use AI Chat to structure the shot list before production. The MCP workflow in this article is a release-ready method, not a live Oakgen connection tutorial.

Who this guide is for

This workflow is for an ecommerce marketer, agency producer, catalog manager, or small brand owner who has a real product and needs controlled image concepts: a clean listing set, lifestyle contexts, seasonal scenes, or ad-ready crops.

It is not a recipe for replacing product truth with plausible-looking pixels. If a shopper needs to see the exact finish, component, fit, texture, safety feature, or included accessory, the final image must communicate the actual SKU.

For the broader catalog process, see MCP image generation for ecommerce teams. Before connecting any creative service, use the MCP security checklist for creative teams to review trust, permissions, approvals, and asset handling.

Research note

This article was checked on July 24, 2026 against Anthropic's official statement about Claude image output, Anthropic's vision documentation, Amazon's current Seller Central image guidance, and Oakgen's deterministic connected-generation contract. The prompts below are workflow templates, not the results of an unperformed Oakgen-to-Claude test.

Anthropic distinguishes image understanding from image generation: Claude can work with uploaded images, while an external tool is needed to render a product photo. That is why the approval and accuracy process remains important even when the conversation feels direct.

Start with the product-photo input shot list

The generator can only preserve details it can see or that you describe unambiguously. A single front photo may be enough for a loose concept, but it is rarely enough for a catalog set.

Product authority

• [ ] Front or three-quarter view in even, neutral light
• [ ] Side and back views when geometry or labeling changes
• [ ] Top or bottom view for closures, bases, and openings
• [ ] Close-up of logo, label, controls, stitching, or texture
• [ ] Color reference photographed beside a neutral gray card when color is critical
• [ ] Packaging and included components shown separately
• [ ] Correct SKU, size, material, and variant recorded in text

Creative direction

• [ ] Required image job: main listing, detail, scale, use, or lifestyle
• [ ] Exact aspect ratio and target placement
• [ ] Camera height, lens feel, and product angle
• [ ] Light direction, softness, shadow, and background
• [ ] Props that are allowed and props that could mislead
• [ ] Details that may change and details that must not
• [ ] Final owner and no-ship conditions

The best reference is often boring: evenly lit, sharp, neutral, and complete. A beautiful source image with clipped edges and hard reflections can hide the facts the new scene needs.

A five-step Claude-to-image workflow

1. Have Claude audit the references

Ask Claude to inventory what is visible and what is missing. It should not infer the unseen back panel, package contents, or finish.

Review these product references as evidence. List visible facts, uncertain details, and missing views. Separate product invariants from creative choices. Do not draft a generation prompt until I answer the uncertainty questions.

This first step catches wrong-SKU work before it becomes an attractive error.

2. Build a shot contract

For every requested image, record:

• job and destination;
• product angle and scale;
• background and lighting;
• approved props;
• crop and safe area;
• visible invariant details;
• declared non-literal elements;
• no-ship conditions.

A “summer lifestyle shot” is not a contract. “4:5 paid-social concept; bottle on pale limestone; hard noon shadow; no fruit, hands, liquid splash, text, or changed label” is much closer.

3. Match a model to the required inputs

Oakgen's planned workflow begins with search_models, then get_model_schema. The agent should look for a model that can express the actual job, including any required reference input. If the schema does not accept what the plan needs, switch the plan or the model. Do not ask Claude to fabricate a field.

Read the model-selection and pricing guide before building a batch.

4. Quote the proof image

The intended get_pricing operation checks the exact request before a paid start. Quote one difficult, representative shot first. It should contain the package details and scene conditions most likely to fail.

The approval record should show the chosen model, dimensions, number of outputs, prompt, references, quoted amount, and the person approving. Only then should start_generation begin the job.

5. Follow the job; do not restart it

Keep the returned handle and use get_generation_status. If the image is taking longer than expected, a repeated start can create a duplicate. If the direction is clearly invalid, cancel_generation is intended to request cancellation; keep checking the original handle until the final state is known.

The safe recovery order is documented in the image generation MCP troubleshooting guide.

Five product-photo prompts to copy and adapt

Each prompt assumes the product reference and invariant list travel with it. Replace the bracketed creative details with real instructions; never replace the product facts.

1. Marketplace-style main image

Create one square product image using the uploaded SKU reference as the product authority. Preserve package geometry, cap, color, label placement, and included components. Show only the product being sold against a clean white background, centered and large in frame, with a physically plausible soft grounding shadow. Add no text, badge, prop, hand, duplicate, or accessory. Return the image for compliance review; do not describe it as marketplace-approved.

2. Honest scale image

Create a 4:5 product scene that communicates real scale using the supplied dimensions. Place the product beside one ordinary, non-included object only as a scale cue. Preserve the exact SKU and do not imply that the prop is included. Use neutral daylight and leave the upper-right area quiet for later layout.

3. Ingredient or material context

Build a landscape editorial scene around the approved material story: [material]. Keep the physical product unchanged. Props may evoke the source material but must not imply an unapproved ingredient, certification, origin, or performance claim. No generated copy inside the image.

4. In-use scene

Show one plausible use moment for [audience and context]. The product must be held and oriented as it is used in reality. Preserve controls, openings, scale, and safety details. If hand anatomy or product contact is wrong, reject the result instead of hiding it with crop.

5. Seasonal variation

Keep the approved base composition and product angle fixed. Change only the surrounding styling to [season or event] using removable props and background color. Do not change packaging, add a limited-edition label, or imply a seasonal SKU that does not exist.

These prompts are specific about refusal conditions because “more detailed” is not automatically safer. A long prompt can still contradict the source image.

Product-accuracy review rubric

Score each dimension as pass, revise, or reject. A reject on a product-truth row cannot be averaged away by beautiful lighting.

| Review dimension | Pass | Revise | Reject | |---|---|---|---| | SKU identity | Correct variant and package | Minor cleanup away from identity | Wrong color, size, model, or package | | Geometry | Shape and proportions match | Small background-edge artifact | Warped, added, or missing physical part | | Label and text | Exact or deliberately composited later | Known typography area masked for finishing | Invented claim, misspelling, or false mark | | Color and material | Plausible match under declared light | Grade needs correction | Material or finish changed | | Contents | Only real included items | Ambiguous prop needs disclosure | Invented accessory or quantity | | Use | Physically plausible | Pose or contact needs adjustment | Unsafe or impossible use | | Scene | Supports the brief | Distracting prop or reflection | Scene creates a false product promise | | Crop and quality | Placement-ready resolution and safe area | Straightforward crop or retouch | Important product detail clipped or unreadable |

No-ship examples

• The bottle is elegant, but the ingredient line has changed.
• The chair looks realistic, but one support has disappeared.
• The food appliance includes a blade not sold with the SKU.
• The “scale” shot makes a travel-size item look full-size.
• A hand covers the one control the buyer needs to understand.
• A lifestyle prop suggests a medical, environmental, or performance claim the brand has not approved.
• The main listing image contains decorative objects the shopper could mistake as included.

Here is where this workflow breaks: generators are optimized to create a convincing image, not to maintain your catalog record. Convincing and true are different review columns.

You can use this rubric with the current Oakgen image generator: attach it to the brief, review the output at full size, and require a written reject reason instead of rolling another random variation.

Marketplace and ad review

Amazon's public Seller Central guidance says listing images should accurately represent the product. Its main-image guidance also addresses white backgrounds, frame fill, overlays, and extra props. Requirements can vary by category and change, so check the current rules for the account and marketplace before upload.

For paid social, review the image in the complete ad: creative, copy, destination, and audience. A visually accurate photo can still support a misleading claim. Platform acceptance is not a substitute for the brand's own review.

Common mistakes

Asking for a full set before proving identity

Run the hardest representative shot first. Batch production multiplies both good direction and hidden defects.

Treating generated typography as finished packaging

When exact label text matters, preserve it from a verified source or plan a controlled compositing and retouching step.

Hiding uncertainty in the prompt

If the back, contents, or dimensions are unknown, get the facts. Do not turn “probably” into a photorealistic assertion.

Approving at thumbnail size

Inspect labels, joints, reflections, edges, fingers, and repeated objects at full resolution, then inspect the crop at its real placement size.

Calling a concept a final photograph

Keep provenance and approval status in the asset manifest. Decide explicitly whether disclosure or labeling is required for the intended channel and jurisdiction; do not assume.

Frequently Asked Questions

Can Claude generate product photos?

Claude does not render generated photos in chat. It can analyze references, plan a shot, and direct a connected external image service through MCP; that service creates the pixels.

Can I connect Oakgen MCP to Claude now?

Not as a public production connection. Oakgen has not released a public endpoint, credential flow, launch date, or dated Claude compatibility result.

What reference photos should I upload?

Provide a clean front or three-quarter pack shot plus any side, back, top, texture, label, and color views needed to understand the product. Include only images you are allowed to use.

Can AI product photos be used as marketplace main images?

Check the marketplace's current rules and review the output against the actual SKU. For example, Amazon says the main image must accurately represent the product and meet specific background, framing, and overlay requirements.

How do I keep a product label accurate?

Treat the original pack shot as the authority, mark label text and placement as invariants, reject altered copy, and plan manual compositing or retouching when the generator cannot preserve exact typography.

Why should the agent request a quote before generation?

The quote lets the operator approve the exact model, dimensions, output count, and reference inputs before paid work starts rather than discovering the cost after a batch.

What product images should never ship?

Reject images with changed package geometry, wrong variants, invented accessories, unreadable or altered labels, misleading scale, unsafe use, impossible reflections, or unapproved claims.

Sources and Further Reading

• Can Claude produce images? — Anthropic's current explanation of Claude's image-output boundary.
• Vision with Claude — official guidance for image inputs and analysis.
• Amazon Seller Central product-image guidance — current official requirements summary and troubleshooting direction.
• Model Context Protocol introduction — official host, client, server, and tool concepts.

When the references are ready, build the first controlled product concept in Oakgen. Prove product truth before you pay to multiply the scene.