Claude MCP Setup for a Creative Production Workflow

Claude MCP lets Claude discover and call tools provided by an external server. For creative production, that can turn a conversation into a controlled workflow: read the right brand context, check available credits, generate an asset, and retrieve the result.

But a connected badge is not enough. A working setup has four separate proofs:

1. Claude can see the server.
2. The intended account is authenticated.
3. The expected read and generation tools are visible.
4. A small generation finishes and returns a retrievable asset.

This guide uses Claude Code's documented remote-server flow and Advibly as the creative server. The same principle applies elsewhere, but setup steps and permission controls can differ between Claude surfaces.

Key Takeaways

• Treat connection, authentication, tool discovery, and generation as separate proofs. A connected badge confirms only the first layer.
• Start with the narrowest useful scope, verify the exact account and brand through read-only calls, then approve one low-cost image test.
• Keep generation IDs and review every returned asset. Expand into a campaign only after the first paid proof is retrievable and accurate.

Before you connect anything

You need:

• Claude Code with MCP support;
• an Advibly account;
• an onboarded brand with current product and audience details;
• available generation credits; and
• the remote HTTP endpoint or setup command shown on Advibly's MCP setup page.

Advibly uses OAuth rather than a copied API key. That makes authentication simpler, but it does not tell you what the connected account can access. You still need to verify the account, brands, credits, and live tool list after sign-in.

Choose the narrowest configuration scope

Claude Code supports three MCP configuration scopes: local, project, and user. The right choice depends on who needs the connection and where its configuration should live.

Start with local unless you have a clear reason not to. A creative server can spend credits and change persistent brand context; convenience is not a compelling permissions model.

Anthropic's Claude Code MCP documentation is the authority for current commands, supported transports, and scope behavior. Use the exact remote HTTP setup instruction supplied by Advibly rather than retyping its endpoint from memory.

Connect and authenticate

1. Open Advibly's MCP setup page while signed in to the account you intend to use.
2. Copy the current Claude Code setup instruction or remote HTTP endpoint.
3. Add the server at your chosen scope using Anthropic's current MCP command format.
4. In Claude Code, open /mcp and complete the browser OAuth flow.
5. Return to Claude Code and confirm that the connected identity is the account you expected.
6. Run claude mcp list, note the configured Advibly server name, then inspect that connection with claude mcp get <name>, substituting the configured server name for <name>.

At this point, you have proved configuration and authentication. You have not yet proved that the creative workflow works.

Use a seven-step verification ladder

The fastest way to debug MCP is to test one layer at a time. If you jump from connection to a multi-asset campaign, one failure can have half a dozen plausible causes. The ladder below turns that mystery into a specific check.

The live tool list is the decisive evidence at step four. Advibly's documented MCP surface includes brand reads, credit checks, image and video generation, and asset retrieval. Tool availability can change with the server version or account permissions, so treat documentation as an expectation, not proof of what your current connection exposes.

Run reads before writes

Your first tool calls should not generate anything. Ask Claude to:

1. list the accessible brands;
2. identify the intended brand by its exact ID;
3. retrieve that brand's saved context; and
4. check the available credit balance.

Review the returned product name, audience, voice, colors, references, and offer details. If the context is wrong, stop there. A generation model cannot rescue the wrong product facts with sufficient enthusiasm.

This read-only preflight also confirms account scope. If Claude can see an unexpected brand, or cannot see the one you need, the problem belongs to authentication or permissions, not creative direction.

Make the first generation deliberately small

The minimum useful paid test is one static image. It is cheaper and simpler to inspect than a multi-shot video, while still exercising the full chain from brand selection to generation and retrieval.

Before running it, require Claude to show:

• the selected brand ID;
• the intended format and aspect ratio;
• the generation tool it plans to call;
• the inputs it will send; and
• the point at which credits will be spent.

Then approve one generation. Preserve the returned generation ID, poll or retrieve the result, and inspect the actual asset, not Claude's description of it.

Check:

• product and offer accuracy;
• logo and text integrity;
• visual consistency with the selected brand;
• aspect ratio and intended placement;
• policy-sensitive claims; and
• whether the asset can be retrieved again by ID.

A confident chat response without a generation ID and retrievable output is not a successful test.

Worked example: test a launch campaign without paying for the whole mistake

Suppose the goal is two static launch concepts and one short product-video concept for an existing brand.

First, connect the server and verify OAuth. Ask Claude to list brands, select the exact brand ID, retrieve its saved context, and check credits. Do not choose a brand by a similar display name alone.

Next, ask Claude to describe the proposed sequence of tool calls and identify which calls spend credits. Approve one static concept only. Once it finishes, retain its generation ID and inspect the output for product facts, offer accuracy, text rendering, logo use, and aspect ratio.

If the direction is wrong, correct it before generating the other assets. If it is sound, approve the second static and the video as separate units. Preserve every generation ID so one failed asset can be retried without restarting the campaign.

Finally, review all three outputs. Treat scheduling or publishing as a separate operation. Confirm that your live MCP connection exposes the required action before asking Claude to use it; a product's broader publishing capability does not prove that the same action is available through your current server connection.

This sequence spends a little evidence before it spends the full budget. That is the useful difference between an agent workflow and an expensive chain reaction.

Put approvals at consequential boundaries

The MCP specification's security guidance emphasizes consent, clear tool behavior, data protection, and access controls. In practice, approval before every harmless read creates friction without much protection. Approval matters most where the action has a durable consequence.

Require explicit approval before:

• the first credit-spending generation;
• a batch or multi-asset fan-out;
• any persistent change to saved brand context; and
• scheduling or publication.

You can allow ordinary brand and asset reads once the correct account and brand have been verified. Avoid automatic execution until the minimum generation test has passed and the workflow's paid calls are clear.

Diagnose the layer, not the whole workflow

Do not reconnect everything when one generation times out. If the server, authentication, reads, and first generation already passed, the fault is later in the chain. Retry the failed unit by ID or correct its inputs instead of rebuilding the connection.

When the dashboard is the better choice

MCP is not automatically the superior interface. The Advibly dashboard is the stronger alternative when you need one asset, want visual control over brand and project selection, or cannot verify the tools exposed in Claude. It has fewer integration layers and makes the selected workspace visible.

Claude MCP becomes more useful when the work already starts in Claude and requires a repeatable sequence: retrieve brand context, inspect credits, produce several related assets, retain IDs, pause for approval, and continue in the same working session.

Choose the dashboard for the shortest one-off path. Choose MCP when orchestration and repeatability are worth the additional verification.

Where Advibly fits

MCP standardizes how applications expose tools and context to models; it does not guarantee the quality or completeness of a creative workflow. The server still needs useful product capabilities behind the connection.

Advibly's role is specific: it turns saved brand and product context into images, videos, ads, carousels, posts, and campaign assets through an AI creative generator. Its installable creative skills can add a repeatable workflow for formats such as UGC-style ads, explainers, restyling, claymation, 3D animation, and collage motion.

Use a skill only after the connection ladder passes. A readable workflow can organize the work, but the live server must still expose the required tools, and a human should still approve spend and publication.

If your goal is repeatable creative production from Claude, connect Advibly at the narrowest useful scope, prove one retrievable image, and only then expand the run.

Sources

• Anthropic: Connect Claude Code to tools via MCP
• Anthropic: Model Context Protocol
• Model Context Protocol: Security and Trust & Safety
• Advibly: MCP setup
• Advibly: Agent Skills
• Advibly public skills repository