A useful creative MCP workflow for marketing agencies does not begin with “make 20 ads.” It begins with one signed-off campaign brief, a list of facts that cannot change, and a testing question. The agent can then help find an image model, read its accepted inputs, show the price, start approved jobs, track them, and return organized assets.
The agency still owns the consequential decisions: whether the brief is accurate, whether the work is worth buying, whether a product is represented honestly, and whether anything is ready for a client.
| Stage | Agent can coordinate | Human must decide | |---|---|---| | Intake | Turn the brief into fields and flag gaps | Approve claims, audience, offer, rights, and exclusions | | Planning | Propose a controlled variation matrix | Choose the actual testing question | | Preflight | Find models, inspect inputs, request a quote | Approve model, payload, and spend | | Production | Start jobs and follow their handles | Stop a weak direction before it becomes a batch | | Review | Sort outputs against a rubric | Approve, reject, or request a precise revision | | Delivery | Apply names and assemble a manifest | Confirm final crops, policy, and client sign-off |
If you need to produce the current campaign now, shape the brief in Oakgen AI Chat, generate stills in the AI image generator, and build UGC-style concepts in the UGC ads workflow. The connected workflow discussed below is a future-facing operating model, not a live Oakgen MCP offer.
Oakgen's public production MCP has not launched. The planned first surface is limited to image generation with six operations. This guide does not provide an endpoint, credential flow, plan requirement, host promise, or launch date.
Who this workflow is for
This is for the person who has to turn a client call into usable work by Friday: an agency producer, performance creative lead, account director, paid social manager, or hands-on founder.
It fits teams that need repeatable image concepts and variants but cannot afford a free-roaming agent that changes the offer, invents product details, or spends against a vague prompt. It is less useful when the job is one hero visual that deserves close art direction from an experienced designer.
The mistake is treating volume as the objective. Twenty unrelated pictures are not a test set. They are twenty new variables.
Research note
This workflow was reviewed on July 24, 2026 against the official Model Context Protocol introduction, Meta's current photo-ad guidance, and Oakgen's deterministic connected-generation contract. The workflow is a production template, not a performance study. We did not invent CTR lifts, labor savings, or customer results.
For production planning, pair the controlled ad-variation workflow with the agent model-selection and pricing worksheet. If the agency is still choosing a connected service, use the image-generation MCP buyer's guide before adapting the process.
The agency intake template
Before an agent sees the assignment, turn the client brief into this one-page intake. Empty cells are questions, not permission to improvise.
| Field | What belongs here | Example | |---|---|---| | Business objective | The commercial job of the campaign | Acquire first-time buyers for the starter kit | | Audience | One useful segment, not “everyone” | Busy renters with small kitchens | | Offer | Exact approved offer and dates | Starter kit, 15% off through August 31 | | Proof | Approved facts the creative may use | Dishwasher-safe; fits a 30 cm shelf | | Product authority | Source images and SKU record | Front, back, top, package, color reference | | Visual invariants | Details that cannot move | Handle shape, logo position, coral color | | Creative variable | What this batch is allowed to test | Context: tiny kitchen vs. picnic vs. office | | Placement | Crop and safe-area requirements | 4:5 feed still | | No-ship rules | Automatic rejection conditions | Altered logo, invented accessory, false claim | | Owner | Person who can approve spend and output | Named creative director |
Add a rights note for every uploaded reference: who owns it, what it may be used for, and when it should be removed. “The client sent it in Slack” is not a rights record.
From one brief to a controlled batch
1. Separate invariants from variables
The invariant brief is the spine of the batch. Product geometry, approved claims, offer, audience, logo treatment, and placement stay fixed. A useful test changes one substantial element: the visual hook, setting, product angle, or composition.
If the agent changes the headline, setting, audience, crop, color system, and product pose at once, a winner teaches the media team almost nothing.
2. Ask for a plan before images
Have the agent return the matrix, not pixels. The producer should be able to point to every row and explain what it tests.
An instruction worth saving:
Convert this approved brief into a 20-cell image plan. Keep the product, audience, offer, claim, placement, and brand rules fixed. Use four visual hooks and five controlled executions per hook. Name the changed variable in every cell. Stop before model selection or generation.
Reject cells that are merely synonyms. “Premium studio,” “luxury studio,” and “high-end studio” are not three hypotheses.
3. Discover the model from the job
In Oakgen's planned image-first workflow, the agent would use search_models to find relevant options and get_model_schema to inspect the chosen model's real input contract. That order matters. A model that cannot accept the required reference input is not suitable, however fashionable it is.
The schema check should answer:
- Can this job use the product reference in the required way?
- Which fields are required?
- Which aspect ratios or sizes are accepted?
- How many outputs can the request ask for?
- Which convenient-sounding prompt controls do not actually exist?
Do not let the agent invent a parameter because the brief would benefit from it.
4. Quote the exact request
get_pricing is intended to quote the real payload before execution. Quote a small proof set first: one representative cell from each visual hook. Approve the larger batch only after the team sees that the direction is viable.
Cost approval should include the model, input summary, output count, quoted amount, budget owner, and timestamp. A quote for one image is not approval for twenty.
5. Start once, then follow state
After approval, start_generation would begin durable work. Save the returned job handle beside the asset ID. If a response times out or a job takes longer than expected, check it with get_generation_status. Do not press start again because the image has not appeared.
In plain language: retry the lookup, not the purchase.
If the creative direction is plainly wrong, cancel_generation is intended to request cancellation. A request is not proof that work stopped. Keep checking the same job until it reaches an authoritative final state. The image-generation MCP troubleshooting guide covers the diagnostic order.
Build the proof set before the batch
Use Oakgen's current image generator to test one representative concept from each visual hook before producing the full delivery set.
The 20-variation planning matrix
This is the linkable worksheet for the article. Replace the example labels, but preserve the structure: four hypotheses, five executions, one declared variable per row.
| IDs | Visual hook | Five executions | What remains fixed | Review question | |---|---|---|---|---| | A01–A05 | Product in a constrained space | shelf, drawer, corner, counter, carry bag | Product, offer, 4:5 crop, coral palette | Does compactness read instantly? | | B01–B05 | Visible before/after order | clutter/clear, packed/open, hidden/shown, far/close, busy/calm | Same product angle and claim | Is the transformation honest without copy? | | C01–C05 | Hands using the product | unpack, place, hold, clean, store | Same hand styling and environment | Does use clarify function without adding parts? | | D01–D05 | Material and detail proof | texture, join, edge, handle, scale | Same lighting family and crop logic | Does the detail support the approved proof? |
Name files so a strategist can read the test without opening them:
client_campaign_hook-execution_placement_v01_status.ext
Example: northstar_launch_a03_feed45_v01_client-review.png
Never overwrite a rejected file. Status belongs in the manifest and the filename until delivery; the final export can use the client's naming convention.
To pressure-test this matrix in the current product, generate one representative image per hook. Keep the four outputs beside the hypotheses and reject weak directions before expanding them.
Approval checkpoints that keep the workflow accountable
Gate 1: brief lock
The account lead confirms audience, offer, claim, dates, rights, and no-ship rules. The agent may identify ambiguity but may not resolve commercial facts.
Gate 2: proof-set spend
The creative owner approves model, schema-bound inputs, quote, and four representative concepts. This is the cheapest place to kill a weak direction.
Gate 3: batch expansion
Only proven hooks expand. The producer records which prompt and reference set produced each approved direction.
Gate 4: internal quality control
Review at full size and at placement size. Check product shape, labels, hands, reflections, crop safety, brand color, duplicated objects, and unintended text.
Gate 5: client review
Show a curated set with IDs and a one-line hypothesis. Do not ask a client to excavate a generation dump. Separate “accurate but not preferred” from “factually unsafe.”
Gate 6: delivery
The delivery manifest records dimensions, placement, version, approval owner, and source asset IDs. A pretty folder is not traceability.
Client review rules
Give clients three decisions:
- Approve: accurate, on-brief, ready for final finishing.
- Revise: direction is sound; one named defect needs work.
- Reject: violates an invariant or fails the testing hypothesis.
Avoid “thoughts?” It invites a new brief after production.
Also be honest about the output stage. An AI concept may be good enough for a low-risk test and still need typography, pack retouching, legal review, or a real product shoot for the final hero campaign. The workflow should make that distinction visible.
Where the workflow breaks
The brief contains unapproved claims
Stop at intake. An agent should not turn a founder's wish into ad copy or visual proof.
The product reference is weak
More prompt detail does not repair a blurry label, hidden side, or incorrect color. Ask for better source photography.
Every cell changes everything
Return to the matrix. If the team cannot state the variable, the media result will not clarify it later.
The job appears stuck
Preserve the handle and retrieve status. Re-starting blindly risks duplicate work and muddy cost records.
The client wants “more options”
Ask what question the new options should answer. Generate a new row only after a new hypothesis exists.
The image passes review but fails policy
Platform review remains separate from creative review. Meta says ad review may consider the image, text, targeting, and destination. The person publishing the ad must verify the current platform rules.
A practical decision framework
Use an agent-led MCP workflow when the work is repeatable, inputs can be structured, quotes need approval, and jobs must survive beyond one chat response.
Use Oakgen's current web tools when a producer wants direct visual iteration today without setting up a connected tool.
Use a designer-led workflow when typography, novel layout, exact retouching, or brand expression is the main value.
Use a real shoot when accurate physical proof, talent performance, regulated claims, or final hero photography cannot tolerate synthesized ambiguity.
The strongest agency workflow is usually hybrid: agent for planning and coordination, image tools for concept production, specialists for finishing, and humans for every promise the client will make.
Frequently Asked Questions
What is a creative MCP workflow for a marketing agency?
It is a controlled way for an AI host to discover creative tools, inspect accepted inputs, quote work, start image jobs, check status, and return assets while the agency keeps brief, budget, and approval decisions human.
Can an agent create a whole campaign without review?
It can coordinate repetitive image work, but an agency should keep human approval before spend, before client review, and before delivery. Claims, product accuracy, brand fit, and usage rights need accountable owners.
How many ad variations should an agency generate?
Start from the testing question, not a round number. A 20-cell plan is useful only when each cell represents a controlled change and the media team can give every variant a fair test.
How should an agency prevent duplicate image jobs?
Keep the returned job handle, retry the status check rather than the start action, and reuse the same stable request identity when the service supports idempotent starts.
What should stay fixed across creative variations?
Usually the product, offer, audience, core claim, mandatory branding, placement, and measurement plan. Change one meaningful creative variable per test cell whenever practical.
Is Oakgen's public MCP available for agencies now?
No. The public production connection is not released. Oakgen's planned first surface is image-only, and this guide describes the intended workflow without claiming an endpoint, credentials, host compatibility, or launch date.
What can an agency use in Oakgen today?
Teams can currently use Oakgen's web-based image generator, agent chat, image arena, and UGC ad workflows. Those are current product workflows, not evidence that the planned public MCP has launched.
Sources and Further Reading
- Model Context Protocol: introduction — official protocol vocabulary for hosts, clients, servers, and tools.
- Meta photo ad guidance — official guidance on focal points, visual consistency, image quality, and previewing ads.
- Meta ad review guidance — official overview of the inputs Meta may review before an ad runs.
Ready to build the campaign now? Turn the approved brief into image concepts in Oakgen, then carry only the strongest directions into production.
