APIs and MCPs: Why Retail Media Needs Both

The API versus MCP debate keeps surfacing in retail media conversations, and it’s usually framed as a choice. It isn’t. Each does a different job, and dropping either one breaks the model. Here’s how we think about it, and why retail media networks should stop waiting on this decision.

The API is the data layer

The API is how we move retail media data into our warehouse and build the long view our customers need: performance over time, budget pacing, one report across every channel. That role isn’t changing.

The API can’t keep up on its own

Retail media networks are big and move fast, with new ad products and betas coming constantly. Every new endpoint is work on our side. We map it, test it, ship it, and customers see it months later. With APIs alone, our speed is capped by a network’s release schedule plus our backlog. We’re constantly playing catch-up.

A good MCP changes that. It describes its own tools at runtime, so our agent finds a new capability and uses it with no code change from us. You ship a feature to the MCP, our agent uses it the same day. That’s fast value for our mutual customer.

When to use each

These aren’t competing options. They solve different problems.

Use the API when you need cross-channel reporting, historical performance, or budget pacing across retail media networks and publishers. That’s warehouse-level work: structured, predictable data you’re analyzing over time. Recommendations are a good example. You’re pulling data, modeling it, surfacing an insight. That’s an API job, and it saves cost and time. Ask an agent to run models over a large data set and you get long execution times and results that can vary from run to run. The API lets another service do the heavy lifting and hand the agent the right recommendation at the right time.

Use the MCP when you want to implement it once and have new capabilities surface automatically; no rebuild or code change on our end or yours. A new ad product drops, the agent picks it up the same day. It’s also how we handle complex, multi-step agent workflows, like end-to-end campaign setup, where the agent needs to understand what tools exist at runtime, not just execute a fixed sequence of calls.

At Skai, we built our data foundation on the API, and we’re building our agent layer on the MCP. Both are in production. Both matter.

A good MCP, not a wrapper

The value disappears if the MCP just wraps one tool per endpoint. We’ve measured open-source retail media MCPs that load tens of thousands of tokens of definitions before the agent does anything useful. That fills the context window and makes it slower and wrong more often. A good MCP means:

• Semantic tools with real error handling, not one-to-one endpoint mirrors
• Runtime discovery, so new ad products appear to the agent on their own
• A catalog that stays compact as coverage grows
• Safe write paths for bids, budgets, and campaigns, built for human approval
• User-scoped OAuth, so every action traces back to a person, not a shared key
• Parity with the API over time across all supported ad products

Good for retail media networks, and why partners should move now

Fragmented integrations force networks to carry the support load for every individual platform. An MCP model changes that by providing one integration surface that allows any agent, including future ones, to access into your capabilities immediately.

We see partners taking a wait-and-see approach, betting smarter LLMs will eventually make MCPs unnecessary. That misreads the problem. The bottleneck isn’t intelligence. It’s the operational drag and context window exhaustion that comes with unstructured access. Even today’s elite models stumble over poorly built protocols, while a clean MCP makes them far more precise. That difference in performance is only going to grow.

Governance is the other piece. An MCP gives organizations a deterministic harness: the ability to gate specific capabilities and cleanly separate read-only insights from write-heavy operations. Direct API or CLI access alone leaves your controls fragile and your surface area exposed.

Advertisers are already looking for agent-ready infrastructure. Moving now puts the foundation in place before the market shifts. Waiting means a full rebuild when the pressure is highest.

Getting started is simpler than it looks. Start with read-only tools that answer common reporting questions, then layer in write paths built for human oversight. 

We’ve already deployed this architecture at Skai, and we’re ready to help you do the same. Ready to learn more? Contact Skai today.

---

Frequently Asked Questions