Amazon's MCP server lets agencies build a DSP campaign in under an hour

Amazon Ads opened the public beta of its MCP server in early February, an interface that lets AI agents run actions directly on the ad platform through natural language instructions. The goal is to simplify automation of tasks such as campaign creation, settings management and reporting access.

The server is built on Model Context Protocol, an open standard that lets AI agents connect to software and drive its features. It acts as an interpretation layer between AI systems and the Amazon Ads APIs.

Instructions written in natural language get translated into structured API calls. Agents can then run actions on the platform without building specific integrations or stacking point-to-point connections.

One integration is enough to connect custom agents or AI platforms such as Claude, ChatGPT or Gemini. Once wired to the MCP server, those agents can reach a range of Amazon Ads features: creating or editing campaigns, pulling reports, managing account settings, accessing billing data.

The service is paid. Amazon hands out credits to experiment, then charges for server usage at scale. Two agencies, iProspect (dentsu) and Jellyfish, agreed to share early feedback after several weeks of testing.

iProspect started testing as soon as the beta went public. The objective was to explore what the tool can do, with no preconceptions. "There's often a gap between the marketing promise and the operational reality. So the point was to build the most objective learning curve possible," said Dominique Latourelle, associate director adtech solutions at iProspect.

The team simply followed the public MCP documentation and used Kiro, the AWS interface built on Claude, which supports two kinds of interaction. On one side, natural language prompts to talk to the system directly. On the other, a more advanced command line interface where you can insert code, plug in agents, or automatically turn a request into executable technical instructions such as audits or command generation.

The first step was working out what could actually be administered through MCP. "We questioned the machine to identify the actions available through the write APIs: creating or editing campaigns, accessing reports, usable dimensions, billing data," Latourelle said.

The point was to determine whether MCP expands what was already possible or simply replaces the existing APIs. "We quickly realized a significant number of operations were already possible," Latourelle said.

Technically, there are two routes. "You can use Kiro directly, or connect an internal AI agent built on a known LLM that uses Kiro as a technical bridge to query the Amazon Ads servers, whether for DSP or Sponsored Ads," Latourelle said.

That second route is useful for handling documentation or sensitive data while keeping an in-house tool. "The idea would be an internal system that can route requests to MCP-compatible platforms, to save time," he added.

In its tests, iProspect built a full campaign end to end without actually launching it. The point was to compare the time required against a manual API setup.

At this stage, the productivity gain is barely visible. Two things temper that. The technology has been public for four weeks. And since this is still an experiment, Latourelle's team ran repeated back-and-forths to understand error codes and check everything worked as expected.

"We're not trying to move fast: every step gets double-checked. Once the mechanics are fully understood and made reliable, the potential to scale could be significant," Latourelle said.

The next step is testing in live conditions, on two or three campaigns.

Latourelle noted the experiments were far simpler on Sponsored Ads than on Amazon DSP, which suggests Amazon is following Google's pattern of favoring Google Ads over DV360 when rolling out new features.

He also noted that Amazon Marketing Cloud remains out of scope for now. MCP does not yet allow direct use of features such as custom audience creation, overlap analysis between two campaigns, or reach and frequency studies.

"That would be a real game changer," Latourelle said. An AMC-native feature already generates SQL from a prompt. "It works fine for light or exploratory analysis. But as soon as the queries get more complex or heavier, the limits show up fast."

Longer term, the value could be in using these tools to query past studies more easily, compare results over time and centralize analysis into a form of client intelligence hub.

"Late last week, we set up our first Amazon DSP campaign through the MCP server," said Alexandre Barré, adtech and partnerships director at Jellyfish. The campaign was supporting the opening of a Monoprix store in Nice.

"At 5 p.m. on Thursday, the ad sales arm we'd booked pulled out: the reserved inventory had gone to another advertiser. So we had to find a new one fast to keep the campaign alive," Barré said.

The process starts with an internal prompter built on Pencil, which turns the client brief into a structured brief for the agents. That brief goes to a media-and-sales-house agent, which analyzes Jellyfish's display campaign history and identifies relevant ad sales arms. Its recommendation: French press ad sales group 366, through Nice Matin and Monaco Matin.

Barré then went into Claire, 366's new platform, to create the three deals that would carry the campaign. "Once the deals were created, I interacted with Amazon DSP through MCP, which handled the entire campaign setup," he said.

Amazon's Kiro agent automatically generated the campaign structure: flight dates, total and daily budget, device targeting, audiences (Monoprix affinity, upper-income, urban), bidding strategy, and a CPM aligned with the deals.

"In total, barely 50 minutes passed between receiving the brief and getting the campaign live," Barré said. He pointed out that despite a background as a trader, he had not configured a campaign in a DSP in close to six years.

The whole reporting phase, with the campaign starting this week, will also run through MCP. "That could let us increase how often we deliver reports," Barré said.

The use case stays relatively simple: a display campaign in a single city. The value shows up immediately on wider activations. "If the campaign had covered 10 or 20 cities, setup time would have stayed almost identical, where a manual setup would easily have taken several hours," Barré said.

The test covers display for now, but Amazon MCP also supports video campaigns. Jellyfish's retail team says tests are possible on Sponsored Ads too, where the potential is particularly large given the granularity and volume of campaigns.

The point is to cut the time spent on low-value operational work. Beyond the time saved, these tools also open the way to more consistent, standardized measurement built on shared reference points.