Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableAWS Machine Learning Blog→ original

MoEngage Plugs AI Agents Into Its Existing APIs: Glance's Push CTR Up 38%

Before a sale, a marketer opens analytics, then a copy library, then a campaign builder, one after another. MoEngage did not rewrite hundreds of APIs. Instead, it automatically converted its OpenAPI descriptions into MCP tools that an AI agent calls from Claude or ChatGPT. At Glance, the click-through rate (CTR) on push notifications rose 38%, and campaigns launch 50% faster. If a product already has a mature API, it is easier to build the agent layer on top of it.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
MoEngage Plugs AI Agents Into Its Existing APIs: Glance's Push CTR Up 38%
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Key metric #1Workspace and user IDs are passed in headers, so the model never sees them
Key metric #2Paying per token instead of running its own GPUs cut AI costs by more than 50%
Key metric #3Manual work on channel, timing, and frequency dropped by 60–70%

To launch a weekend sale, a marketer opens three tools before writing a single line of copy: analytics, to see what worked last time; a content library, to pick a tone; and a campaign builder. Multiply that by email, push, SMS, and WhatsApp, and most of the week goes to switching between interfaces rather than making decisions. MoEngage, a customer communications platform used by more than 1350 brands, removed this routine with AI agents.

It did not rewrite its hundreds of existing APIs. Instead, it automatically converted their descriptions into "tools" under the MCP (Model Context Protocol) standard, which the agent calls on its own. The result at the customer Glance: the click-through rate on push notifications (CTR, the share of users who tapped a push) rose 38%.

For developers, this is a ready-made blueprint for adding agents to a mature product without touching the backend.

How the APIs became agent tools

MoEngage treated this as an architecture problem. The agent needed access to a customer's live data, and it had to be kept out of everyone else's. The platform's APIs for campaigns, analytics, segments, content, and workflows were already described in OpenAPI 3.0. The team used the FastMCP framework, which reads such a specification and generates MCP tools by itself. Each operation becomes a tool with a schema, a description, and parameter validation. In the example from the article, this takes about ten lines of Python: load the specification, create a client, call FastMCP.from_openapi, and start the server.

The tools fall into five groups: campaigns, analytics, segmentation, content, and automation workflows. The definitions are generated from the specifications, so when an API is updated, the tools are updated with it. The MCP server does not run a model itself. It only exposes the platform's capabilities, and the reasoning is done by whichever agent connects to it. Any MCP-compatible client can connect, including Claude and ChatGPT through the MoEngage connector.

How customers are isolated

In a system serving thousands of brands, the model must not decide whose data to read. MoEngage did three things. The server accepts the same JWT token as the regular APIs, and the AgentCore runtime validates it before the request reaches the server. Workspace and user identifiers travel in request headers rather than in tool arguments, so the model cannot see them and cannot spoof them. Each user session runs in its own isolated environment, and the state of one conversation cannot be read from another. On top of that sit Amazon Bedrock Guardrails with safety and personal-data filters, while an intermediate layer checks the structure of generated content before it is shown in the interface.

The agents can be kept on a short leash. Each one keeps a log: what data it pulled, what it decided, which channels it touched, and what it sent. The mode is set at the agent level: full autonomy, or human confirmation before an action. In the article's example, the agent finds a drop-off at the third onboarding step and pauses the SMS branch only after the user approves.

What came out of it, and the caveats

Merlin AI includes four types of agents: built-in assistants in the studio (copywriter, designer, segment builder, templates, workflows); conversational agents such as Campaign Insights; custom agents for CRM teams; and decisioning agents that pick the offer, channel, and timing for each customer individually. Reasoning and generation use Anthropic's Claude Sonnet models on Amazon Bedrock. The post names the versions as Sonnet 4.6, 5, and 5.5, and MoEngage switches between them through the same API.

The figures the company reports: more than six months in production; at Glance, campaigns launch 50% faster, push clicks are up 38%, and there are roughly three times as many content variants per campaign. Across the platform as a whole, manual work on channel, timing, and frequency fell by 60–70%, and goal-based conversion from decisioning rose by 20–30%. Pay-per-token inference instead of running its own GPUs cut AI infrastructure costs by more than 50%.

These numbers call for caution. The article was written together with the MoEngage team and published on the AWS blog, so it is a vendor's self-report. The measurement methodology is not disclosed: control groups and baseline values are unknown. The cost comparison with "in-house models" also depends on which models and what utilization were taken as the baseline.

The practical takeaway is simple. If your product has a mature API with a specification, you do not need to build the agent layer from scratch. What you need is MCP tool generation from OpenAPI, authorization on the same tokens, identifiers kept out of the model's reach, and a managed runtime for sessions. Separately, check that the AgentCore features you need are available in your AWS region.

Story

Key Questions Answered

How the APIs became agent tools?

MoEngage treated this as an architecture problem. The agent needed access to a customer's live data, and it had to be kept out of everyone else's. The platform's APIs for campaigns, analytics, segments, content, and workflows were already described in OpenAPI 3.0. The team used the FastMCP framework, which reads such a specification and generates MCP tools by itself. Each operation becomes a tool with a schema, a description, and parameter validation.

How customers are isolated?

In a system serving thousands of brands, the model must not decide whose data to read. MoEngage did three things. The server accepts the same JWT token as the regular APIs, and the AgentCore runtime validates it before the request reaches the server. Workspace and user identifiers travel in request headers rather than in tool arguments, so the model cannot see them and cannot spoof them. Each user session runs in its own isolated environment, and the state of one conversation cannot be read from another.

What came out of it, and the caveats?

Merlin AI includes four types of agents: built-in assistants in the studio (copywriter, designer, segment builder, templates, workflows); conversational agents such as Campaign Insights; custom agents for CRM teams; and decisioning agents that pick the offer, channel, and timing for each customer individually. Reasoning and generation use Anthropic's Claude Sonnet models on Amazon Bedrock. The post names the versions as Sonnet 4.6, 5, and 5.5, and MoEngage switches between them through the same API.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…