MoEngage introduced a new layer that lets large language model agents call the platform’s own APIs through a secure, tenant‑scoped MCP connector, hosted on Amazon Bedrock AgentCore. For engineers, the change means the same API surface that powers UI and external integrations can now be invoked by autonomous agents, eliminating the need to duplicate business logic for AI use cases.
Agentic Customer Engagement Architecture
The core of the solution is the MoEngage Model Context Protocol (MCP) Connector, which presents existing segment, campaign, and content APIs as MCP‑compatible tools. Agents run on the AgentCore runtime, which provides hosting, tool‑call planning, and model execution. All reasoning and generation use Anthropic Claude models (Sonnet 4.6, 5, and 5.5) available in Amazon Bedrock. By keeping the connector tenant‑scoped, each brand’s data remains isolated while agents operate on live data.
Four Agent Modes Implemented in Merlin AI
Merlin AI defines four distinct agent categories, each built on the same foundation:
- Embedded agents – Integrated into the Merlin AI Studio UI to perform focused tasks such as copy generation, banner design, segment creation, Jinja/Liquid markup conversion, in‑app template building, and flow construction.
- Conversational agents – Available across campaign, analytics, segment, and flow pages; they interpret natural‑language queries, invoke the appropriate MCP tools, and synthesize multi‑metric analyses.
- Custom agents – Configurable by CRM teams to run continuously on MoEngage data, logging each data pull, decision, and channel interaction. Execution can be fully autonomous or gated by human review.
- Decisioning agents – Operate per‑customer, selecting content, timing, frequency, and channel based on goals and guardrails defined by product and marketing teams.
Implementation and Operational Implications
Because agents call existing APIs rather than new services, the deployment pipeline remains largely unchanged. Engineers must provision AgentCore runtime environments, configure model versions, and ensure the MCP connector enforces tenant boundaries. Operational monitoring should include:
- Model latency and cost metrics from Bedrock.
- AgentCore health checks and scaling policies.
- Audit logs from the MCP connector that record which tenant data each agent accessed.
Security considerations focus on the tenant‑scoped nature of the connector. While the source does not detail specific IAM policies, practitioners should treat the connector as a privileged data‑access layer and enforce least‑privilege principles for any service accounts that host agents. Isolation is achieved at the application level; therefore, additional network segmentation or VPC controls may be advisable to limit blast radius.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Engineers can now expose legacy platform capabilities to LLM‑driven workflows without rewriting business logic, but they must treat the MCP connector as a critical security boundary and instrument comprehensive logging. Monitoring model usage and AgentCore health becomes part of the operational stack, and any tenant‑level data leakage risk must be mitigated through strict access controls on the connector. The pattern demonstrated by MoEngage—wrapping existing APIs as MCP tools and hosting agents on Bedrock AgentCore—offers a reusable blueprint for other SaaS platforms seeking to add agentic features while preserving data isolation.

