Architecture

Cloud AI is powered by mcp-service, a Spring AI MCP server. AI clients connect over SSE / Streamable HTTP (/stream). Keyword routing (api / pipeline / clickstream) selects the backend for each tool call.

Clients

Three client types talk to mcp-service:

Client	Role
Claude Desktop	Local MCP client over SSE / Streamable HTTP
claude.ai	Remote MCP
Cursor	IDE MCP client

mcp-service hub

The service sits in the center and orchestrates:

• dashboard-backend — REST with JWT + OAuth; backed by PostgreSQL (RDS) for auth and application context
• Redis cluster — Cache and RAG vectors
• OpenAI — Embeddings used for retrieval-augmented context
• EdgeTag — Script / tag overview and OAuth-related edge integration
• Kinesis Firehose — Redacted tool-usage logs streamed to S3 (analytics Parquet) for later Athena analysis

Data lake (Athena)

For pipeline and clickstream queries, mcp-service runs SQL against:

• Amazon Athena — Serverless SQL
• Amazon S3 — Data lake storage feeding Athena
• Glue Data Catalog / AWS Glue — Schema catalog and ETL for the lake

Ad & commerce APIs

For live channel and store data, mcp-service calls APIs directly over REST:

• Google Ads (Ads API)
• Meta (Graph API)
• TikTok (Marketing API)
• Shopify (Admin API)

Tool-call logging

Every tools/call is logged (redacted) → Kinesis Firehose → S3 → Athena. That audit path sits alongside keyword routing so operators can see which tools ran and which backend (API, pipeline, or clickstream) handled them.

Technology stack

• Spring Boot 3.3 / Spring AI (MCP starter), Java 17
• AWS SDK (Athena, S3, IAM, Firehose)
• Redis (Jedis) for cache and RAG vectors
• Dedicated Athena workgroup for MCP query cost isolation (see service config ATHENA_MCP_WORKGROUP_NAME)