SAP Business AI Agent Suite

A small, self-contained demo of a Joule-style multi-agent architecture built on a real, standalone Model Context Protocol (MCP) server, grounded by a lightweight RAG knowledge base, and designed around SAP's Clean Core principles throughout.

This is a portfolio project (mock/generic data only — see Clean Core & data below).

Why this exists

This repo focuses specifically on agentic AI architecture and ABAP AI / Clean Core concepts — the areas most relevant to SAP Joule / Business AI extension roles — as a complement to two earlier repos in the same portfolio:

• SAP CPI & BTP Integration Examples — core Integration Suite / iFlow skills.

• IntelliFleet — SAP BTP + CAP + Fiori + Business AI end-to-end application.

Repo #3 goes one level deeper: instead of embedding "MCP-style" tool calls inside one app, it ships a real standalone MCP server any MCP-compliant client could connect to, plus a small multi-agent orchestrator that composes specialized skills the way SAP Joule composes skills internally.

Architecture

                     ┌─────────────────────┐
   user query  ───▶  │   Orchestrator       │  routes by intent
                     │   (planner agent)     │  (rule-based, or Claude Haiku
                     └──────────┬────────────┘   if ANTHROPIC_API_KEY is set)
                                │
        ┌───────────────────────┼───────────────────────┐
        ▼                       ▼                        ▼
┌───────────────┐     ┌───────────────────┐     ┌────────────────┐
│ Diagnostics    │     │ Documentation      │     │ Action          │
│ agent          │     │ agent              │     │ agent           │
│ (reads live    │     │ (RAG over mock SAP │     │ (drafts via MCP │
│  device/order/ │     │  help content —    │     │  tool; never    │
│  incident data)│     │  BTP/CPI/ABAP AI)  │     │  writes back)   │
└───────┬────────┘     └────────────────────┘     └────────┬────────┘
        │                                                    │
        └───────────────────┬────────────────────────────────┘
                             ▼
                  ┌─────────────────────┐
                  │  MCP server (stdio)  │  get_device_status, get_customer,
                  │  src/mcp-server/     │  get_order, list_open_incidents,
                  └──────────────────────┘  draft_incident_note

1. Standalone MCP server (src/mcp-server/)

A real Model Context Protocol server built with @modelcontextprotocol/sdk, run as its own process over stdio — not tool functions embedded inside one app. It exposes five tools over mock data (src/mcp-server/data/mock-data.ts):

Any MCP-compliant client can connect to this server independently of the agents in this repo (e.g. point Claude Desktop's MCP config at src/mcp-server/server.ts via tsx).

2. RAG knowledge base (src/rag/)

A small curated set of mock SAP-help-style markdown docs covering BTP, CPI/ Integration Suite, ABAP AI & Clean Core, and Joule-style agent architecture (src/rag/knowledge-base/). Retrieval (src/rag/retriever.ts) uses a dependency-free TF-IDF + cosine-similarity search — fully offline, no embedding API calls. The documentation agent grounds its answers in the retrieved excerpts, and (optionally) asks Claude to synthesize a short answer from them; with no ANTHROPIC_API_KEY, it returns the excerpts directly.

3. Multi-agent orchestration (src/agents/)

• orchestrator.ts — the planner. Routes a query to one of three specialized agents. Uses deterministic keyword rules by default; if ANTHROPIC_API_KEY is set, uses Claude Haiku for more flexible natural- language routing instead (same dual-mode pattern used in IntelliFleet).

• diagnosticsAgent.ts — calls MCP read tools to answer operational questions ("what's the status of DEV-001?").

• documentationAgent.ts — answers conceptual questions via RAG.

• actionAgent.ts — drafts a note via the MCP draft_incident_note tool.

Clean Core & data

Every tool the agents can call is the mock equivalent of a released SAP API (a released OData service / wrapped BAPI / event API) — nothing in this repo reads or writes a simulated core table directly, and draft_incident_note only produces draft text; it never persists anything. See src/rag/knowledge-base/abap-clean-core.md for the underlying principle this mirrors.

All data (src/mcp-server/data/mock-data.ts, the knowledge-base docs) is generic/mock. No real Cognizant/Electrolux client code, data, or IP appears anywhere in this repo.

Running it

npm install
npm run cli

Runs three example queries end-to-end (spawns the MCP server as a subprocess, routes each through the orchestrator). Pass your own query:

npm run cli -- "What is the status of DEV-002?"

By default this runs fully offline/deterministic — no external API calls, no cost. To enable Claude-assisted routing and documentation synthesis:

cp .env.example .env   # then fill in ANTHROPIC_API_KEY

To run the MCP server standalone (e.g. to point another MCP client at it):

npm run mcp-server

Tests

npm test

Runs offline against the real MCP server subprocess (test/mcp-server.test.ts), the retriever (test/retriever.test.ts), and the orchestrator's routing (test/orchestrator.test.ts) — no API key required.

Tech stack

Node.js + TypeScript, @modelcontextprotocol/sdk, @anthropic-ai/sdk (optional, dual-model Haiku/Sonnet cost-conscious pattern), a dependency-free in-process TF-IDF retriever, zod for tool schemas, tsx for running TypeScript directly, Node's built-in test runner.