Your AI assistant is brilliant — and handless
Claude can explain ABAP Cloud clean-core rules better than most documentation. It can draft a RAP behavior definition from memory.
But ask it to do something in your DEV system — create the class, run the unit test, show you the ATC findings — and it hits a wall. It can talk about your SAP system all day. It can't touch it.
MCP (Model Context Protocol) is the fix: an open standard that works like a USB port for AI. Plug in an MCP server, and the assistant gains that server's abilities — reading systems, creating objects, running tests.
The SAP MCP ecosystem exploded in 2026. Here's the landscape as of this month, and how to actually use it.
The MCP servers that matter
abap-adt-mcp — the full ABAP workbench for AI
The headline act. abap-adt-mcp (open source, MIT) gives AI agents full ABAP development capabilities through SAP's ADT HTTP services — 173 tools in the 2.0 release, with nothing to install inside SAP.
The killer demo: type "In DEV, create class ZCL_INVOICE_CHECK in package ZSANDBOX, add a unit test, run it, and show me any ATC findings" — and watch the AI search the system, pick the transport, create the class, run the tests, and report back with real results.
Install is one npm command (Node.js 22.12+). Works with Claude Desktop, Claude Code, Cursor, VS Code — anything that speaks MCP.
SAP ABAP MCP — the governance-first option
If 173 tools make your security team nervous, sap-abap-mcp takes the opposite approach: governance first. Runs locally, keeps SAP credentials in your OS keychain.
Standout feature: it blocks writes to any profile classified as production. Dev and QA can be fenced into package allowlists. For team rollouts, start here.
abap-mcp — no SAP system required
Built on abaplint (the open-source ABAP parser), so it works on ABAP source anywhere — a git checkout, an abapGit export, a code review, CI. No system, no credentials, one install command.
Every snippet linted against Clean ABAP as you write; RAP business objects scaffolded from validated generators. Perfect for learning and CI pipelines.
@ui5/mcp-server — for the frontend side
SAPUI5 has its own: @ui5/mcp-server with 9 specialized tools — app scaffolding (freestyle, Fiori Elements, CAP full-stack), integration cards, project inspection, linting.
SAP's official ADT MCP server
SAP ships an official ADT MCP server inside its ADT clients. For SAP-supported workflows, start there; the community servers above are the power tools.
Setting it up: the 10-minute version
- Install —
npxthe server you picked. Nothing to install inside SAP; it talks to standard ADT HTTP services. - Connect — register it in your MCP client, add your SAP connection (host, client, user). Credentials stay in your local keychain.
- Scope it — point at DEV first. Package allowlists on (default
$TMPin stricter servers). Read-only mode until you trust it. - Try the loop — create a small class, write a unit test, activate, run it, summarize ATC findings. Watch every tool call before approving.
Guardrails: how not to get fired
- Never point an AI agent at production. DEV first, QA maybe, PRD never.
- Approve tool calls. Keep "ask before executing" on — especially for writes and transports.
- Respect the SAP API policy. These are dev tools, not a backdoor to business data.
- Review generated code like a junior's pull request. Fast and confident is exactly when mistakes slip through. ATC + ABAP Unit on everything.
What to actually use it for
The highest-value uses: ATC sweeps across a package with fix suggestions. Unit test generation for legacy classes that have none. RAP scaffolding from a data model sketch. Code reviews against Clean ABAP before transport release.
MCP turns the AI assistant from a brilliant explainer into a junior developer with infinite patience. Still needs supervision — and still the biggest productivity shift in the SAP workflow this year.
Tried any of these? Hit a snag with the ADT setup? Drop a comment — happy to troubleshoot.