Free test · 8 steps · about 10 minutes

AI SDLC for SAP

How to use AI in SAP ABAP teams, from Jira ticket to transport

In a 180-task ABAP benchmark, Claude solved 37% of tasks on the first try and 84% after five rounds of compiler and test feedback. How much of that your SAP team gets depends on its AI software development lifecycle. Answer 8 steps about your AI tools, IDE, SAP release and rules, and get an AI SDLC proposal that fits your environment: a 0 to 10 score for how much of your SAP development cycle AI can run, what to change first, the install steps for your setup and how your ABAP developers use each of 10 AI skills.

Hand-drawn timeline titled 10 SAP AI skills: Evaluation, Development, Testing and Deployment, with spec-grill and reverse-engineer at Evaluation, implement, refactor, cds-rewrite and bugfix at Development, test-case and code-review at Testing, and transport-release and documentation at Deployment.

No sign-up. Your proposal shows on this page.

What 10 out of 10 looks like

  1. A developer opens a Jira ticket and AI sharpens the specification with them.
  2. In one IDE, AI reads the live code from DEV, the Z tables and where-used.
  3. It writes the change, activates it, runs the syntax check, ATC and ABAP Unit, and fixes what failed.
  4. It assigns the change to a transport. A person reviews it and releases the transport; QAS and PRD follow your process.

Nothing is copied between windows. A person approves every step.

Step 1 of 8

raises your scorelimits it or needs checkingblocks part of the cycleno effect

Which AI tools to use in my SAP teams?

The tool decides how AI gets into the IDE and to the MCP server.

Which AI tool will your ABAP developers use?

The AI assistant your developers will work with. Claude Code and GitHub Copilot both work as agents: they read files, run tools and change code, not only chat.

Why it matters and what your answer changes

Both reach the ABAP MCP server, by different routes. Copilot runs inside Eclipse; Claude Code runs in a terminal next to it and, on ADT 3.62 or newer, reads and writes through SAP’s server with no add-on. The install guide and the skill scenarios change with your choice. “Not decided yet” is scored as Claude.

Claude Code
Anthropic’s coding agent. It runs in a terminal or VS Code, reads a whole project, runs commands and edits files. Included in Claude Team and Enterprise.
GitHub Copilot
GitHub’s coding assistant. It has an official Eclipse plugin with ABAP support and can use Claude, GPT and other models.
Is Claude Code switched on for your developers?

Claude Code is the part of Claude that works on code: it reads a whole project, runs commands and edits files. The chat in the Claude app is a different thing.

Why it matters and what your answer changes

Yes: AI works as an agent on your code and can connect to an MCP server. No: only copy-paste into the chat is left, and the score stays at 2 or 3 whatever else you answer.

Claude Code
Anthropic’s coding agent. It runs in a terminal or VS Code, reads a whole project, runs commands and edits files. Included in Claude Team and Enterprise.

Source: Claude plans

An admin turns it on in the Claude admin settings. Without it, only chat and copy-paste are left.

Is MCP allowed for Copilot in your GitHub organization?

MCP lets Copilot call outside tools, here SAP’s server. In a company GitHub account, an admin has to allow it. It is off by default.

Why it matters and what your answer changes

Yes: Copilot can read and change ABAP through SAP’s server and run checks. No: Copilot only sees the file open in the editor, with no live checks.

MCP (Model Context Protocol)
An open standard that lets an AI tool call tools in another system, such as “read this class” or “run the tests”. Without it, AI sees only what you paste.

Source: Copilot for Eclipse

It is off by default. An organization admin turns it on.

Ten AI skills across the SAP development lifecycle

An AI SDLC for SAP has one skill per stage of a change, from the Jira ticket to the transport. How much each skill does on its own depends on your score; the result spells out every skill step by step for your setup. All ten are free and open source as a Claude Code plugin and GitHub Copilot prompt files: github.com/marian-kamenistak/sap-ai-skills.

  1. 1Specifyspec-grillJira ticket to a specification scored 8/10+
  2. 2Understandreverse-engineerHow the existing code works
  3. 3BuildimplementCode, activation, checks
  4. 4Testtest-caseABAP Unit and QAS scenarios
  5. 5Reviewcode-reviewATC plus Clean ABAP and risks
  6. 6ImproverefactorCleaner code, same behaviour
  7. 7Modernizecds-rewriteData logic moved into CDS views
  8. 8DocumentdocumentationDocs for Confluence or Jira
  9. 9Releasetransport-releaseTransport, QAS and PRD checklist
  10. 10Run and fixbugfixDump to cause to fix, back to Specify
1 to 10: one change from Jira ticket to production. A bug found in production starts the cycle again at 1. A person approves every step.

Toolsets · checked 10 Nov 2026

8 AI toolsets for SAP teams

Your release, your IDE and your security rules decide which setup fits. Each drawing shows how the tools connect, from the developer to the DEV system. The first one gives AI the most of the cycle; the last one works when MCP is not allowed yet.

Not sure which one is yours? The test picks it from your answers.

  1. The full stack

    Most of the cycle
    Hand-drawn flow of the full AI SDLC stack for SAP: a Jira ticket goes to Claude Code, then to the SAP MCP server and SAP DEV, then ATC and ABAP Unit checks, and a person releases the transport.
    Step by step
    1. Jira ticketmcp-atlassian, read only
      MCP
    2. Claude Code + 10 SAP AI skillsin the Eclipse terminal
      MCP, localhost
    3. SAP ADT MCP server + arc1ADT 3.62+, arc1 adds where-used
      ADT
    4. SAP DEVAI writes only to one Z package
      runs
    5. ATC + ABAP UnitAI fixes the code from the results
      hands over
    6. A person releases the transportQAS and PRD follow your process

    Release 7.31 SP04 or newer. On 7.40, test one write on DEV first.

    Reads: yesWrites: yes

    Fits when you want AI from the Jira ticket to the transport, with a person approving each step.

  2. Claude Code or Codex in Eclipse

    Hand-drawn flow: a developer works with Claude Code or Codex inside Eclipse ADT, which talks to SAP's MCP server (ADT 3.62 or newer) and through it to SAP DEV.
    Step by step
    1. Developer
      prompt
    2. Claude Code or Codexin the Eclipse terminal
      MCP, localhost
    3. SAP ADT MCP serverADT 3.62+, Edit File tool on
      ADT
    4. SAP DEVAI writes only to one Z package

    Release 7.31 SP04 or newer, 7.40 included

    Reads: yesWrites: yes

    Codex connects to the same server: url and token in config.toml.

    Fits when you run an older release like 7.40, where community servers can’t write.

  3. GitHub Copilot inside Eclipse

    Hand-drawn flow: a developer chats with GitHub Copilot inside Eclipse ADT, Copilot calls SAP's MCP server in the same window, and the server reaches SAP DEV.
    Step by step
    1. Developer
      chat
    2. GitHub Copilot, agent modeinside Eclipse ADT
      MCP
    3. SAP ADT MCP serverMCP allowed by a GitHub org admin
      ADT
    4. SAP DEVAI writes only to one Z package

    Release 7.31 SP04 or newer

    Reads: yesWrites: partly

    agent edits on ABAP files still have open bugs

    Fits when your company already pays for Copilot and wants one window, no terminal.

  4. Community MCP server, no Eclipse

    Hand-drawn flow: a developer uses Claude Code or Codex in VS Code, which talks to a community MCP server and on to SAP DEV, with writes from release 7.51.
    Step by step
    1. Developer
      prompt
    2. Claude Code or Codexin a terminal or VS Code
      MCP
    3. Community MCP server on the laptopvibing-steampunk or mcp-abap-adt
      HTTPS, /sap/bc/adt
    4. SAP DEVthe developer's own SAP user

    Release 7.51 or newer to write, 7.50 reads

    Reads: yesWrites: yes

    Fits when you are on S/4HANA and your developers prefer VS Code to Eclipse.

  5. Central MCP server, run by IT

    Hand-drawn flow: three developer laptops connect over VPN with a token to one central MCP server, which keeps an IT log and reaches SAP DEV.
    Step by step
    1. Claude Code, Codex or Copiloton each developer laptop
      VPN, personal token
    2. One MCP server in your networkIT chooses the tools and keeps the log
      ADT
    3. SAP DEVwrites only where IT allows

    Release 7.51 or newer to write with a community server

    Reads: yesWrites: yes

    Fits when security wants one place to control and log what AI can do in SAP.

  6. Your own small MCP server

    Hand-drawn flow: a developer works with Claude Code, which calls your own MCP server with approved tools only, and the server reaches SAP DEV.
    Step by step
    1. Developer
      prompt
    2. Claude Code
      MCP
    3. Your MCP serveronly the tools you approve
      ADT
    4. SAP DEVAI writes only to one Z package

    Release Same as community servers: reads from 7.50, writes from 7.51

    Reads: yesWrites: yes

    Fits when you want full control of the tools and can spend a few days building it.

  7. SAP's own: Joule for Developers

    Hand-drawn flow: a developer uses Joule for Developers inside Eclipse ADT, paid in SAP AI Units, connected to the SAP system.
    Step by step
    1. Developer
      asks
    2. Joule for Developersinside ADT: completion, explanation, unit tests
      paid in SAP AI Units
    3. SAP systemSAP supports the MCP tools with this licence

    Release ABAP Cloud: S/4HANA and BTP ABAP, not NetWeaver 7.x

    Reads: yesWrites: partly

    suggests code, the developer accepts it

    Fits when you build on ABAP Cloud and want support from SAP.

  8. No MCP yet: abapGit files

    Start here
    Hand-drawn loop without MCP: code goes from SAP DEV by abapGit export into Git files, then to Claude, Codex or Copilot, and a person pastes it back into SAP DEV.
    Step by step
    1. SAP DEV
      abapGit export
    2. Files in a Git repository
      reads files
    3. Claude Code, Codex or Copilotin VS Code
      a person pastes it back
    4. SAP DEVa person runs ATC and the tests

    Release Any release abapGit runs on (7.02+)

    Reads: partlyWrites: no

    AI sees only the exported package

    Fits when security says no to MCP for now. Start here and show the results.

Tool list · checked 10 Nov 2026

Tools for an AI SDLC in SAP

What SAP teams use to run each step of the SDLC with AI, from the Jira ticket to production. Every link was checked on 10 November 2026.

ToolTypeWhat it does and where it helpsLink
Official SAP tools 13
ABAP Development Tools (ADT) for EclipseOfficial

SAP's Eclipse IDE for ABAP, where developers write, debug and transport code. It also runs SAP's MCP server.

SDLC step: Development

ADT for VS CodeOfficial

SAP's free ABAP extension for VS Code with the MCP server built in. It covers fewer object types than Eclipse so far.

SDLC step: Development

SAP ADT MCP serverOfficial

Lets an AI agent search, read, edit, test and transport ABAP through ADT, and read short dumps. From ADT 3.62 it reads and edits source for Claude Code too, and writes only to packages on its allow-list. SAP states support from release 7.31 SP04, ECC included.

SDLC step: DevelopmentTestingDeploymentProduction

Joule for Developers, ABAP AIOfficial

SAP's AI assistant inside ADT: code completion, code explanation and ABAP Unit test generation. Paid in SAP AI Units.

SDLC step: DevelopmentTesting

SAP-ABAP-1 (generative AI hub)Official

SAP's own ABAP model in the generative AI hub, tuned mainly to explain existing ABAP code.

SDLC step: Evaluation

gCTS (Git-enabled CTS)Official

Keeps ABAP changes in Git and moves them between systems, so each AI-made change has a history you can review.

SDLC step: Deployment

ABAP Test Cockpit (ATC)Official

SAP's built-in quality checks for ABAP. Run them on AI-written code before the transport is released.

SDLC step: TestingDeployment

ABAP UnitOfficial

The unit test framework built into ABAP. The AI writes the tests, ABAP Unit shows whether the code does what the ticket asked.

SDLC step: Testing

ABAP CleanerOpen source (SAP)

Applies 100+ cleanup rules with one keystroke in Eclipse or VS Code. Tidies AI-written code before review.

SDLC step: Development

Clean ABAP style guideOpen source (SAP)

SAP's style guide for ABAP. Turn it into the rules file your AI agent reads before it writes code.

SDLC step: Development

CAP MCP serverOpen source (SAP)

Lets an AI agent search your CAP model and the CAP documentation while it builds the service.

SDLC step: Development

UI5 MCP serverOpen source (SAP)

Gives an AI agent the UI5 API reference, UI5 guidelines and manifest checks when it builds a UI5 app.

SDLC step: Development

Fiori MCP serverOpen source (SAP)

Helps an AI agent create and change SAP Fiori elements apps from a prompt.

SDLC step: Development

AI coding assistants 4
Claude CodeCommercial

Anthropic's AI coding agent for the terminal and VS Code. For ABAP it connects to SAP's MCP server or a community one.

SDLC step: EvaluationDevelopmentTesting

GitHub Copilot for EclipseCommercial

Copilot inside Eclipse ADT: ABAP completion, chat, and agent mode with MCP. Agent edits on ABAP files still have open bugs.

SDLC step: Development

Claude Code for Eclipse (unofficial)Community

Unofficial plugin that brings Claude Code into Eclipse. It does not see open ADT editors yet.

SDLC step: Development

Nova IntelligenceCommercial

An agentic platform for SAP. It reads the custom code, processes and integrations, keeps the documentation current, and writes and tests code changes in the system.

SDLC step: EvaluationDevelopmentTesting

MCP servers and plugins 12
arc1 ADT MCP extensionCommunity

Adds read-only tools to SAP's ADT MCP server in Eclipse: where-used, package contents, the inactive-objects list, and source reading on ADT 3.60 and 3.61. Block its arc1_sap_http_post tool, which can post to any ADT endpoint.

SDLC step: EvaluationDevelopment

vibing-steampunkCommunity

Bridges ADT to MCP: read code, debug, run tests and read dumps. Writes need 7.51 or newer, or a small extension installed on the SAP side.

SDLC step: DevelopmentTestingProduction

mcp-abap-adt (mario-andreschak)Community

Sixteen read-only tools for ABAP code, DDIC and table contents. A low-risk first step, though table reads send business data to the model.

SDLC step: Evaluation

mcp-abap-abap-adt-apiCommunity

Over 120 ADT tools in one server: edit, activate, transport, debug and refactor ABAP from an AI agent.

SDLC step: DevelopmentDeployment

mcp-abap-adt (fr0ster)Community

Create, read, change and delete ABAP objects on ECC, S/4HANA and BTP ABAP Cloud. Needs BASIS 7.50 or newer.

SDLC step: Development

abap-adt-mcp (williansaez)Community

Read and write source, run unit tests and ATC, manage transports and read dumps, with guard rails set per system on the server.

SDLC step: DevelopmentTestingProduction

ABAP FS (vscode_abap_remote_fs)Community

VS Code extension that opens an ABAP system as files. It has AI tools for Copilot and an MCP server for Claude Code and other agents.

SDLC step: DevelopmentTesting

mcp-sap-docsCommunity

Gives an AI agent search over SAP Help and the ABAP keyword documentation, so its answers rest on the docs.

SDLC step: EvaluationDevelopment

SAP AI MCP servers (index)Community

A list of SAP MCP servers and SAP AI skills, kept up to date by the community. Start here to find more.

SDLC step: Evaluation

bw-modeling-mcpCommunity

Lets an AI agent read, create and change BW/4HANA objects such as aDSOs, transformations and DTPs. On BW 7.5 on HANA it reads transformations and DTPs only.

SDLC step: Development

mcp-sap-guiCommunity

Lets an AI agent work in SAP GUI for Windows: read screens, fill in fields and run transactions. Saves wait for your approval.

SDLC step: TestingProduction

hana-cli MCP serverOpen source (SAP)

SAP sample command line tool for HANA with an MCP server, so an agent can list tables and views and run queries.

SDLC step: Development

Quality and code in files 6
abapGitCommunity

Exports ABAP to Git as plain files, so an AI agent can read whole packages and you get a diff to review.

SDLC step: DevelopmentDeployment

abaplintCommunity

Static checks for ABAP in files that run in CI. A fast, repeatable gate on AI-written code.

SDLC step: Testing

SAP Enterprise Continuous Testing by TricentisCommercial

Codeless, model-based test automation for SAP and the apps around it. It generates test cases from a plain-language description, paid in SAP AI Units, and updates the tests when SAP changes.

SDLC step: Testing

PanayaCommercial

Shows what a change affects in custom code, transports and business processes. An agent fixes ABAP custom code with a person in the loop, and every fix is tested before go-live.

SDLC step: DevelopmentTestingDeployment

WorksoftCommercial

Codeless automation that tests SAP business processes end to end, in SAP GUI, Fiori and the systems around them. Worksoft Impact uses SAP change and usage data to pick the tests a change needs.

SDLC step: TestingDeployment

Int4 SuiteCommercial

Test agents for SAP Integration Suite. You describe a test in plain language, and the agent runs it with real messages captured from production.

SDLC step: Testing

Backlog and specs 4
mcp-atlassianCommunity

Connects an AI agent to Jira and Confluence, Cloud or Server/Data Center, so it reads the ticket and the spec.

SDLC step: Evaluation

calmcpCommunity

Connects an AI agent to SAP Cloud ALM: requirements, user stories, defects, test cases and analytics. It only reads until you switch writing on.

SDLC step: EvaluationTesting

Atlassian Rovo MCP serverCommercial

Atlassian's hosted MCP server for Jira and Confluence Cloud. The agent reads tickets and writes specs and pages.

SDLC step: Evaluation

Azure DevOps MCP serverCommercial

Microsoft's MCP server for Azure DevOps: work items, wiki, test plans, repos and pipelines.

SDLC step: EvaluationTestingDeployment

Skills 1
SAP AI skillsCommunity

10 skills that take an ABAP change from a Jira ticket to a released transport, for Claude Code and GitHub Copilot. People approve the spec, the diff and the transport.

SDLC step: EvaluationDevelopmentTestingDeployment

SAP AI workshop

Set it up with your team in one day

The AI for SAP teams workshop runs on your own code and backlog: up to 12 people for 2,400 EUR, or the whole IT team of up to 50 for 3,900 EUR, excluding VAT. It starts from files and moves to the ABAP MCP server where your IT approves it. The case behind it is an 18-person SAP team that cut median implementation time from 19 days to about 5.5.

The sample deck of a workshop day, 73 slides: install guide, four group guides with the exact prompts, one slide on how each skill works.

Guides

Why AI in SAP ABAP needs an MCP server

Without MCP, AI sees only the snippet someone pastes into it. It doesn't know your Z tables, includes or naming rules, and it never learns whether the code passed the syntax check, so it writes textbook ABAP that doesn't fit your system. With an ABAP MCP server, AI reads the live code, asks the system for DDIC and where-used, runs the checks itself and fixes its code from the results. That loop is where the gap between 37% and 84% comes from, measured on 180 ABAP tasks in the TH Köln benchmark and its extended model leaderboard (Claude Opus 4.6, April 2026).

37%of ABAP tasks solved on the first try
84%after five rounds of compiler and unit-test feedback

SAP's own MCP server, per SAP Help, runs inside Eclipse ABAP Development Tools 3.60 and newer, on localhost behind a token, and SAP states support down to release 7.31 SP04, on-premise included. From ADT 3.62 it reads and edits source for CLI clients such as Claude Code, and it writes only to the packages you put on its allow-list. Nothing is installed on SAP and AI works under the developer's own SAP user.

The best AI for ABAP: Claude Code or GitHub Copilot

Both connect to the same SAP server. GitHub Copilot for Eclipse runs inside ADT and reads ABAP through the IDE, and you can pick Claude models in it. Claude Code runs in the Eclipse terminal with your Claude licence. On ADT 3.62 or newer it reads and edits code through SAP's server too; on 3.60 and 3.61 it needs the community add-on arc1 to read and can't change existing objects through SAP's server. The test asks which one you use and builds the install guide for it.

When a local MCP server isn't allowed

  • A central MCP server inside your network, behind the firewall and reached over VPN. IT gets one place for rules and logs. Chat connectors that run in the cloud can't reach it; Claude Code on the laptop can.
  • Your own MCP server with only the tools you choose, for example read source, where-used, ATC and tests, and no table contents.
  • No MCP. Export code to files with abapGit and let AI work on them. Reading, checks and moving code back stay manual, and the score stops around 4 out of 10.

Why rewrite ABAP reports to CDS views

Many Z reports still read data the old way: a SELECT, then a loop with more SELECTs inside, joins and totals built in ABAP. A CDS view moves that work into the database. The cds-rewrite skill does the move with AI and proves the result stays the same.

  • Speed. HANA joins, filters and totals the data where it lives, so far less data travels to the ABAP server. Reports with SELECTs in loops are usually the biggest winners.
  • One definition, many users. The same view feeds the ABAP report, a Fiori app, an OData API and analytics.
  • Authorization in one place. A CDS access control declares who sees which rows once, instead of AUTHORITY-CHECKs scattered through programs.
  • Testable. CDS views have their own test double framework, so the rewrite comes with tests that prove the output.
  • Where SAP is going. RAP, ABAP Cloud and SAP's clean-core guidance all build on CDS views.

Not everything belongs in a view. Updates and complex procedural rules stay in ABAP, and a rewrite that doesn't make the report faster or simpler isn't worth shipping, which is why the skill measures runtime before and after. CDS view entities need release 7.55 or newer (S/4HANA 2020); from 7.40 SP05 to 7.54 the skill writes classic DDIC-based CDS views (DEFINE VIEW), and 7.31 has no CDS.

SAP and AI terms in plain English

For anyone answering the test who doesn't work in SAP every day. Each question in the test explains the terms it uses.

Claude Code
Anthropic’s coding agent. It runs in a terminal or VS Code, reads a whole project, runs commands and edits files. Included in Claude Team and Enterprise.
GitHub Copilot
GitHub’s coding assistant. It has an official Eclipse plugin with ABAP support and can use Claude, GPT and other models.
MCP (Model Context Protocol)
An open standard that lets an AI tool call tools in another system, such as “read this class” or “run the tests”. Without it, AI sees only what you paste.
ADT MCP server
SAP’s own MCP server, built into Eclipse ADT 3.60 and newer. From 3.62 it reads and edits source for CLI clients such as Claude Code, and writes only to the packages you allow. It runs on the developer’s laptop (localhost:2234) behind a token; SAP states support from release 7.31 SP04.
ADT (ABAP Development Tools)
SAP’s free plugin that turns Eclipse into the editor for ABAP. It is the modern way to develop on S/4HANA.
SAP GUI and SE80
SAP GUI is the classic desktop client for SAP. SE80 is its built-in development workbench, older than ADT.
SAP HANA Studio
An older Eclipse-based tool for HANA database work. SAP no longer develops it (SAP Note 2465027).
Release (SAP_BASIS)
The version of SAP’s technical platform, for example 7.50 or 7.58. It decides which language features and tools work.
SICF and /sap/bc/adt
SICF is the SAP transaction where web services are switched on and off. /sap/bc/adt is the group of services Eclipse ADT and every ADT-based MCP server use.
Basis
The SAP administration team: systems, users, settings and transport imports.
Transport
A package of changed objects that moves a change from development to test to production. It is the only way code reaches production.
DEV, QAS, PRD
The development, quality (test) and production systems. Code is written only in DEV and travels forward by transport.
ChaRM
Change Request Management in SAP Solution Manager: transports tied to an approved change document, with its own approvals.
Z objects
Your company’s own code. Names start with Z or Y, everything else is SAP standard.
CDS view
Core Data Services: a view defined as source code (define view entity …) that joins, filters and totals data inside the database. It is the base of Fiori apps, analytics and RAP on S/4HANA.
DDIC (ABAP Dictionary)
The definitions of tables, structures and data types in SAP. Structure, not data.
Short dump (ST22)
SAP’s crash report: the failing line, the call stack and variable values. Transaction ST22 lists them.
ATC (ABAP Test Cockpit)
SAP’s built-in quality gate that runs static code checks, often before a transport is released.
ABAP Unit
SAP’s built-in unit test framework. Test classes live next to the code they test.
abapGit
An open-source tool that exports ABAP objects to files and imports them back. It runs as one program inside SAP and works without internet in offline mode.
arc1
A community Eclipse plugin that adds read-only tools to SAP’s MCP server: where-used, package contents, the inactive-objects list, and source reading on ADT 3.60 and 3.61. Block its arc1_sap_http_post tool, which can post to any ADT endpoint.
Community MCP servers
Open-source MCP servers that talk to SAP’s ADT services directly, without Eclipse, for example vibing-steampunk or mcp-abap-adt. Some can read and write code; some can read table contents, which you should switch off.
Read-only Jira token
A personal access token with read rights only. AI can read tickets with it but can’t change them.
WSL
Windows Subsystem for Linux: Linux running inside Windows. Its own network makes localhost on Windows hard to reach.

Back to the test

FAQ

AI SDLC for SAP: common questions

What is an AI SDLC for SAP?

An AI SDLC for SAP is a way of building ABAP changes where AI agents work at every stage: specification, analysis, code, tests, review, documentation and transport, with a person approving each step. In practice it runs through an MCP server in Eclipse ADT and one AI skill per stage of the lifecycle.

How do you use AI in SAP ABAP development?

Give an AI coding agent such as Claude Code or GitHub Copilot access to your code and to the system’s checks. The strongest setup is SAP’s ADT MCP server in Eclipse: AI reads live code, writes changes in DEV, runs ATC and ABAP Unit and assigns transports, while a person approves every step.

What is the best AI for ABAP coding: Claude Code, Copilot or Joule?

In a public 180-task ABAP benchmark, general frontier models solved far more tasks than SAP’s own ABAP model: Claude Opus 4.6 reached 84% after five rounds of feedback, SAP-ABAP-1 about 20%. Claude Code and GitHub Copilot both connect to SAP’s MCP server; Joule for Developers runs on SAP AI units.

Does the ABAP MCP server work on-premise and with ECC?

Yes. SAP’s ADT MCP server ships in Eclipse ABAP Development Tools 3.60 and newer, and SAP states support down to release 7.31 SP04, on-premise and cloud, ECC included. It runs on the developer’s laptop, uses the ADT connection the developer already has, and needs nothing installed on the SAP server.

Can Claude or ChatGPT connect directly to an SAP system?

Claude Code can, through an MCP server: SAP’s own server in Eclipse, a community server that talks to the ADT services, or one your team builds. It runs on the developer’s laptop, so it reaches an on-premise system over your network. Chat connectors that run in a vendor’s cloud can’t reach a server behind your firewall.

Will AI replace SAP ABAP developers?

AI takes over the typing: boilerplate, copy-paste and first drafts of tests and documentation. It still needs someone who knows the business process and the system, and in a good setup a person approves every change before it reaches production. In one 18-person SAP team the same people delivered faster, from 19 to about 5.5 days median.

Can Claude Code connect to SAP’s ADT MCP server?

Yes. SAP’s ADT MCP server runs inside Eclipse ABAP Development Tools on the developer’s machine, at localhost:2234 behind a bearer token, and Claude Code connects to it with one claude mcp add command. From ADT 3.62 the server reads and edits source for Claude Code too; on 3.60 and 3.61 reading needs the community add-on arc1. SAP supports the tools only with a Joule for Developers licence.

Is there an open-source set of AI skills for ABAP?

Yes. The ten skills on this page, from spec-grill to bugfix, are open source under the MIT licence at github.com/marian-kamenistak/sap-ai-skills, as a Claude Code plugin and as GitHub Copilot prompt files. In Claude Code, install them with /plugin marketplace add marian-kamenistak/sap-ai-skills and then /plugin install sap-abap@sap-ai-skills.

Updated 10 November 2026.

Tell us what you need

No seat is booked yet. We come back with dates, price, and an honest read on whether this class fixes your actual problem. If it doesn't, we say so.

Which language?
Which classes interest you? optional

AI²

Bestseller

Masterclasses

The ladder

For companies

Your details reach the Academy team. No newsletter, no list, never sold.