Using Claude Code and AGENTS.md with SAP ABAP

By ABAP for Agents · Updated 22 October 2026

Claude Code works in your terminal or editor and reads instruction files from the project folder at the start of each session. For ABAP that's useful, because the things the agent can't guess (your release, your namespace, which objects exist) can be written down once. The same file can serve Codex, Cursor and GitHub Copilot too, if you use AGENTS.md.

AGENTS.md: one file for every agent

AGENTS.md is an open convention: a Markdown file in the repository root with instructions for coding agents. No frontmatter, no special syntax. Several tools read it:

• OpenAI Codex and other agents that follow the convention
• Cursor, in the project root and in subdirectories
• GitHub Copilot, where the nearest AGENTS.md in the directory tree takes precedence
• Claude Code, which can read it directly in recent versions, or through an import from CLAUDE.md

If a team uses more than one tool, keeping the ABAP rules in AGENTS.md means one file to maintain instead of four.

The CLAUDE.md import

Claude Code's own file is CLAUDE.md. According to Claude Code's documentation, when a CLAUDE.md exists in your working directory or above it, Claude reads that instead of AGENTS.md by default. The simple, version-independent setup is a CLAUDE.md that imports AGENTS.md with the @ syntax and adds anything Claude-specific below it:

@AGENTS.md

## Claude Code
- Task skills live in .claude/skills/: abap-clean-core-review.
- Prefer plan mode for changes that touch more than one ABAP object.

This is the exact CLAUDE.md in our free lite rules. The imported file is loaded at launch together with CLAUDE.md.

What to put in AGENTS.md for ABAP

Claude Code's documentation recommends keeping each CLAUDE.md under 200 lines, because longer files consume more context and reduce adherence. Imports don't change that: an imported AGENTS.md counts too. So focus on what the agent can't work out from the code:

# ABAP instructions

## Target
- Language version: Standard ABAP. Product: S/4HANA 2023 (SAP_BASIS 7.58).
- Never mix syntax from different releases in one answer.

## SAP objects
- Never invent function modules, BAPIs, classes, CDS views or tables.
- Name every SAP object you rely on and flag the ones to verify.

## ABAP SQL
- Escape host variables with @. Guard FOR ALL ENTRIES with
  IF itab IS NOT INITIAL and select all key fields.

## Self-check before answering
- Table expressions don't set sy-subrc. Use line_exists( ) or OPTIONAL.
- BAPI errors are in the RETURN table. No COMMIT WORK in RAP handlers.
- No !=, ++ or "text" literals. Every statement ends with a period.

Replace the target line with your own system. If you don't know the release of every system the code will run on, tell the agent to assume the lowest plausible one.

Skills for longer procedures

Some instructions only matter for one kind of task, such as a code review against Clean ABAP and the clean core levels A to D. Loading that checklist into every session wastes context. Claude Code skills solve this: each skill is a folder in .claude/skills/ with a SKILL.md file, and its description tells Claude when to load it.

---
name: abap-clean-core-review
description: Review ABAP code for Clean ABAP, ABAP Cloud readiness and
  clean core level (A-D). Use when the user asks for a code review.
---
# Task: Clean ABAP + clean core review
Return a table: # | Severity | Line/object | Finding | Why | Fix.
1. Correctness: unguarded FOR ALL ENTRIES, SELECT SINGLE without the
   full key, COMMIT WORK in RAP or in loops, BAPI RETURN not read.
2. Clean core level of every SAP object used. Mark unknown ones
   "verify in Cloudification Repository". Do not guess.

Ask "review this class for clean core" and Claude pulls the skill in.

Path-scoped rules (optional)

If your repository mixes ABAP with other code, Claude Code also supports .claude/rules/*.md files with a paths field in the frontmatter. A rule with paths only loads when Claude works with matching files:

---
paths:
  - "src/**/*.abap"
---
# ABAP SQL rules
- Escape every host variable with @ ...

Rules without paths load at launch, like CLAUDE.md. Rules in .claude/rules/ are specific to Claude Code, so other agents won't see them.

Check that the files loaded

Run /context in a Claude Code session and look under memory files, or open /memory, which lists the instruction files Claude Code knows about. If your AGENTS.md doesn't seem to be read, the usual cause is a CLAUDE.md on the path that doesn't import it.

Then test with a release-specific prompt:

Write ABAP for SAP_BASIS 7.40 SP05 that reads ZORDERS
for a list of keys.

With the rules active, the answer should state the release, declare the target table instead of using inline @DATA in the SQL (which needs 7.40 SP08), and guard the FOR ALL ENTRIES.

Connecting to the SAP system

SAP's ABAP development tools for VS Code include an ADT MCP server that external agents like Claude Code can connect to. Per SAP, it's off by default, needs an explicit system allow-list, and is read-only unless you enable editing. If you connect it, say how the agent should use it:

- If an ABAP MCP server is connected, use it to look up objects,
  signatures and release state, and to run the syntax check, ATC
  and ABAP Unit before answering.
- Read-only by default: never use write tools without explicit consent.

That turns "never invent objects" from a hope into something the agent can actually check. It still doesn't replace your own review: keep ATC and ABAP Unit as the gate before anything is transported.

For the mistakes these files should cover, see ABAP mistakes AI coding assistants make. Using Cursor or Copilot as well? See the Cursor rules guide and the Copilot instructions guide.

More guides