ABAP AI Code

ABAP AI Code

An agentic AI coding assistant that runs inside the SAP system. Written in ABAP with an HTML UI, it reads, explains, reviews and — with the optional write add-on — creates, changes and deletes ABAP repository objects. No code export, no laptop toolchain, no ADT required.

Demos: Z_ABAP_AI_CODE2 walkthrough · Combobulating

---

Table of contents

• 1. Why this exists — comparison with Claude Code / Codex
  • 1.1 Side-by-side
  • 1.2 When to use which
• 2. What you get
• 3. Repository layout and the plugin pattern
• 4. Installation
• 5. Configuration
  • 5.1 Providers (ZAICODE_PROVIDER)
  • 5.2 API keys (Z_ABAP_AI_KEYS)
  • 5.3 Prompt / schema folder
• 6. First run
  • 6.1 Selection screen
  • 6.2 The three-pane window
• 7. Tool catalog
• 8. Writing your own tool
• 9. Logs and observability
• 10. Security model
• 11. Known limitations
• 12. Legacy AGENTS flow
• 13. Links, TODO, license

---

1. Why this exists — comparison with Claude Code / Codex

Claude Code, Codex CLI and friends are excellent — if your code is on a laptop, in Git, and you are allowed to send it to a vendor from there. In a lot of SAP shops none of those three is true: the code lives only in the system, ADT/abapGit may not be installed, and exporting sources to a developer machine is a compliance conversation nobody wants to have.

ABAP AI Code is that class of tool, re-implemented where the ABAP code actually is. You start a report in SAP GUI, type a question in your own language, and an agentic loop (LLM → tool call → ABAP execution → result → LLM → …) works on live repository objects.

1.1 Side-by-side

	Claude Code / Codex CLI	ABAP AI Code
Where it runs	Your laptop / a container	Inside the SAP system, SAP GUI report Z_ABAP_AI_CODE2
What it sees	Files in a Git working copy	Live repository objects (READ REPORT, SEO_*, TADIR) — always the current version
How it changes code	Writes files; you commit	Writes via standard SAP APIs (RPY_*, SEO_*), with syntax check, transport entry (RS_CORR_INSERT) and activation
Reviewing a change	git diff in a terminal	Built-in HTML diff with per-hunk approve / decline, decline notes and an AI discussion thread per hunk
Prerequisites	Node.js, Git, code exported to disk	An SAP system. Nothing else — no ADT, no Eclipse, no Git working copy
Auth / secrets	Vendor account or API key on the laptop	API key AES-256 encrypted per SAP user in ZAICODE_APIKEY, unlocked by a password that is never stored
Model choice	Vendor's own models	8 preconfigured providers (Anthropic, OpenAI, Mistral, Gemini, Groq, Cerebras, OpenRouter, NVIDIA); model list pulled live from /v1/models
Extending it	MCP servers, hooks — all outside SAP	ABAP plugin classes implementing ZIF_AI_TOOL, discovered automatically; they run in the system with full ABAP API access
Write access	Always on	Opt-in: the base package is read-only; write tools ship as a separate package you install only where changes are intended
ABAP awareness	Generic code model	ABAP-native: class/method-level reads, structure tree, McCabe + Halstead metrics, breakpoints, "run this program"
Data leaving the system	The whole working tree is reachable	Only what a tool explicitly sends: one object's source, one method, one diff

1.2 When to use which

• Use Claude Code / Codex when your ABAP is already mirrored to Git and your policy allows a cloud agent to read that repository. The tooling around it is richer.
• Use ABAP AI Code when the code must not leave the system, when developers work in SAP GUI, when there is no abapGit mirror, or when you want AI assistance bound to SAP authorizations, transports and user identity.
• They are not exclusive. Nothing here prevents you from keeping an abapGit mirror and running both.

---

2. What you get

Concrete capabilities, from a user's point of view:

Ask about code in plain language. "What does ZCL_X do?", "Where is the bug in Z_REPORT?", "Make me a table of the methods and what they do." Answers come back in your language — the answer language follows the question language.

Read at the right granularity. A whole class, a single method (ZCL_X=>METHOD), a program, several objects in one call (comma-separated), or everything matching a pattern (ZCL_AITOOL_*). Over-broad requests ask you for confirmation instead of quietly reading half the system.

Get a real code review. review_sap_code runs a dedicated reviewer prompt (~15 KB of ABAP-specific review rules) against the live source and returns findings — not a generic "looks fine to me".

See the object, not just the text. Answers that resolve to one object open in a real ABAP editor pane with syntax highlighting, a structure tree (sections / attributes / methods, or events / forms / modules), breakpoints you can set, and a "Run" button that starts the program via its selection screen.

Measure it. One click gives McCabe cyclomatic complexity and full Halstead metrics (η1/η2, volume, difficulty, effort, estimated bugs) per method and per class.

Approve changes hunk by hunk. When the write add-on is installed, a change never lands silently: you get a diff, you approve or decline each hunk, you can leave a note or argue with the AI about a specific hunk, and only then is it saved, transported and activated.

Keep the receipts. Every LLM request, response, thinking block, tool call and tool result is written to a per-session log folder, plus a per-question token table and a Claude-Code-compatible session.jsonl.

Bring your own model. Switch provider and model on the selection screen; the model dropdown is filled from the provider's live /v1/models, so a new model is available the day it ships. Anthropic extended thinking is supported with a configurable budget.

---

3. Repository layout and the plugin pattern

The platform is split into two abapGit repositories along a single line: can it change your code?

Repository	Package contents	Install it when
ABAP-AI-Code (this repo)	Engine, agentic runner, UI, diff/review cluster, metrics, crypto, key management, and the read-only tools (read_sap_object, review_sap_code, show_code_example)	Always — it is the base
ABAP-AI-CODE-TOOLS	The write tools (create_sap_object, modify_sap_object, delete_sap_object) and ZCL_CODE_OBJECT_SAVER, which performs the actual write / delete / activate	Only on systems where AI-assisted code changes are intended

This split is possible because every tool is a plugin:

Layer 1  <tool_name>.json   OpenAI function definition — the contract the LLM sees
Layer 2  <tool_name>.md     prompt fragment — behaviour rules for that tool
Layer 3  ZCL_AITOOL_<X>     ABAP class implementing ZIF_AI_TOOL — does the work

ZCL_AI_TOOL_FACTORY discovers implementers at runtime by querying SEOMETAREL for classes that implement ZIF_AI_TOOL or inherit from ZCL_AITOOL_BASE. Both the tools JSON array and the tool section of the system prompt are assembled from that same registry, so:

• installing a package adds its tools to the LLM's toolset and to the prompt — no core code is touched;
• not installing it means the LLM is never told the capability exists, so it cannot attempt it and then fail;
• the base has no compile-time dependency on the write package: the runner and the UI call the saver purely by dynamic name (CALL METHOD ('ZCL_CODE_OBJECT_SAVER')=>…), and write paths fail gracefully with "read-only platform" when it is absent.

Registration is fail-closed: a tool whose .json or .md companion file is missing from the prompt folder is not registered, and a message names the missing file. A forgotten file surfaces as an explicit warning instead of a tool that silently misbehaves.

Design rationale and the full envelope specification: FABLE_NEW_ARCHITECTURE.md.

---

4. Installation

Prerequisites

• abapGit — to pull the repositories into the system.
• An API key for at least one supported provider.
• HTTPS connectivity from the SAP system to the provider, configured in SM59. The setup is described step by step in this SAP Community post.

Note: AVE / ABAP Code Reviewer is no longer a required dependency. The diff/review cluster it used to provide now lives in this repository (ZCL_CODE_ACR_*, ZCL_CODE_POPUP_DIFF). Install AVE only if you want it as a standalone tool.

Steps

1. abapGit → Online → clone https://github.com/ysichov/ABAP-AI-Code into a Z package. Pull and activate.
2. (Optional, write access) clone https://github.com/ysichov/ABAP-AI-CODE-TOOLS into a second package. abapGit does not resolve dependencies, so the base must be installed first.
3. Copy the TOOLS folder of this repository to the machine that runs SAP GUI — e.g. C:\soft\GITHUB\ABAP-AI-CODE\TOOLS. These .json / .md files are read from the frontend at runtime; they are deliberately not abapGit objects, so prompts can be edited in a text editor without a transport.
4. If you installed the write add-on, copy its companion files (create_sap_object.*, modify_sap_object.*, delete_sap_object.*) into the same folder.
5. Run Z_ABAP_AI_CODE2 once and either point p_tools at your folder and save a variant, or change the default in the program's INITIALIZATION block.

---

5. Configuration

5.1 Providers (ZAICODE_PROVIDER)

Providers are pure customizing — table ZAICODE_PROVIDER, maintainable in SM30/SE16. On first run of Z_ABAP_AI_KEYS it is seeded with:

Provider	Base URL	Wire format
ANTHROPIC	https://api.anthropic.com/v1	Anthropic (ANTHROPIC flag set)
OPENAI	https://api.openai.com/v1	OpenAI (CACHEKEY flag set)
MISTRAL	https://api.mistral.ai/v1	OpenAI-compatible
GEMINI	https://generativelanguage.googleapis.com/v1beta/openai	OpenAI-compatible
GROQ	https://api.groq.com/openai/v1	OpenAI-compatible
CEREBRAS	https://api.cerebras.ai/v1	OpenAI-compatible
OPENROUTER	https://openrouter.ai/api/v1	OpenAI-compatible
NVIDIA	https://integrate.api.nvidia.com/v1	OpenAI-compatible

Three fields drive behaviour: URL (base endpoint), ANTHROPIC (speak the Anthropic wire format instead of OpenAI), CACHEKEY (provider accepts OpenAI's prompt_cache_key). Adding any other OpenAI-compatible provider is one table row — no code change.

5.2 API keys (Z_ABAP_AI_KEYS)

Run Z_ABAP_AI_KEYS to store a key:

• Provider / Name — several named keys per provider are supported (e.g. DEFAULT, TEAM).
• Key — the provider API key.
• Password (twice) — minimum 10 characters. This password is never stored anywhere.

The key is AES-256 encrypted into ZAICODE_APIKEY, with the AES key derived from the password and sy-uname. See Security model.

5.3 Prompt / schema folder

p_tools on the selection screen points at the frontend folder holding <tool_name>.json and <tool_name>.md. It also holds two files the runner reads directly:

• tool_runner.md — the orchestrator's own system prompt (routing rules: read before modify, delete-vs-modify, read-only vs write, answer in the user's language…). Per-tool fragments are appended to it by the factory.
• delete_misuse_redirect.md — the template used to bounce a delete_sap_object call back to modify_sap_object when the user only asked to remove part of an object.

Edit any of these in a text editor and the change takes effect on the next run. This is the main tuning surface of the whole system.

---

6. First run

6.1 Selection screen

Z_ABAP_AI_CODE2 opens with only the password visible. Enter it and press Enter — that unlocks the rest of the screen. The staged reveal is deliberate: without the password nothing can be decrypted, so nothing else is worth showing.

Field	Meaning
p_pwd	Password for your stored key. Masked. Mandatory
p_prov	Provider. Only providers you have a key for are listed; a single-key user is snapped straight to theirs
p_name	Which of your named keys to use (auto-selected when you have exactly one)
p_model	Model. Filled live from the provider's /v1/models, refreshed only when provider / key / password changes. Defaults to a haiku model when available
p_tools	Frontend folder with the .json / .md tool files
p_temp	Temperature (also changeable at runtime from the toolbar)
p_maxt / p_nomax	Max output tokens, or no limit
p_think / p_thbud	Anthropic extended thinking and its token budget (hidden for other providers)
p_log	Folder for session logs
p_wipe	Clear the password from memory right after the key is decrypted. Leave on in PROD; switch off on a dev box to avoid retyping

6.2 The three-pane window

• Left — question. A text editor for your prompt, and the toolbar: Ask, History (all messages of the session), Temp (change temperature live), New session (clear the conversation).
• Middle — progress + structure. A live step log: each LLM call and each tool call with its label, elapsed seconds and token counts (in / out / cached). Below it, the structure tree of the object currently open; double-click a node to jump to it in the editor.
• Right — answer + code. Either the rendered HTML answer, or a real ABAP editor showing the object. In editor mode the toolbar gains Back (return to the list), Metrics (McCabe + Halstead) and, for a program, Run. Breakpoints can be set directly.

When a tool produces a change, this pane switches to a diff: each hunk carries approve / decline buttons, a note field and an AI thread so you can question a specific hunk before deciding. Nothing is written until you approve.

Under the hood, run() is a bounded agentic loop — up to 8 iterations of LLM → tool calls → execute → feed results back, ending when the model answers with no tool calls. Identical idempotent calls (the same read or review twice) are served from a per-question cache instead of being re-executed.

---

7. Tool catalog

Read-only tools, shipped in this repository:

Tool	Class	What it does
read_sap_object	ZCL_AITOOL_READ	Reads PROG / CLAS / METH / FUNC sources. Accepts ZCL_X=>METHOD, comma-separated lists and ZCL_* wildcards. Asks for confirmation above 10 objects; reports "not found" as a hard error instead of letting the model hallucinate
review_sap_code	ZCL_AITOOL_REVIEW	Reads the object itself and runs it through the reviewer prompt. Also the tool used for "find the bug / explain what's wrong" — diagnosis without modification
show_code_example	ZCL_AITOOL_EXAMPLE	Generates an ABAP example for a topic. Display only — never saved, never reviewed

Write tools, shipped in ABAP-AI-CODE-TOOLS:

Tool	Class	What it does
create_sap_object	ZCL_AITOOL_CREATE	Creates a program or class that does not exist yet
modify_sap_object	ZCL_AITOOL_MODIFY	Changes an existing object — including removing a method, a FORM or comments
delete_sap_object	ZCL_AITOOL_DELETE	Deletes a whole object. Guarded: a request to remove part of an object is redirected to modify_sap_object

All three route through ZCL_CODE_OBJECT_SAVER, which does the syntax check, the transport entry and the activation.

---

8. Writing your own tool

Four steps, no core changes:

1. Schema — TOOLS/my_tool.json, a standard OpenAI function definition:

   {
     "type": "function",
     "function": {
       "name": "my_tool",
       "description": "What it does and when the model should reach for it.",
       "parameters": {
         "type": "object",
         "properties": { "object_name": { "type": "string", "description": "…" } },
         "required": ["object_name"]
       }
     }
   }

2. Prompt fragment — TOOLS/my_tool.md, the behavioural rules appended to the system prompt when (and only when) this tool is installed.

3. ABAP class — inherit from ZCL_AITOOL_BASE and redefine two methods:

   CLASS zcl_aitool_mine DEFINITION PUBLIC INHERITING FROM zcl_aitool_base CREATE PUBLIC.
     PUBLIC SECTION.
       CONSTANTS c_tool_name TYPE string VALUE 'my_tool'.
       METHODS zif_ai_tool~get_tool_name REDEFINITION.
       METHODS zif_ai_tool~execute       REDEFINITION.
   ENDCLASS.

   execute receives the raw JSON arguments; get_json_attribute( ) from the base pulls out a flat attribute. Fill rs_result: xml_payload (what goes back to the model), error_text on failure, and — for a tool that changes code — save_required, object_type, object_name, original_source, final_source, so the runner triggers the diff / approve / save flow.

4. Activate. The factory finds the class on the next run. Nothing to register.

Two conventions worth remembering:

• Tool name = schema file name = md file name. c_tool_name is the single binding point between all three layers.
• Never call the LLM client directly. Use mo_context->ask( ). It routes the sub-call through the runner, so it appears in the progress panel, is logged and is counted in the token totals. mo_llm->ask( ) bypasses all of that and makes the tool invisible.

To disable a tool without deleting it, redefine zif_ai_tool~is_active to return abap_false.

---

9. Logs and observability

With p_log set, each session gets its own folder <log_path>/<date>_<time>/ containing, per step:

Suffix	Content
Q	The raw request sent to the provider (JSON) and the assembled prompt (TXT)
LLM	The raw response, plus a readable Markdown rendering
THINK	The extended-thinking block, when enabled
TOOL	Tool call arguments and results
A	The final answer
SUBQ / SUBLLM	Requests / responses of tool sub-agent calls

Plus two aggregates:

• _tokens.md — one row per question: input / output / cached tokens.
• claude_compatible/session.jsonl — the same session as Claude-Code-style JSONL, one JSON object per line, so a Claude session reader ingests it directly.

The middle pane shows the same information live while the loop runs.

---

10. Security model

API keys. Stored AES-256 encrypted in ZAICODE_APIKEY. The AES key is derived from the password and sy-uname:

• the password is never persisted — not in the database, not in the source. Someone reading the table or this repository cannot decrypt anything;
• the username is embedded in the plaintext and verified on decrypt, so a copied row cannot be replayed under another user;
• format v2 (current) uses a random per-record salt and a random IV with stronger key stretching. Legacy v1 rows still decrypt; re-saving a key upgrades it.

Password handling in the UI. The cached model-list state stores a per-session salted SHA-256 token of the password, never the password itself, so neither a debugger nor a dump reveals it. With p_wipe on, the password is cleared from memory as soon as the key is decrypted — the blast radius of a debug session shrinks to one key rather than the master password.

Write access. Read-only by default. The write package is a separate installation decision, and even where it is installed nothing is saved without an explicit per-hunk approval, a syntax check and a normal transport entry.

What leaves the system. Only what a tool explicitly sends to the provider: the source of the object being read, reviewed or changed, plus your prompt. There is no background indexing and no repository-wide upload.

---

11. Known limitations

Stated plainly, so nothing surprises you:

• Navigation and the structure tree use plain text search, not real ABAP parsing — good enough to jump around, not a substitute for ADT.
• The agentic loop is capped at 8 iterations per question; very large multi-object tasks need to be split.
• Prompt / schema files live on the frontend, so each developer needs a local copy (this is also what makes them editable without a transport).
• The UI is SAP GUI + CL_GUI_HTML_VIEWER; it is a desktop-SAP-GUI tool, not a Fiori app.
• There is no MCP ecosystem — extensions are ABAP classes.
• Answer quality depends heavily on the .md prompt files. They are meant to be tuned.
• tool_runner.md mentions a web_search tool that has no implementation in this repository; treat that line as aspirational until a plugin provides it.
• The write add-on ships delete_sap_object.json but no delete_sap_object.md — under the fail-closed rule that means ZCL_AITOOL_DELETE refuses to register until the prompt fragment is added.

---

12. Legacy AGENTS flow

The original implementation used string proto-tags ({AGENT:CODE_SEARCH PROG Z_TEST}) and a chain of specialised agents, driven from the AGENTS folder (Object_detector.md, task_orchestrator.md, class_processor.md, code_review.md, final.md, …).

It still runs, in parallel with the new flow: copy the AGENTS folder locally and point the legacy program's path at it. New work should use Z_ABAP_AI_CODE2 and the TOOLS folder — the reasons are laid out in FABLE_NEW_ARCHITECTURE.md.

---

13. Links, TODO, license

• Write add-on: https://github.com/ysichov/ABAP-AI-CODE-TOOLS
• Architecture design doc: FABLE_NEW_ARCHITECTURE.md
• Roadmap: TODO.md
• SM59 / connectivity setup: SAP Community — ABAP Code Reviewer
• Standalone code reviewer (optional): https://github.com/ysichov/AVE
• License: MIT