abap-mcp

Make your AI coding agent an expert SAP ABAP & RAP consultant — local by default. No SAP system, no credentials, one command to install.

abap-mcp is a Model Context Protocol server that gives any AI coding agent — Claude Code, GitHub Copilot, Cursor, Codex, Windsurf — real ABAP senses, built on abaplint (the open-source ABAP parser/linter). The agent brings the reasoning; abap-mcp brings ground truth: a deterministic parser, objective scores, and validated generators that keep the AI honest. It works on ABAP source wherever your agent works — a git checkout, an abapGit export, a code review, CI — long before anything reaches a system.

• New to ABAP or RAP? Your agent becomes the senior consultant looking over your shoulder:

every snippet linted against Clean ABAP as you write (lint_abap), every finding explained with its rationale and examples (explain_abap_rule), and scaffold_rap_bo starts you from a canonical, self-validated RAP business object instead of a blank editor.

• Senior ABAP consultant? It's your assessment and task engine: point your agent at a repo and

get a scored, categorized ABAP Cloud readiness report with an A–D tech-debt grade (check_cloud_readiness), a phased migration backlog with efforts and exit criteria (plan_cloud_migration), an objective before/after verdict on every rework (compare_abap), released-API replacements (check_released_api), and CI gates that hold the line while the migration proceeds.

The default stdio edition is 100% local: the analysis engine makes zero network calls and reads no user files. Sources go in as text and structured findings come back. The released-API list and abaplint rule data are package-bundled. An optional guarded Streamable HTTP edition makes the same tools available to remote clients such as ChatGPT web; in that mode, source text is sent to the machine you host. See privacy and data handling.

Why this exists

Every other ABAP MCP server is either a bridge to a live SAP system (ADT/RFC — needs credentials, a system, and trust) or a documentation search. But AI coding agents spend most of their time where the files are — editing abapGit repos, reviewing diffs, generating code — long before anything reaches a system. This server gives agents the missing feedback loop at that layer:

• "Does this ABAP parse? Is it clean? How does it perform?" → lint_abap (+ focus packs)
• "Fix this block automatically — casing, obsolete syntax." → fix_abap (deterministic, parser-guaranteed)
• "How far is this classic report from ABAP Cloud? Grade it." → check_cloud_readiness (A–D)
• "Plan the migration — what do we tackle first?" → plan_cloud_migration (phased backlog)
• "This class has no tests — start me a harness." → scaffold_abap_unit (failing-by-default)
• "What depends on what? Migrate which object first?" → get_object_dependencies (+ Mermaid)
• "Did this rework make the code better or worse?" → compare_abap
• "Is MARA a released API? What do I use instead?" → check_released_api
• "Start me a correct RAP business object." → scaffold_rap_bo
• "What's in this 4,000-line class? Draw it." → get_abap_outline (+ Mermaid)

Install — 60 seconds

The only requirement is Node.js 22+. No SAP system, no credentials, no API keys. New to AI tooling (or ABAP)? Follow the from-zero walkthrough in [docs/INSTALL.md](docs/INSTALL.md) — or let the installer do it:

npx abap-mcp setup     # detects VS Code / Claude Code and registers the server; guides Eclipse

Otherwise, pick your client:

Codex CLI · Codex IDE extension · ChatGPT desktop — one command configures all three on the same Codex host:

codex mcp add abap-mcp -- npx -y abap-mcp

Restart ChatGPT desktop or the IDE extension, then use /mcp to confirm abap-mcp is connected. In ChatGPT desktop you can alternatively open Settings → MCP servers → Add server, choose STDIO, and enter command npx with arguments -y abap-mcp.

Codex plugin (MCP server + review/mentor/migration skills)

codex plugin marketplace add palimkarakshay/abap-mcp
codex plugin add abap-mcp@abap-mcp

ChatGPT web cannot start a local npx process. It needs a deployed HTTPS /mcp endpoint or a secure development tunnel. This repository ships the HTTP entry point and hardening controls, but does not operate a public service. Follow the OpenAI setup and self-hosting guide.

Claude Code

claude mcp add abap-mcp -- npx -y abap-mcp

VS Code (Copilot agent mode) — one click on the Install in VS Code badge above, or one command:

code --add-mcp '{"name":"abap-mcp","command":"npx","args":["-y","abap-mcp"]}'

For a whole team, commit examples/vscode/mcp.json as .vscode/mcp.json in your abapGit repo — everyone who opens the folder gets the server offered automatically.

Eclipse (via the GitHub Copilot plugin) — Copilot Chat in Eclipse speaks MCP: open the Copilot menu → Edit preferences → MCP, and add the same "abap-mcp" server block (command npx, args ["-y", "abap-mcp"]). Steps and prerequisites: GitHub's MCP docs, Eclipse tab. Stock ADT without Copilot has no MCP client today.

Cursor · Windsurf · any MCP client — add this to its mcp.json / .mcp.json (project or global):

{
  "mcpServers": {
    "abap-mcp": { "command": "npx", "args": ["-y", "abap-mcp"] }
  }
}

Claude Desktop — Settings → Developer → Edit Config, add the same mcpServers block to claude_desktop_config.json, restart.

From a clone (contributing / hacking):

npm install && npm run build
claude mcp add abap-mcp -- node /path/to/abap-mcp/dist/cli.js

Connected? Ask your agent "list your ABAP tools" — you should see eighteen by default, lint_abap through check_rap_behavior, plus the opt-in run_abap_unit when ABAP_MCP_ENABLE_RUN=1 is set.

First things to ask

• "Here's my functional spec — build the RAP implementation from it." (or run /mcp__abap-mcp__abap-from-spec)
• "Lint this class against ABAP Cloud and explain the worst finding like I'm new to ABAP."
• "How cloud-ready is this repo? Grade it and plan the migration in phases."
• "Is MARA a released API? What do I use instead?"
• "Scaffold a RAP BO for entity Booking on table zbooking, draft enabled."
• "I reworked zcl_pricing — compare old vs new: did it actually get better?"

Guided workflows built in (MCP prompts)

The consultant's playbook ships both as four MCP prompts and as four Codex plugin skills. In Claude Code the prompts appear as slash commands (/mcp__abap-mcp__…); prompt-capable clients list them natively, while the Codex plugin discovers the corresponding skills:

CLI — same engine, no AI required

Every tool is also a subcommand, so it works in terminals and CI where no MCP client exists:

npx abap-mcp lint src/                          # lint files or whole directories
npx abap-mcp fix src/ --write                   # apply abaplint's deterministic auto-fixes in place
npx abap-mcp lint src/ --focus Performance      # themed pass: Performance | Security | Styleguide
npx abap-mcp lint src/ --rules-file org.json    # your org's abaplint rule pack, same engine
npx abap-mcp readiness src/ --fail-below 80     # repo-level ABAP Cloud readiness, scored + graded A–D
npx abap-mcp plan src/                          # phased migration backlog: work items, S/M/L efforts, exit criteria
npx abap-mcp unittest src/ --out tests/         # failing-by-default ABAP Unit skeletons for untested classes
npx abap-mcp deps src/ --mermaid                # dependency graph (+released-API flags) as a Mermaid diagram
npx abap-mcp compare old/ new/                  # rework verdict: findings resolved/introduced, grade movement
npx abap-mcp scaffold --entity Travel --table ztravel --key travel_id --out ./out
npx abap-mcp outline src/zcl_monster.clas.abap  # navigate big objects (--mermaid for a diagram)
npx abap-mcp released MARA I_Product --edition btp  # released-API status + successor, per edition (s4hc|btp|pce)
npx abap-mcp explain exit_or_check              # rule rationale
npx abap-mcp release --since 2605 rap           # bundled ABAP Cloud/RAP release-delta knowledge (dated, sourced)
npx abap-mcp knowledge "clean core level C"     # search the bundled SAP knowledge base (Clean Core, ATC, SAP-AI)
npx abap-mcp aisdk --scenario ZDEMO_AI --interaction string --out ./out  # scaffold a Generative AI Hub call
npx abap-mcp agent-rules --target Cloud --paired sap-adt-mcp --run       # print the AGENTS.md rules block
npx abap-mcp rapcheck src/ --release 2508       # check RAP behavior/service definitions (BDEF/SRVD); exit 1 on errors
npx abap-mcp unittest --run src/                # EXECUTE ABAP Unit tests offline (open-abap kernel, no DB/CDS/EML); exit 1 on failure

Directories are walked recursively (abapGit naming), batched automatically, and readiness merges batches into one scored, categorized repo report. Exit codes are CI-friendly (1 on error findings / failed threshold).

Agentic workflows, recipes & CI

[docs/examples/abap2xlsx-assessment.md](docs/examples/abap2xlsx-assessment.md) — what all of this produces on a real, well-known open-source repo (100 files, three commands, no system). [docs/COOKBOOK.md](docs/COOKBOOK.md) — practical recipes: the fix-until-clean loop, PR review without a transport, whole-repo migration triage, CI gates, per-persona use cases. [examples/claude-code/](examples/claude-code/) — drop-in agentic workflows that turn the tools into that over-the-shoulder consultant: an abap-code-reviewer subagent, an abap-cloud-migrator sweep loop (readiness score as the loop condition), an abap-tech-debt-assessor (the graded A–D client deliverable), per-repo .mcp.json, and a GitHub Actions quality gate for abapGit repos.

Agentic loop

The full offline loop an agent can run before anything reaches a system:

1. Scaffold — scaffold_rap_bo, scaffold_abap_unit, or scaffold_abap_ai_sdk for a

validated starting point instead of a blank editor.

1. Fix, then lint — fix_abap (deterministic mechanical pass) followed by lint_abap until

findings are zero or consciously waived.

1. Run — run_abap_unit (opt-in via ABAP_MCP_ENABLE_RUN=1, or always through

abap-mcp unittest --run) executes the ABAP Unit tests and feeds real pass/fail results back into the loop — TH Köln's ABAP benchmark (arXiv 2601.15188) measured this alone moving Claude Opus 4.5 from 31.6% to 78.7% task success across feedback rounds.

1. Readiness — check_cloud_readiness / plan_cloud_migration to grade the result and size

what's left; compare_abap proves the rework didn't regress.

1. Hand off — pair with SAP's official ADT MCP server or abap-adt-mcp for the in-system half

(write, activate, ATC, a real ABAP Unit run) — see Related projects.

Tools

Knowledge base & resources

A dated, cited SAP knowledge bundle ships with the package (curated 2026-09-10): 100 ABAP Cloud / RAP release-delta rows (2502–2608 trains + ABAP Platform 2025, plus pre-2502 chronology corrections so an agent can't present an established feature as a 2025-26 novelty), 55 Clean Core entries (Levels A–D, release contracts C0–C3, the real ATC vocabulary), and 7 SAP-AI decision cards (SAP-ABAP-1, Generative AI Hub orchestration, the ABAP AI SDK, Joule for Developers, SAP's official ADT MCP server, …). Every row is abap-mcp's own original summary over a cited source with a confirmed / reported confidence flag — never copied SAP prose (see the licensing boundary in docs/DESIGN.md).

Reach it as tools (explain_abap_release, search_sap_knowledge; CLI release / knowledge) or as MCP resources a host can attach directly to a conversation: abap-mcp://knowledge/manifest (provenance for every bundled file), .../release-deltas, .../clean-core, .../sap-ai (whole files), and .../release/{id}, .../clean-core/{id}, .../sap-ai/{id} (individual cards, each returned as both application/json and text/markdown).

Honesty box — what this is not

• Not ATC. The objective readiness score is still language-level: statements ABAP Cloud

removed. Released-API coverage is now partial, offline, and edition-aware: check_released_api and the releasedApiFindings in readiness reflect SAP's published Cloudification lists as of the bundled snapshot date — one per SAP edition (s4hc Public Edition, btp, pce Private Edition / on-prem; edition param, default s4hc) — they cover tables and function modules referenced in your source, not every API, and are only as current as the snapshot. A target system's own ATC check ("Usage of Released APIs (Cloudification Repository)" for Public Edition / "Usage of APIs (Cloudification Repository)" for Private Edition & on-prem, via check variants such as ABAP_CLEAN_CORE_DEVELOPMENT / ABAP_CLEAN_CORE_READINESS) remains authoritative; treat an "absent from the list" result as "not released as of the snapshot", not as proof. The readiness report's grade is our own blocker-density banding (gradeMeaning: "blocker-density"), not SAP's Clean Core Level A–D — see cleanCoreVocabulary on the report for the real ATC vocabulary (src/data/atc-vocabulary.json).

• Scaffold validation is tiered. Generated classes and CDS views are round-tripped through

abaplint at Cloud level before they're returned (the generator and the linter share one parser). Behavior and service definitions are outside abaplint's checked surface, so they run through abap-mcp's own RAP checker instead and are labeled validated: "rap-checker" when it returns nothing at error or warning severity; the metadata extension has no checker at all and stays "template". Each generated file is labeled validated: "abaplint" | "rap-checker" | "template", and ADT activation is still the final arbiter.

• The RAP checker is our own grammar, not SAP's parser. check_rap_behavior (and the rap/…

findings merged into lint_abap) reads BDEF/SRVD with a tokenizer and grammar this project wrote from SAP's published RAP BDL feature tables, the ABAP keyword documentation, and a 102-file corpus of Apache-2.0 SAP sample sources. Every report carries grammarVersion, rulesVersion and a scopeNote saying what that does and does not prove: it cannot see DDIC tables, behavior-pool classes or CDS field types, and constructs it does not recognise are reported as info, never as errors. Rules derived from anything weaker than SAP documentation are capped at warning and carry their provenance in the message. Every shipped rule is documented in docs/RAP-RULES.md.

• `run_abap_unit` is not SAP's kernel. It transpiles to JavaScript and executes on the

open-abap kernel (opt-in, ABAP_MCP_ENABLE_RUN=1; always on in the CLI via unittest --run) — no database (any ABAP SQL aborts the method), no CDS, no EML/RAP runtime, no AMDP, no authority checks, no locks, no ATC, no activation. A green run is evidence about pure logic, never proof of system behavior; constructs it cannot execute come back in unsupported, never as a silent pass.

• `scaffold_abap_ai_sdk` is checked against abap-mcp's own stubs, not SAP's real API. Generated

classes are round-tripped through abaplint against bundled IF_AIC_* / CL_AIC_* declarations this project wrote from SAP's documentation, not SAP source — labeled validated: "abaplint-syntax", meaning the ABAP parses, not that it matches SAP's actual ABAP AI SDK signatures exactly. The manual ISLM setup it depends on (SAP_COM_0A69, INTS/INTM) is listed in setupSteps, never performed.

• Text-in only, by design. The analysis engine does no user-filesystem walking or outbound

network access; it parses strings you explicitly pass. With stdio those strings are handled by the local child process. With Streamable HTTP they travel to the endpoint operator, so use HTTPS, authentication, and the documented privacy controls. For whole directories, use the CLI below, abaplint in CI, or the mcp-kit wrap-abaplint recipe.

Develop

npm install
npm run check     # typecheck + tests + build + routing eval — the CI gate
node dist/cli.js  # stdio MCP server
npm run start:http # guarded Streamable HTTP on http://127.0.0.1:3000/mcp
npx @modelcontextprotocol/inspector --cli node dist/cli.js --method tools/list

Tool and prompt descriptions are CI-graded (a rubric test enforces verb-first names, when-to-use, non-goals, described params, worked examples — the mcp-kit discipline).

Design

The decision log — why offline, why abaplint, why a dual-parse readiness diff, why the scaffolder validates its own output, what was deliberately left out — lives in docs/DESIGN.md.

Online companion: abap-mcp-genai

A separate, opt-in binary — abap-mcp-genai, never imported by the default server — wraps SAP's Generative AI Hub orchestration service using your own SAP AI Core tenant: explain_with_sap_abap_1 (ask SAP's own ABAP-tuned model to explain a class, method or snippet) and list_genai_hub_models. Your ABAP source leaves this machine when you use it — it goes to SAP infrastructure under your organization's own SAP AI Core contract, never to abap-mcp's authors. Read docs/GENAI.md in full before enabling it.

Related projects

• abap-kit by Furkan Coşgun (MIT) — a CLI that

scaffolds complete offline ABAP projects (abapGit layout, abaplint + transpiler config, the open-abap runtime) and can execute ABAP Unit tests locally via the abaplint transpiler. Complementary by design: abap-kit builds and runs the project harness; abap-mcp gives your AI agent the analysis, planning and scaffolding tools inside it — abap-kit create + abap-mcp setup is a complete offline ABAP+AI workspace. Its transpile-and-run pipeline previewed what run_abap_unit (v0.11) now ships natively in this server, on the same abaplint transpiler + open-abap-core foundation.

• SAP's official ADT MCP server

(GA, ships inside ADT for Eclipse and VS Code; 20 tools across 8 toolsets, 18 licence-free) — writes, activates and tests ABAP in a real system (abap_creation-create_object, abap_activate_objects, abap_run_unit_tests, abap_run_atc, …), with no offline lint, readiness grade or released-API cross-check of its own.

• abap-adt-mcp by Willian Saez (MIT) — the

community equivalent, 173 online tools over the same ADT REST API, same live-system trust model.

• Pairing. abap-mcp verifies offline, before and after: lint_abap / check_cloud_readiness /

compare_abap on code before it is written to a system, then again on what SAP's official server or abap-adt-mcp just wrote, activated or ran ATC against. Neither online server needs abap-mcp, and abap-mcp never needs credentials — see the recipe in docs/COOKBOOK.md and get_abap_agent_rules for the rule text to paste into a repo's AGENTS.md.

Credits

• abaplint by Lars Hvam — the parser and rule engine

underneath every tool here (MIT).

• SAP/abap-atc-cr-cv-s4hc — SAP's official ABAP

Cloudification Repository (object release list), Apache-2.0. The bundled released-API snapshot (src/data/released-apis.json, refreshed with each weekly release — tool output carries its snapshotDate) is a compact transform of that data, redistributed under Apache-2.0 with attribution; see docs/DESIGN.md and scripts/build-released-api-index.mjs for the pipeline.

• mcp-kit — the production-MCP patterns this server

follows (typed tool specs, transport discipline, description lint).

MIT © Akshay Palimkar. Not affiliated with or endorsed by SAP SE. "SAP", "ABAP" and "RAP" are trademarks of SAP SE; this is an independent open-source tool for developers working with them.