claude-for-abap

MCP server giving Claude (and any MCP-compatible client) live access to SAP systems via ADT.

Read source, search the repository, run syntax checks, run unit tests, run ATC, diff the same object across landscapes, edit and activate ABAP — all from a chat window or an autonomous agent. No add-on installation on the SAP stack required.

Why

SAP development is full of repetitive read-the-source / check-the-callers / diff-the-system work. AI assistants are great at exactly that kind of task — but only if they can reach the system. ADT (ABAP Development Tools) is the HTTP API that Eclipse uses; it ships with every modern NetWeaver and S/4 system. This server speaks ADT on behalf of the agent so the agent can do real work against your real systems, with the same auth and scoping you'd give a developer in Eclipse.

What's in the box

27 high-level tools wrapped around the most common ADT endpoints, plus a generic escape hatch for anything else, plus 5 user-invokable Clean Core prompts that turn the tool surface into outcome-shaped slash commands (see Clean Core prompts below).

¹ Experimental tools target ADT endpoints (SNOTE, SM12 enqueues, SM36/SP01) that classic NetWeaver does not expose; on such systems they return available:false with a fall-back hint rather than failing. They work where the backing service exists (typically S/4HANA).

Multi-system aware. One config, many SAP systems (DEV / QAS / PRD or landscape-wide); switch with the system argument or compare across two with adt_compare_source / adt_transport_diff.

Safe by default. A readOnly flag (global or per-system) blocks every write method. Read-only POST queries (search, where-used, package tree) remain allowed so agents can still discover.

Robust. Per-request timeout. CSRF token negotiation with auto-retry on 403. Self-signed cert opt-out. Optional debug tracing to stderr.

Structured errors. ADT's <exc:exception> envelopes are parsed into { type, message, namespace } so failed calls don't dump XML into the agent's context window.

Install

# global
npm install -g claude-for-abap

# or run without installing
npx claude-for-abap

Requires Node.js 22.19+ (undici v8, used as the HTTP client, requires this minimum).

Configure

Create your config:

mkdir -p ~/.sap-adt-mcp
cp config.example.json ~/.sap-adt-mcp/config.json
$EDITOR ~/.sap-adt-mcp/config.json

The server searches in this order:

1. $SAP_ADT_MCP_CONFIG (absolute path)

2. ~/.sap-adt-mcp/config.json

3. ./config.json (cwd at server start)

Sample config

{
  "defaultSystem": "DEV",
  "readOnly": false,
  "systems": {
    "DEV": {
      "host": "https://sap-dev.example.com:44300",
      "client": "100",
      "language": "EN",
      "user": "DEVELOPER",
      "password": "env:SAP_DEV_PASSWORD",
      "rejectUnauthorized": false
    },
    "QAS": {
      "host": "https://sap-qas.example.com:44300",
      "client": "200",
      "user": "DEVELOPER",
      "password": "env:SAP_QAS_PASSWORD"
    },
    "PRD": {
      "host": "https://sap-prd.example.com:44300",
      "client": "300",
      "user": "READONLY",
      "password": "env:SAP_PRD_PASSWORD",
      "readOnly": true
    }
  }
}

Per-system options

Read-only mode

readOnly: true (top-level or per-system) refuses any unsafe HTTP method. Whitelisted read-only POST endpoints (nodestructure, search, usagereferences, parsers, checkruns) remain available so agents can still discover and analyze without being able to modify.

Recommended: set readOnly: true for QAS and PRD profiles. Keep DEV writable.

Self-signed certificates

Many internal SAP systems use self-signed certs. "rejectUnauthorized": false disables TLS validation for that profile only. Don't set this on PRD.

Connect a client

Claude Code (CLI)

claude mcp add sap-adt -- npx claude-for-abap

Pass secrets through the registration:

claude mcp add sap-adt \
  --env SAP_DEV_PASSWORD=... \
  --env SAP_PRD_PASSWORD=... \
  -- npx claude-for-abap

Claude Desktop

Edit claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "sap-adt": {
      "command": "npx",
      "args": ["-y", "claude-for-abap"],
      "env": {
        "SAP_DEV_PASSWORD": "..."
      }
    }
  }
}

Quit and restart Claude Desktop fully (system tray → Quit) for the change to apply.

Validate before connecting

npx claude-for-abap --validate-config

Loads the config and pings every system; exits non-zero if any are unreachable or rejecting credentials. Run this first when troubleshooting.

Tools

Object-source CRUD

Quality

Discovery

Cross-system

Runtime errors

Data

Transports

Escape hatch

adt_request — direct ADT REST call. Use this when a niche endpoint isn't covered by a high-level tool. Handles auth / CSRF / cookies / sap-client automatically.

Object types

Friendly aliases (any of either column work):

Clean Core: prompts + reference

The server ships an opt-in Clean Core layer for SAP S/4HANA work. There are two pieces, and they are deliberately separate:

• Five MCP prompts (src/prompts.js) — the operational surface. The user invokes them as slash commands. Each one pairs a slice of the Clean Core framework with the relevant adt_* tools so the model can act on a real system, not just lecture about levels.

• Long-form reference (skills/abap-clean-core/) — the framework's full text: Stay Clean / Get Clean playbook, A/B/C/D level deep-dive, Cloudification Repository state semantics, ABAP Cloud allowed/forbidden lists, the SAP Application Extension Methodology (3 phases), governance practices, KPI calculations, ATC exemption process. Read once, link to it from PRs, hand to a new team member. The prompts above quote what they need; the reference is everything else.

Design choices

• Opt-in, not auto-firing. Clean Core is an S/4HANA discipline. ECC developers should not have it imposed on them. Nothing fires unless the user types the slash command.

• ECC applicability check baked into every prompt. The first thing each prompt body asks the model to do is verify the target system is S/4HANA. On ECC, it backs off and offers help in classic-ABAP idioms with no level labels.

• Tone is descriptive, not judgmental. "I know it's Level D, just ship it" is honored. The agent ships, marks the level, sketches the Level A refactor for later, and moves on.

The prompts

In Claude Code (assuming you registered the server as sap-adt), the exact commands are:

The slash-command name structure is determined by the MCP client: the mcp__<server-alias>__ prefix is added automatically based on the alias you used when registering the server. If you registered the server with a different alias (e.g. claude mcp add cc -- npx claude-for-abap), the prefix changes accordingly (/mcp__cc__clean_core_grade).

In Claude Desktop, prompts appear in the slash-command picker — same naming.

Argument flow examples

Atomic prompts (grade, review) take all their arguments inline and return a structured analysis:

You: /mcp__sap-adt__clean_core_grade object:ZCL_PRICING type:class system:DEV
Agent: → adt_get_source { ... }
       → adt_run_atc { ... }
       Verdict: Level C. Two SELECTs from MARA without using the released
       I_Product view; one CALL FUNCTION to internal FM RV_PRICE_PRINT.
       Refactor sketch: replace SELECT with `from I_Product`; encapsulate
       the RV_PRICE_PRINT call in a Z-class so the dependency is localised.

Mode-loading prompts (refactor, create, design) optionally take a seed; without one, they wait for the user's natural-language follow-up:

You: /mcp__sap-adt__clean_core_create
Agent: I'm in Clean Core CREATE mode (Level A by default). What do you
       want to build, on which package and system?
You:   A small Fiori list-report of overdue invoices, package ZFIN_REPORTS,
       system DEV.
Agent: Plan: a CDS view projecting I_OperationalAcctgDocItemCube for items
       with NetDueDate < today; a behavior definition; a service binding
       exposing it to Fiori Elements list-report. Three objects. Confirm?

Or with a seed argument so the request is one-shot:

You: /mcp__sap-adt__clean_core_create requirement:"Fiori list-report of
     overdue invoices" package:ZFIN_REPORTS system:DEV

Read the long-form reference

skills/abap-clean-core/ is the canonical source for everything the prompts quote and more. If you're setting up Clean Core governance for a real program — KPI baselines, ATC exemption discipline, maturity assessment, on-stack vs side-by-side trade-offs at the architecture level — that's where the depth lives.

The directory is structured as one entry point (SKILL.md) plus four deep-dive files in references/:

• references/levels-detailed.md — Cloudification Repository state values, released local vs released remote APIs, reclassification dynamics, per-anti-pattern remediation

• references/decision-framework.md — fit-to-standard, the SAP Application Extension Methodology in detail, on-stack vs side-by-side triggers, hybrid patterns, worked scenarios

• references/governance.md — Stay Clean / Get Clean playbook, the four KPIs and how to compute them, ATC exemption process, maturity assessment, multi-year roadmap

• references/abap-cloud-rules.md — full allowed/forbidden lists, RAP / CDS / business object interfaces / Custom Fields, prebuilt services, classic-to-cloud migration patterns

You can install the reference as an actual auto-loading Claude skill by copying or symlinking skills/abap-clean-core/ into your ~/.claude/skills/ — but that's an explicit choice. The default behavior of this repo is opt-in, prompt-only.

Examples

See examples/ for end-to-end agent workflows: project discovery, class audit, cross-system release verification, where-used-driven refactor, test triage.

Quick taste:

You: "Compare class ZCL_PRICING between DEV and PRD on the live systems."
Agent: → adt_compare_source { systemA: "DEV", systemB: "PRD",
                              object: "ZCL_PRICING", type: "class" }
       Returns: { identical: false, stats: { added: 14, removed: 9 },
                  diff: "..." }
       Then narrates the meaningful changes.

Architecture

MCP client (Claude Desktop / Claude Code / custom)
         │  stdio (JSON-RPC)
         ▼
  ┌──────────────────────────────────┐
  │  src/server.js                   │  CLI + MCP dispatch (thin)
  │   ├─ src/tools/*.js              │  one module per category:
  │   │                              │    connection, source, quality,
  │   │                              │    lifecycle, discovery,
  │   │                              │    cross-system, transports,
  │   │                              │    runtime, data, request
  │   ├─ src/object-uris.js          │  type alias → ADT URI map
  │   ├─ src/node-structure.js       │  package tree XML parser
  │   ├─ src/object-references.js    │  <objectReference> parser
  │   ├─ src/dump-feed.js            │  runtime-dumps Atom parser
  │   ├─ src/data-preview.js         │  Data Preview XML parser + SELECT guard
  │   ├─ src/diff.js                 │  unified diff (LCS)
  │   ├─ src/adt-error.js            │  <exc:exception> parser
  │   ├─ src/lock.js                 │  ADT lock acquire / release
  │   └─ src/adt-client.js           │  HTTP client: auth / CSRF / cookies / timeout
  └──────────────────────────────────┘
         │  HTTPS
         ▼
   SAP system (ADT REST: /sap/bc/adt/...)

Two runtime dependencies: @modelcontextprotocol/sdk (the MCP wire protocol) and undici (HTTP with custom TLS dispatcher). Everything else is stdlib.

Multi-step editing pattern

For most edits, adt_set_source is enough — it acquires the lock, writes, and releases. For workflows that touch the same object multiple times within a single turn (e.g. apply N method-level patches, then activate), use the sticky-lock pattern:

1.  adt_lock { object: "ZCL_X", type: "class" }              → returns lockHandle
2.  adt_set_source { object: "ZCL_X", type: "class",         (repeat as needed)
                     source: "...", lockHandle: "<handle>" }
3.  adt_activate { objects: [{ name: "ZCL_X", type: "class" }] }
4.  adt_unlock { object: "ZCL_X", type: "class",
                 lockHandle: "<handle>" }

The lockHandle parameter on adt_set_source skips internal lock/unlock when present.

Caveats

• NetWeaver release variation. A few ADT endpoints (especially around transport requests, ATC, and object-create XML shapes) have small shape differences across NW 7.5x, S/4 on-prem, and Steampunk. If a high-level tool fails with HTTP 4xx, the tool description notes which endpoint it hits — fall back to adt_request with the right path / content type for your release.

• ABAP Cloud (BTP / Steampunk). Only on-prem ADT 7.5x+ has been actively exercised. Steampunk uses a stricter object-type allowlist (only released / public APIs) and some collection endpoints differ. Steampunk users: please open an issue with the endpoints that differ; PRs welcome.

• Object-create XML shapes. The body templates target modern on-prem releases. Some older systems require additional attributes (adtcore:masterLanguage, etc.) — open an issue if your system rejects the create payload, and include the response body.

• DDIC primitives (tables, data elements, domains) are not creatable via adt_create_object — these need a richer DDIC-specific payload that we haven't generalised. Use adt_request for now.

Troubleshooting

"failed" in Claude Desktop's MCP server list. Run claude-for-abap --validate-config from a terminal. If that prints OK, the server is fine and the issue is in your claude_desktop_config.json (wrong path or env).

403 with x-csrf-token: required. Should self-heal — the client refetches the token and retries. If it persists, you likely have an SSO / front-end auth in front of ADT that breaks Basic auth; check your ICM and SAP web dispatcher rules.

Read-only mode refuses an obviously-read endpoint. It's probably a POST endpoint not in the whitelist. Open a PR or issue with the path; we'll add it. Or temporarily flip the system to readOnly: false.

SAP_ADT_MCP_DEBUG=1 traces every request and response (status, latency, URL, request headers minus Authorization) to stderr. The MCP client shows stderr in its server log, so check there.

Contributing

See CONTRIBUTING.md — bug reports, new tool coverage, NetWeaver compatibility notes, docs, examples all welcome.

License

MIT — see LICENSE.