sap-adt-mcp

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

Previously published as claude-for-abap — that package still works but is deprecated; new installs should use sap-adt-mcp.

# global
npm install -g sap-adt-mcp

# or run without installing
npx sap-adt-mcp

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.

Audit log

Every write the server performs against SAP (POST/PUT/DELETE/PATCH — locks, source updates, activations, transport operations) is appended to a local JSONL file, including which MCP tool triggered it and, for blocked attempts in read-only mode, the violation itself. Reads and read-only queries are not logged. Nothing leaves your machine — this is your local answer to "what exactly did the AI change?".

Default location: ~/.sap-adt-mcp/audit.log. One JSON object per line:

{"ts":"2026-06-11T12:00:00.000Z","tool":"adt_set_source","host":"https://...","sapUser":"DEVELOPER","method":"PUT","path":"/sap/bc/adt/programs/programs/ztest/source/main","status":200,"ok":true,"transport":"E4DK900123"}

Configure or disable:

{ "audit": { "enabled": false, "path": "/var/log/sap-adt-mcp/audit.log" } }

…or set SAP_ADT_MCP_AUDIT=0 (also accepts false/no/off).

Automatic error reporting

The server sends small, redacted reports to the maintainer so defects get found and fixed. This is on by default and the server prints a notice saying so on startup. There are three channels:

1. Crash — a tool handler throws an unexpected error.
2. ADT error — a tool returns a non-2xx ADT response that the classifier flags as a likely tool bug (406/415 content negotiation, malformed requests, server dispatcher blow-ups). User/business-side responses (401/403/404, lock and enqueue conflicts, data-preview SQL errors) are not reported.
3. Agent-reported — the calling agent files a defect the other two channels can't see (wrong data in a successful response, an ignored parameter, a missing capability) via the adt_report_issue tool.

What is sent: the sap-adt-mcp version, Node version, OS, the tool name, and the error/finding with a fingerprint for de-duplication, plus an anonymous install id (random bytes, cached at ~/.sap-adt-mcp/install-id) that lets repeat reports from the same install be grouped for triage — it identifies neither you nor your system. Before anything leaves your machine it is scrubbed of hostnames, users, passwords, tokens, IPs, and emails; tool arguments and free-text fields are redacted the same way. Reports go to a relay the maintainer owns, which files/de-dups a GitHub issue — the relay holds the GitHub credentials, never this package.

Turn it all off:

{ "reporting": { "enabled": false } }

…or set SAP_ADT_MCP_REPORT=0 (also accepts false/no/off). Finer control:

Local control panel

A small HTML button panel for the read-only tools — search, grep, get_source, read_table, ATC, where-used, packages, transports, dumps, inactive objects — so you can poke at SAP from a browser without going through an agent.

The trick: the panel is served from inside the MCP process itself, reusing the same tool handlers. So it is reachable only while a session keeps the MCP connected — close the session (or disconnect the MCP) and the process exits, taking the panel down with it. There is no standalone server to leave running.

Open it from a session (easiest). Just ask the agent to open it — that calls the adt_open_panel tool, which starts the panel on demand and opens the URL in your browser. adt_close_panel stops it. In Claude Code the bundled /panel command does the same (/panel, /panel url, /panel close). Nothing listens until you ask — the socket opens only on that call.

Or auto-start at boot. Set it in config or env and it comes up with the server:

{ "panel": { "enabled": true, "port": 0 } }

…or SAP_ADT_MCP_PANEL=1 (SAP_ADT_MCP_PANEL_PORT pins a port; port: 0 / unset picks a random free one). On boot the server prints the URL, e.g.:

[sap-adt-mcp] panel: ready (read-only) → http://127.0.0.1:39555/?t=<token>

Safety: bound to 127.0.0.1 only, gated by a per-boot random token in that URL, and limited to a curated read-only allowlist — no write tool (set_source, activate, delete, lock, transport release) is reachable from a button, regardless of config. Each tool's form is rendered from its live input schema, and the system selector at the top targets any configured system.

Connect a client

Claude Code (CLI)

claude mcp add sap-adt -- npx sap-adt-mcp

Pass secrets through the registration:

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

Claude Desktop

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

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

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

Validate before connecting

npx sap-adt-mcp --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

Debugger

External ABAP debugger over ADT's /sap/bc/adt/debugger/* API. Set a breakpoint, wait for a session to hit it, then inspect the stack/variables, step through, and (guarded) change values. Requires debug authorization on the backend; minimum NW 7.31 SP04 + Kernel 7.21 (fine on S/4).

Typical flow:

1.  adt_debug_set_breakpoint { object: "ZREPORT", type: "program", line: 42 }   → id
2.  adt_debug_listen {}          → ARM IT FIRST, then trigger the run
3.  (trigger: a Fiori/OData call, an ADT classrun, an RFC, a job)
4.  listen returns caught:true → already attached (or caught:false → call again)
5.  adt_debug_stack {}           → where execution paused
6.  adt_debug_variables { names: ["sy-subrc", "lv_total"] }
7.  adt_debug_step { kind: "over" }   → advance; or set_variable / continue
8.  adt_debug_stop {}            → resume the debuggee + cleanup

What an external breakpoint traps: HTTP/ICF (Fiori, OData, ADT), RFC and background sessions — not your own SAP GUI dialog session. Running the report from SE38 with F8 traps nothing and opens no debugger, which reads like a broken feature but isn't. To debug dialog-only code, run it inside an ICF session with a throwaway IF_OO_ADT_CLASSRUN wrapper and trigger POST /sap/bc/adt/oo/classrun/<class>:

METHOD if_oo_adt_classrun~main.
  DATA lt_list TYPE TABLE OF abaplist.
  " There is no screen in an ICF session, so list output must go to memory.
  SUBMIT zoy_dbg_target EXPORTING LIST TO MEMORY AND RETURN.
ENDMETHOD.

SUBMIT opens a fresh program context, so you debug the target report on its own.

Timing: arm adt_debug_listen before triggering the run — for a fast request (Fiori/OData) arming it after clicking is already too late. A trapped debuggee then waits only ~15–20 s for the attach before resuming on its own, so listen attaches immediately with no round trip in between. A caught: false can therefore mean three things: the run hasn't happened yet, the listener was armed too late, or that line never executes in this flow — keep a known-good control breakpoint to tell a wrong breakpoint apart from a setup problem.

Read-only mode: inspection (breakpoints, listen, stack, variables) is allowed; the WRITE tools above (step, goto-stack, set-variable, watchpoints) are refused under readOnly: true. Debugging another user's session (requestUser) is off by default; enable it per system with "debug": { "allowRequestUser": true } (needs backend debug authorization).

Data

Transports

Control panel

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):

Skills (packaged workflows)

The raw tools are building blocks; skills/ ships ready-made workflows for Claude Code that orchestrate them. Each SKILL.md lists its own prerequisites (minimum NetWeaver release, required authorizations, read-only compatibility).

To use one, copy its folder into your project's .claude/skills/ (or ~/.claude/skills/ for all projects):

cp -r node_modules/sap-adt-mcp/skills/dump-triage .claude/skills/
# or from a clone: cp -r sap-adt-mcp/skills/dump-triage .claude/skills/

Claude Code picks them up automatically; they trigger when the conversation matches (e.g. "is E4DK900123 safe to release?" → transport-release-gate).

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 sap-adt-mcp), 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.

Large objects (thousands of lines)

A multi-thousand-line class or program can exceed the per-call I/O cap, so its source cannot be passed inline. Keep the content on disk instead — it never enters the agent context:

1.  adt_get_source { object: "ZCL_BIG", type: "class",
                     outputFile: "/tmp/zcl_big.abap" }   → writes the file, no inline source
2.  (edit /tmp/zcl_big.abap locally)
3.  adt_set_source { object: "ZCL_BIG", type: "class",
                     sourceFile: "/tmp/zcl_big.abap",
                     transport: "E4DK900123" }           → MCP reads the file and PUTs it

Note: assigning to a transport works headless, but TR creation (adt_create_transport) routes through a GUI dialog on some systems and can fail with a 500 — pass an existing TR id instead.

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 sap-adt-mcp --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.