Connects an AI assistant to an ADT-enabled SAP ABAP system through a governed capability-discovery and invocation gateway.
Search available SAP capabilities by keyword, name, category, risk level, and pagination cursor.
Describe one or more capabilities (up to 10) to retrieve exact input schemas, schema hashes, and optionally output schemas.
Invoke read-only capabilities using a described capability name, schema hash, and arguments.
Invoke write capabilities with schema-hash validation for scoped changes.
Invoke destructive capabilities when exposed, marked as destructive and not read-only.
Supports reading ABAP, inspecting check results, and making scoped changes through discovered capabilities; actual operations depend on what the connected SAP system exposes.
Provides tools to connect to SAP ABAP systems via ADT HTTP API, enabling management of ABAP objects, searching, reading source code, and more.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@sap-abap-mcp Search for ZCL_DEMO class in DEV100 and show its source."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Connect your AI assistant to an ADT-enabled SAP system to read ABAP, make scoped changes and inspect actual check results.
Install Node.js 22/24 and connect to your company VPN.
Run npx -y @coaspe/sap-abap-mcp@2.2.0 onboard (npx.cmd on Windows).
Verify and save your SAP connection, register your host, then restart the app and verify an actual SAP system read.
Host | Setup command |
Claude Code |
|
Codex |
|
VS Code |
|
Cursor |
|
Gemini CLI |
|
Claude Desktop |
|
Version 2.2.0 includes company CA files, six-host registration and Windows SAP
Logon import. Omit --apply to preview a registration; existing entries are
preserved and existing settings files are backed up. Claude Desktop MCPB users
should avoid a duplicate registration. On Windows replace npx with npx.cmd.
CA profiles use file format version 3 and need 2.2.0 or later. Existing npm registrations pin their package version: restart alone does not update a 2.1.0 registration. Use registration doctor and reviewed repair to update an existing registration to 2.2.0 before using a CA profile.
See setup details, including Linux environment-only credentials and registration recovery.
2.1.0 fixes noninteractive startup, native macOS credential storage and registration recovery. Windows DPAPI explicitly loads its native security module even when Node inherits a PowerShell 7 module path. Bundled npm/npx Windows launchers are resolved without a shell while preserving custom wrappers. This release also reduces the runtime package. Patched vendor bytes are preserved across Windows and Unix checkouts. Existing five gateways still reach 120 capabilities.
Related MCP server: dassian-adt
Repeated source patches use the same current-source view, preserve literal $
sequences and activate after unlock. Transport release supports ADT background
runs; the release also improves RAP response recovery, V4 unpublish errors,
old open-transport discovery, CTS-key object assignment, Dictionary checks and
top-level package creation. See the eight fixes and validation limits.
These paths have local automated coverage. Live B4D write/release acceptance has not been established by this release. Older npm 2.1.0 still has the reported source-patch limitations; upgrading the package and its pinned host registration is required to use the fixes.
Select the existing sap-change-object prompt, with the configured systemId,
exact Z program/class as target, and desired behavior in goal. It reuses
source, semantic/public API, dependency, quality and transport features.
Production/read-only policy, package restrictions, transport requirements and exact-source checks remain enforced. In 2.2.0, source writes recheck the same current-source view under lock, without an explicit inactive probe. A changed current version stops the write. Requested activation follows successful unlock and is skipped on syntax errors. Uncertain multi-object writes stop with saved/uncertain/unattempted identities; there is no automatic rollback or atomic-transaction claim. Inspect state before retrying. Zero tests, unsupported checks and partial findings cannot establish completion.
For a read-only evaluation, keep readOnly: true and expose only search,
describe and invoke_read gateways in the host. Do not execute write, activation,
Unit, ATC, debugger or profiling operations during that evaluation.
See workflow prompts and T1–T5 criteria.
Scope | Evidence and limit |
Clean npm tarball, macOS arm64, Node 24 | Eight launch/profile combinations, Unicode/spaces, help, invalid commands and shutdown passed on the reliability candidate |
macOS Keychain | Actual throwaway legacy-item/UTF-8/9,000-byte storage, renewal and deletion passed |
Windows read-only field procedure | Owner reports all supplied steps passed without candidate source changes; sanitized operation records were not supplied |
Windows DPAPI / npx.cmd automation | Windows CI passed: native storage/module-shadow regression, bundled launchers and eight installed-tarball protocol cases |
Native Desktop MCPB / host restart | Standalone bundle checks and an owner's overall field report remain distinct from retained app-level evidence |
Current 2.x live-SAP acceptance | not verified by retained operation-level records; owner-reported read-only completion is recorded separately |
Historical live SAP / model comparison | 2026-08-05 results are historical; no identical-model T1–T5 completion rate or product ranking is established |
Candidate measurements reduced compressed npm size from 26.29 to 13.33 MB and installed dependency bytes from 118.89 to 91.81 MB. Three warm protocol sessions measured 346–351 to 223–239 ms. Cold install was 12.92 to 16.06 seconds; no cold speed improvement is claimed. See exact snapshots and limits.
Symptom | Next action |
MCP closes immediately | Keep |
| Complete local setup and reconnect; this state is not an accepted SAP installation |
Removed Node/cache or old registered version | |
Certificate untrusted / expired / hostname mismatch | Keep verification enabled; use the company CA PEM file or ask the administrator to fix the certificate/URL. See TLS trust |
SAP URL unknown | On Windows use SAP Logon import to find candidate URLs, then check ADT and confirm the endpoint with your administrator |
SAP authentication/authorization failure | Renew locally or ask the administrator for ADT permission; avoid blind repeated logins |
Inactive conflict / uncertain write | Read current active/inactive state and reconcile with its owner before resuming |
Transport assure and CI, full reference, CLI, profiles, changelog.
Reliability validation and field procedure. Checkout manifest version: 2.2.0.
Normal clients should omit both --api-version and --toolsets.
PRIVACY.md: no default remote telemetry. Credentials remain in OS-protected storage. Feedback exports are voluntary; public adopter names and quotes require separate explicit permission. Downloads and automated tests do not establish active users, task success or independent adoption.
Get exact input schemas and hashes: name or names (up to 10).
Inputs: name, names, includeOutputSchema.
Invoke a destructive capability with its described schemaHash and arguments.
Inputs: name (required), schemaHash (required), arguments.
Invoke a read capability with its described schemaHash and arguments.
Inputs: name (required), schemaHash (required), arguments.
Invoke a write capability with its described schemaHash and arguments.
Inputs: name (required), schemaHash (required), arguments.
Find capabilities; browse categories if keyword search misses.
Inputs: name, risk, limit, query, +2 more.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 5 tools
Each tool has a clearly distinct role: search discovers capabilities, describe retrieves schemas/hashes, and the three invoke tools are cleanly separated by effect class (read/write/destructive). An agent can select the right tool based on intent without overlapping functionality.
All tools follow the same `sap.capability.<action>` namespace with lowercase and snake_case suffixes. The invoke variants consistently append the effect type, creating a predictable and readable pattern.
Five tools neatly cover the generic capability gateway pattern: discovery, introspection, and execution split by effect class. Each tool earns its place with no redundant or missing meta-operations.
The set covers discovery (search), schema retrieval (describe), and execution across all effect classes (read/write/destructive), forming a complete lifecycle for invoking SAP capabilities without obvious dead ends.
MCP server unifying ERPs, CRMs, APIs and knowledge base for Claude, ChatGPT and Gemini.
MCP server that lets your AI assistant build and deploy apps, hosting each one at a live URL.
Cloud-hosted MCP server for secure AI access to enterprise data sources via CData Connect AI.
Hosted MCP server connecting AI assistants to 9,000+ apps and 40,000+ actions via Zapier.
