For manufacturers and energy companies running S/4HANA

An SAP MCP server for your own agent, running under SAP authorisations rather than a service account

An SAP MCP server gives Claude or another agent named operations on S/4HANA, such as reading a bill of material or tracing a change number, under the asking person's SAP authorisations. As of 19 September 2026, SAP ships MCP for Joule and for developers, not a governed server for your own agent. Pexon builds that server.

Checked 19 September 2026

What SAP ships for MCP today, and what each piece does not do

What SAP shipsWhat it doesWhat it does not do
Claude in SAP Business AI and Joule (SAP and Anthropic, 20 May 2026)Claude runs inside SAP's platform; Integration Suite APIs can be converted into MCP tools for Joule agentsGive an agent outside SAP, in Claude or your own tenant, a governed way into S/4HANA
SAP HANA Cloud MCP support (early 2026, per trade press)Lets agents query the database layerGo through application logic or S/4HANA authorisation objects, which is where business rules live
ABAP MCP server (generally available at Sapphire 2026)Helps developers write and analyse ABAP codeRead or change business data; it is a developer tool
Open-source developer servers for CAP, UI5 and MDK on GitHubSupport AI-assisted development of SAP applicationsTouch a production system at all

No PLM-, BOM- or change-specific MCP server from SAP or the open-source community was found on the check date. SAP is moving quickly; this table is dated for that reason.

SAP's MCP work brings agents into SAP's platform. An agent that lives next to your PLM and MES still has no governed door into S/4HANA.

Why SAP is hard for agents

Four places an SAP connection fails its review

01 · Authorisations

A technical user sees too much

SAP decides access through authorisation objects bundled into PFCG roles. A server that logs in as one technical user returns whatever that user may read, for everyone who asks. The fix is the person's own identity carried through to SAP, for example by principal propagation over SAP BTP, or a technical role cut so narrowly that its reach can be written on one page.

02 · Released APIs

Your edition decides the interface

S/4HANA on-premise, private cloud and public cloud do not release the same APIs. A tool that reads through an OData service on one system needs an RFC or BAPI on another, and the tool list has to be checked against your release before anyone promises it.

03 · Your data model

Z-fields and custom logic

Most industrial SAP systems carry customer fields and custom function modules that no public API describes. An agent reading only the standard entities will answer confidently from half the record. We map the custom fields a tool depends on and name them in its description.

04 · Licensing

Digital Access is a contract question

Whether an agent's calls count as indirect or digital access depends on your SAP contract. It is a cost risk nobody should guess at. We ask your SAP account team with the tool list and expected volumes in hand, and the answer goes into the build plan.

What we build

A first tool list, written as the manifest your basis team signs

Six named operations for the case the research points to first in manufacturing: engineering data that lives in a PLM and in SAP at the same time. Two are backed by OData APIs we could source; the rest are resolved against your release in the blueprint.

# sap-mcp-server: tool manifest, draft for blueprint week 1
# Every tool runs as the asking user. No tool posts, releases or deletes.
tools:
  get_bom:                      # material, plant, BOM usage -> items with quantities
    backend: API_BILL_OF_MATERIAL_SRV    # OData, S/4HANA
    mode: read
  get_change_record:            # change number -> status, validity, objects
    backend: API_CHANGEMASTER            # OData, S/4HANA
    mode: read
  where_affected_by_change:     # change number -> open production orders, stock, routings
    backend: OData or BAPI, confirmed in blueprint
    mode: read
  compare_ebom_to_mbom:         # PLM structure vs SAP BOM -> list of deviations
    backend: get_bom + PLM connector
    mode: read
  find_similar_materials:       # description, classification -> candidate duplicates
    backend: OData or BAPI, confirmed in blueprint
    mode: read
  draft_change_request:         # proposed change -> draft a person releases in SAP
    backend: none until approved
    mode: propose
audit: every call logged with user, tool, arguments, SAP response code

The manifest is the contract. If a question needs an operation that is not on it, the answer is a new line in the manifest and a review, not a wider role.

What we build, and what we test it against

What we build

  • The server: an MCP server in your environment, publishing the agreed tool list and nothing else, with typed inputs and documented results.
  • The identity path: the user's Entra ID or other directory identity mapped to their SAP user, so SAP's own authorisation check decides every answer.
  • The audit trail: one log line per tool call with the person, the operation and SAP's response, in the log your security team already reads.
  • The handover: code in your repository, tests against a fixture of your data, and a runbook, so the server outlives the pilot team.

What we test it against

  • Authorisation trace: the same question asked as two users with different roles must return two different answers.
  • Release check: every tool's backend confirmed as released on your edition before the tool is built.
  • Licensing answer: SAP's written answer on how the calls are licensed, attached to the build plan.
  • Failure cases: a locked user, an expired token and a missing authorisation each return a clear refusal, not partial data.

When this is the wrong page

If everyone who needs the assistant works inside SAP all day, Joule with SAP's own agents is the shorter path, and SAP maintains it. If the question is about numbers across SAP, such as margin by product line or stock over twelve months, an MCP server calling APIs per question is the wrong shape; that is a governed SAP data layer for AI. An SAP MCP server is right when an agent outside SAP has to act on operational records, one named operation at a time.

Start with the Readiness Blueprint.

Two weeks, €4,900 fixed price. Week one agrees the tool list with your SAP owners and checks every backend against your release. Week two tests whether the user's identity reaches SAP's authorisation check, and gets the licensing question answered by SAP. You keep the manifest and the costed build plan, whoever builds it. All prices are net and exclude VAT.

What we need from you: a named SAP basis contact, a sandbox or quality system we may call, and one business owner who can say which questions the agent should answer first. What happens after: a use-case pilot builds the server, and it runs in your environment under your operations team, or ours if you prefer.

Related

Questions SAP teams ask before the first tool call

Doesn't the SAP and Anthropic partnership already cover this?

No, it covers the opposite direction. The announcement of 20 May 2026 brings Claude into SAP's Business AI platform and Joule, where SAP decides which tools exist. If your agent lives outside SAP, in Claude, in your own tenant or next to a PLM and an MES, it still needs a server that exposes SAP operations to it and carries the user's authorisations through.

Which SAP releases can an MCP server reach?

Any release that exposes an interface, which is all of them. S/4HANA publishes OData APIs on the SAP Business Accelerator Hub; older landscapes and custom logic are reached through RFC and BAPI. What differs is which APIs are released on your edition: on-premise, private and public cloud do not expose the same set, so the tool list is checked against your system in the first week rather than assumed.

Does an agent calling SAP trigger Digital Access licensing?

We do not know for your contract, and nobody outside SAP can tell you reliably. Access by third-party systems is governed by your SAP licence terms, and those vary between contracts. We put the concrete question, with the tool list and expected call volumes, to your SAP account team during the blueprint, before anything is built, and record the answer in the build plan.

What does an SAP MCP server cost?

The two-week Readiness Blueprint is fixed at €4,900: the agreed tool list, the authorisation model, the licensing question answered by SAP and a costed build plan. A first production server is usually a use-case pilot from €15,000. Model tokens are billed by Anthropic or your cloud provider, and any SAP licence effect by SAP. All prices are net and exclude VAT.

What will you not do on an SAP MCP server?

We will not give an agent posting rights in finance, and we will not connect through a user with broad authorisations to make a demo work. Changes are proposed as drafts a person releases in SAP. We have also not built an SAP MCP server for a named customer yet, so the first engagement is where the tool list is proven against your system, not where a product is installed.

Next step

Not a sales call. An architecture call.

Thirty minutes with the architect who would run the engagement, with your SAP basis lead in the room if you like.