Agent access to your systems

The embedded preview needs a system that accepts basic auth: the documentation states that pure SSO or SAML without a basic-auth fallback is not supported, and points at `external` mode, which opens the app in the normal browser with the existing SAP session. On the web only the language half runs; the preview with its proxy, traffic log and screenshot, Ctrl+F3 activation, the render gate, the workspace-wide check and the MCP servers are desktop-only. The agent app tools send nothing until `abap2ui5.agent.enableAppTools` is switched on in User settings, because an event they fire runs as the configured user and changes data on the system for real.

Every gate here runs off-system, so what the class does on a real system and what the view does on the oldest UI5 release it has to serve are what the `pitfalls` catalogues carry and no local run can see. It runs abap2UI5 app classes, so ABAP that is not one has nothing here to deploy into. In system mode the sandbox tools are not offered, and an `app_act` event runs on the system as the configured user.

It writes and it administers — ETL extraction and database user management are not read operations, and the same connection that profiles a column can move data out of the tenant. Masking is only as good as the configuration naming the fields, so an unlisted sensitive column is an unmasked one.

A system that reports itself productive, or whose role the probe cannot prove non-productive, is write-locked with no override, so changing code in production is not something this does. The permission ladder constrains the package, object name and transport request the server itself passes on a write, and is not a sandbox around ABAP it executes: `abap_run`, `abap_test` and the opt-in `core.eval` run under the connected technical user's SAP authorisations, which is where the boundary actually sits. Data preview, dump variable contents and driving a transaction by batch input stay off in every mode, `admin` included, until each is named explicitly.

Configuration over an API you must already have entitled and provisioned; and the smallest of these repositories, so what it covers is what its config file happens to name.

It administers a live Cloud Foundry landscape: `Apps`, `Organizations`, `Spaces`, `ServiceInstances` and `SecurityGroups` all carry create, update and delete alongside list and get. What a caller can reach is set by the role they hold, and the three templates in `xs-security.json` are MCPViewer for read, MCPEditor for read and write, and MCPAdmin for full administrative access.

It administers a real account through Core Services, so what it can change is what the caller can change. And the UI half depends on hosts that render mcp-ui or MCP Apps resources; elsewhere those five tools return markup nobody displays.

It puts create, update and delete on business entities behind natural language. Whatever the destination's principal may change, an agent may change — and "create a new customer" is a sentence, not a form somebody signed off.

A thin declaration over CPI's APIs: it reaches what those APIs expose and nothing else, and it inherits whatever the tenant's own authorisations allow.

Neither the package manifest nor the repository tree declares a licence.

Design-time only: no message monitoring, no deployment, no runtime status. That is a different server in this same section, and the two do not overlap.

This is the database, not an application in front of it, so what an agent can read is whatever the connection user can read — the guardrails bound the shape of a query, not the sensitivity of what it returns. Grant the connection the rights you would grant a person.

Not for production: the README states local and development instances only, and the server refuses a HAC URL whose host is not private unless `HAC_ALLOW_REMOTE=true` overrides that guard. Masking applies to `config_get` only, so `logs_tail` and `monitoring_info` output, thread dumps included, is unmasked and may carry sensitive data, and whatever a tool returns goes to the MCP client and its model. Certificate checks are disabled whenever `HAC_INSECURE_TLS=true` is set, which the README names as typical for local HAC.

Its maker offers it commercially.

It acts on a real account. Authorisation is the caller's own — single sign-on through the SAP Default Identity Provider, across the global accounts that identity reaches — so what it can change is what you can change, and an agent invoking it is doing administration rather than proposing it. No public source, so what it exposes is what the documentation says it exposes.

Neither the package manifest nor the repository tree declares a licence.

Neither the package manifest nor the repository tree declares a licence.

It deploys and undeploys, and it enumerates credentials, keystores and OAuth2 configurations. Listing security material is a read, and it is still a read of the thing that protects everything else the tenant talks to.

Functional boundary Business Source License 1.1, which is source-available rather than open source: reading and evaluating are fine, and production use is restricted until the licence's change date. That is a procurement question, not a technical one, and it is the thing to settle first. Read-only also means read-only — no modelling, no loads.

Maintainer claim The maintainer claims alignment with SAP API Policy v4/2026 and documents a read-only posture. Treat that as a maintainer claim, not an atlas compliance decision, and verify the deployment pathway and intended data use.

It reads the landscape inventory, not the systems in it — nothing here reaches a running application, and nothing here changes anything. A landscape description is also commercially sensitive in its own right: versions and hosts are what a targeting exercise starts from.

OData v2 only, and untouched since August 2025 — the same author's [Go port](artifact:oisee-odata-mcp-go) is where v4 support and the tool-explosion work went.

Universal mode trades discoverability for size: one generic tool is harder for a model to pick correctly than a named one, which is why it is not the default. On a small service, standard mode is the better answer.

It fronts APIs you already have; nothing here reads a system's source or model, and a config file is only as good as the entity sets somebody wrote into it. OData V2 is the query surface it names.

Real SAP calls need the proprietary [SAP NetWeaver RFC SDK](external:sap-nw-rfc-sdk) and the `pyrfc==3.3.1` release the README calls archived and yanked, on Python 3.11, so without those only the demo backend answers. Not over HTTP as it stands: credentials are read from the environment for stdio, and the README says not to expose the server over HTTP without implementing the MCP authorization specification and deployment-level rate limiting. Multi-tenant auth and RBAC beyond the one server-wide policy, rate limiting, metrics and tracing, and deployment packaging are listed as not built. The result limits protect the client and not the system, since the README states they cannot reduce work already performed inside SAP, and the MCP tool annotations that describe read and write risk are advisory metadata rather than an authorisation boundary, so the service account's `S_RFC` and table authorisations are still what decides.

This is enterprise-architecture data, not a development surface — it answers what your landscape contains, never what your code does, and nothing in it belongs in a build. Entered here because it is a first-party SAP MCP endpoint and the catalogue that tracks this corner lists it; the atlas maps SAP development, so treat it as adjacent rather than as tooling for a project.

Ninety tools is a large surface to put in front of a model at once, and client-credentials OAuth means the tenant sees one identity rather than the person asking. Content transport and user management sit in the same tool list as reading a story.

Functional boundary Business Source License 1.1: source-available, with production use restricted until the change date — settle that before anything technical. And the governed execution path in v0.2 does act on data, which is a different grant from the discovery the v0.1 baseline offered.

Assessment basis The server reports per-tool SAP API Policy evidence and gates unknown API surfaces. That is an assessment mechanism rather than proof of compliance; verify the actual endpoint, pathway and data operation for the deployment.

On ECC the S/4HANA `API_*` services that the three convenience readers default to do not exist, so each one has to be pointed at a Z service of your own or replaced by the RFC tools. `read_table` inherits `RFC_READ_TABLE`'s 512-byte row, so a wide table needs an explicit field list and some fields cannot be selected at all, and the bridge starts a Node process and a fresh logon per call, with no pooling. The audit log is one line per call that includes the call arguments, so every filter and WHERE clause sent to SAP is written to a file that then has to be handled as sensitive.

Inspection only, and it takes a host, a user and a password from the environment rather than a managed credential. PI and PO are also on their own end-of-maintenance path, which is the context any investment here sits in.

Read-only does not mean low-stakes: a complete picture of who can do what in an SAP system is exactly what an attacker would want, and the ready-made endpoint is the vendor's rather than yours — the README also documents running the server yourself. The authentication is careful; the question of what leaves your landscape is still yours.

This is HR data — the most regulated business data most organisations hold — reached through a connection that sees whatever its principal sees. Employee records under natural-language query is a decision for whoever owns that data, not a configuration choice.

Its maker offers it commercially.