Everything in BTP/BAIP

It stops at the `AuthInfo` and the raw bearer token, and the README lists what stays with the consuming server: rate limiting, the scope and tool policy, the MCP transport, the SAP HTTP client's headers, and helmet-style hardening. The one piece of hardening it constrains rather than defers is the cross-origin opener policy, which it deliberately leaves unset because popup-based OAuth breaks under it. It is ESM-only on Node 22 or later and shares the host's Express and MCP SDK instances as peer dependencies, so a server on Express 4 cannot take it.

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.

Any run nobody is watching: the drawing tools work only while the draw.io canvas at `http://localhost:3000/` is open and connected, so there is no diagram to produce on a build agent or a server. Nothing checks the result either, because the maintainer keeps a diagram linter on the [roadmap](https://github.com/etosin/arc-draw#-roadmap) rather than in the release. The maintainer marks the status early at v0.x, installs it from source, and records testing with Claude Code alone.

Archived The declared repository is archived.

Functional boundary With one SAP system the README says to point the client at that arc-1 directly. Version 1 exchanges tokens within a single subaccount, so backends spread across subaccounts are out; it keeps its session map in memory, so more than one instance needs sticky sessions or a shared store. It routes deterministically and runs no model, which is a different thing from an assistant that reasons across systems. The `/all` endpoint drops the per-connection isolation, and the maintainer's answer to that is a PROD backend set read-only at the SAP end rather than a tool surface that hides the writes.

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.

Its README says it plainly and so does this entry: this is not a simulation, every tool call reaches a live BW system, and write operations produce real changes. BW modelling objects are the definition of a warehouse — the blast radius of a wrong one is every report built on it.

The README tells you to fork before cloning, because Claude rewrites `references/` and `SKILL.md` in place as you use it, so a clone shared between people changes under all of them. Installing means `git` and `ln` on macOS or Linux and a symlink into `~/.claude/skills`. Fine positioning stays manual: the coordinates Claude writes can be rough apart from the centre coordinates of elements joined by a straight arrow and the containment relationships.

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.

There is nothing to clone here: "Access is provided through a registration page. Once registered, you will receive access to the GitHub repository containing the agents, deployment guides, and documentation", and that repository answers 404 to anyone who has not registered, so neither the agent code nor the deployment guides can be read before the form is filled in. AWS states the agents are published as an unsupported solution, with customers responsible for testing and securing the code before deploying in production, and that they are intended for non-production SAP environments rather than production systems. The documentation link in the README is still a placeholder.

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.

It answers about the object catalogue, not about your code: nothing here reads your system, so it will tell you a table is not released and not that you used it. That check is ATC's, and this is what makes ATC's answer actionable.

It stops before the artifact: in the README's own words it does not generate or deploy complete iFlow artifacts, so building and shipping the integration flows themselves stays with a person and a tenant. Knowledge beside the work rather than a generator, and it reads nothing from a running tenant.

An early alpha release whose README says to pin the version and test recipes against non-production data first. HANA synchronisation is snapshot-only and does not reconcile deleted rows, so a document for a row deleted in the source stays in the index. A recipe's field allow-list does not decide what the service account may read: the README states that field allow-listing does not replace SAP authorization, and that the package infers no SAP module semantics, authorization scope or safe fields from table names. Only Business Partner has a built-in recipe, every other dataset needs a `KnowledgeRecipe` written for it, and [Qdrant](external:qdrant) local mode is stated to be for development and smaller indexes.

Its README calls it early-stage — the core workflows run end-to-end against real systems and packaging, documentation and polish are stated as still evolving. A Rust binary rather than an npm install, so the platform build matters.