mcp-sap-datasphere-server

SAP Datasphere MCP Server

v1.0.0 — Read-only. API Policy v4/2026 aligned. Sibling to SAPBDCMCP.

An open-source Model Context Protocol server that lets AI agents — Claude, Cursor, or any MCP-compatible client — safely explore an SAP Datasphere tenant. Discover spaces, list assets, preview rows, profile columns, search by column name, and summarize analytical models — all through a clean, read-only tool surface that respects SAP's API governance posture.

MCP host (Claude Desktop, Cursor, ...)
        │
        ▼
sap-datasphere-mcp  ──>  FastMCP server  ──>  tools/registry  ──>  category facades
                                                                   │
                                                                   ▼
                                          DatasphereClient (httpx)  ──>  /api/v1/datasphere/* (modern)
                                                                   ──>  /api/v1/dwc/* (legacy fallback)
                                          policy / audit / redaction interceptor chain

Highlights

• 24 read-only tools across connectivity, catalog, query, discover, profile, summarize, governance.
• 5 MCP Prompts and 4 MCP Resources — first SAP MCP server to ship either.
• SAP API Policy v4/2026 aligned — documented posture, optional audit log, redaction layer, Integration Suite MCP Gateway as the recommended enterprise deployment path.
• Sibling to SAPBDCMCP — same maintainer, same naming, same license model. Run them side-by-side in the same MCP host.
• No write tools, ever. Hard guarantee for the 1.x line.
• OAuth 2.0 client_credentials (HTTP Basic, with body fallback). mTLS-bound client_credentials via IAS is a documented Tier C posture on the roadmap.
• stdio + streamable-HTTP transports; optional bearer auth on HTTP.

Install — pick one

# uv / uvx — fastest, no global install
uvx mcp-sap-datasphere-server

# pip / pipx
pip install mcp-sap-datasphere-server        # or:  pipx install mcp-sap-datasphere-server

# npm (npx-wrapper bootstraps the Python package via uvx)
npx -y @rahulsethi/sap-datasphere-mcp

The console script is still sap-datasphere-mcp — that's what you run, wire into Claude Desktop, and put in MCP host configs. The PyPI package name is mcp-sap-datasphere-server.

Full install guide and Claude Desktop / Cursor wiring is in public_docs/INSTALLATION.md.

Configure

Four env vars are required, the rest are optional. Copy .env.example or set-datasphere-env.example.ps1 to a local file and fill in:

export DATASPHERE_TENANT_URL=https://<your-tenant>.<region>.hcs.cloud.sap
export DATASPHERE_OAUTH_TOKEN_URL=https://<your-uaa>/oauth/token
export DATASPHERE_CLIENT_ID=<your-technical-client>
export DATASPHERE_CLIENT_SECRET=<your-client-secret>

Optional knobs you'll reach for most often:

Full env table: see public_docs/INSTALLATION.md.

Quick start

sap-datasphere-mcp --version    # sap-datasphere-mcp 1.0.0
sap-datasphere-mcp              # starts the stdio MCP server

Then add this to your MCP host config (Claude Desktop / Cursor):

{
  "mcpServers": {
    "sap-datasphere": {
      "command": "sap-datasphere-mcp",
      "args": [],
      "env": {
        "DATASPHERE_TENANT_URL": "https://...",
        "DATASPHERE_OAUTH_TOKEN_URL": "https://.../oauth/token",
        "DATASPHERE_CLIENT_ID": "...",
        "DATASPHERE_CLIENT_SECRET": "..."
      }
    }
  }
}

Try a first prompt: "Show me the SAP Datasphere spaces in my tenant."

A richer one: "Profile the EMPLOYEES asset in the HR_SPACE space." — this drives our profile_dataset MCP Prompt automatically.

Tool catalog

24 tools organized as:

• Connectivity (5) — ping, diagnostics, tenant_info, whoami, plugins_status
• Catalog (5) — list_spaces, list_assets, get_asset, list_columns, space_overview
• Query (3) — preview, relational, analytical
• Discover (3) — assets, assets_with_column, assets_by_column
• Profile (2) — schema, column
• Summarize (4) — asset, space, column_profile, compare_assets
• Governance (2) — api_policy_check, audit_tail

Every tool is prefixed datasphere_<category>_. Every tool is readOnly; none are destructive. Full reference with args + return shapes: public_docs/TOOLS.md.

MCP Prompts (first in the SAP ecosystem)

MCP Resources

datasphere://space/{space_id}
datasphere://space/{space_id}/asset/{asset_name}
datasphere://space/{space_id}/asset/{asset_name}/schema
datasphere://space/{space_id}/asset/{asset_name}/sample

Let your MCP host pre-load context without a tool call.

Architecture (high level)

The boot path is transports/stdio_server.py → server.create_server() → tools.registry.register_all() → category facade modules → the 22 async implementations in tools/tasks.py. Each tool call goes through a small interceptor chain: policy gate → audit start → tool → redaction → audit commit.

Full architecture details and decision log are in public_docs/ — see the quick links.

API Policy v4/2026 posture

This server consumes only SAP-documented Datasphere Catalog and Consumption APIs via OAuth 2.0 client_credentials (the technical-user pattern). It does not scrape, reverse-engineer, or proxy non-public APIs. It is read-only by construction — there are no write/admin tools and there will never be any in the 1.x line.

For production deployments, we recommend wrapping this server behind the SAP Integration Suite MCP Gateway, which inherits SAP-sanctioned metering, agent identity, audit, and rate limiting. See public_docs/SAP_API_POLICY.md for the full disclosure, the three-tier deployment guide, and the datasphere_governance_api_policy_check tool you can call to print the deployment's current posture.

Family

SAPBDCMCP — the same maintainer ships a sibling MCP server for SAP Business Data Cloud at https://github.com/rahulsethi/SAPBDCMCP. Same naming conventions (<product>_<category>_<verb>), same license model, same read-only enterprise posture. If you operate across Datasphere and BDC, run them side-by-side in Claude Desktop and switch between them mid-conversation — neither your agent nor you have to context-switch.

A 2-for-1 family discount is available on commercial licenses for both.

Migration from v0.3.x

In short:

pip install --upgrade mcp-sap-datasphere-server

The PyPI distribution name is unchanged (mcp-sap-datasphere-server) — upgrade in place. The 22 old tool names continue to work through v1.1 as deprecation aliases. Update your Claude Desktop / Cursor configs to the new datasphere_<category>_<verb> names before v1.2 (~Q4 2026). Full guide with the rename table: public_docs/MIGRATION.md.

Versioning

Current: 1.0.0 — SemVer-stable through the 1.x line. Tool names, argument shapes, prompt names, and resource URIs are public contract.

Recent releases:

• 1.0.0 — graduation: family alignment, API path migration, governance layer, MCP prompts + resources, BSL 1.1 relicense. (This release.)
• 0.3.1 — cleanup: examples folder, dead code removal, docs reorg.
• 0.3.0 — analytical querying, deterministic summaries, TTL cache, configurable caps.
• 0.2.0 — metadata, discovery, profiling extensions, mock mode.
• 0.1.0 — initial GitHub release.

Full changelog: CHANGELOG.md.

License

Business Source License 1.1 for v1.0+. Versions v0.3.1 and earlier remain MIT-licensed. Converts automatically to Apache 2.0 on 2029-01-01.

• Personal, research, academic, and internal evaluation use: free.
• Commercial use (paid consulting, embedding in a vendor product, hosted SaaS): requires a separate commercial license.

The path is friendly and the terms are reasonable — see COMMERCIAL_LICENSING.md for how to get one.

Contributing

Contributions accepted under the same license. No CLA required. See CONTRIBUTING.md.

Issues / contact

• Bugs / feature requests: https://github.com/rahulsethi/SAPDatasphereMCP/issues
• Commercial licensing: open a Discussion with title [Commercial License Inquiry] <your-org>.