sac-workbench-mcp

sac-workbench-mcp

An unofficial Model Context Protocol server that lets an AI assistant (Claude Desktop, Claude Code or any other MCP client) work with SAP Analytics Cloud: build and change stories, charts, KPIs, themes and CSS, create planning models and dimensions, write plan values, load data, and use the public REST APIs.

⚠️ Disclaimer — read before use

• This project is not affiliated with, endorsed, sponsored or supported by SAP SE. "SAP" and "SAP Analytics Cloud" are trademarks of SAP SE; they are used here only to describe what the software works with.
• Large parts of this server use undocumented, internal SAP Analytics Cloud interfaces (the requests the SAC web UI itself sends). They were found by recording the UI, not from SAP documentation.
• These URLs, payloads and behaviours are not a stable API. They may change or break with any SAC update, without notice.
• SAP does not support this software or any change it makes to your tenant. Do not open SAP support tickets for problems caused by it.
• Use at your own risk. Test in a sandbox tenant or folder first, keep backups, and review what the assistant is about to change. The authors accept no liability (see LICENSE).

📖 Featured Blog Posts

sac-workbench-mcp is built at NextLytics, a consultancy for data platforms, analytics and planning. More on SAP Analytics Cloud on the NextLytics blog.

1. Trick or Theme: Dark Mode in SAP Analytics Cloud — a use case built with this server: the dark theme, the story CSS and the dashboard in the post were created and tuned through the MCP server. https://www.nextlytics.com/blog/trick-or-theme-dark-mode-in-sap-analytics-cloud
2. Controlling SAP Analytics Cloud with Claude: a home-built MCP server — the introduction to this server: what it does, how it works and what we learned building it. https://www.nextlytics.com/blog/controlling-sap-analytics-cloud-with-claude-a-home-built-mcp-serve

Sister project for SAP BW/4HANA: bw-modeling-mcp — Agentic AI meets SAP BW.

What's new — v0.3.0

First public release. Highlights: complete dashboards in one call (on Datasphere models and on SAC models — planning, Seamless and analytic), declarative chart editing (type, feeds, filters, hierarchy, colours), restricted and calculated measures, KPI tiles as Numeric Point, layout checks per screen width, planning models with dimensions, hierarchies and data reloads, and Datasphere model search. See CHANGELOG.md.

What it can do

178 tools in these areas, plus 3 expert tools for extending the server (see below) — the full list with parameters is in TOOLS.md, how it works inside in ARCHITECTURE.md:

Try it

Once the server is connected, just ask in plain language. A few prompts to start with (use your own model and folder names):

Build me an overview dashboard on the model SALES_PLAN in my folder "Test": four KPI tiles for revenue and quantity, actual vs. plan, two bar charts by region and by product, and a table by product. Use the Horizon template and show the numbers in millions.

Which models can I build stories on? Show me the dimensions and measures of SALES_PLAN.

Create a dark theme for our stories: dark grey background, white text, blue as accent colour, and apply it to the story "Sales Overview".

In SALES_PLAN, raise the 2026 plan for Germany by 5 % in the private version and show me before and after — don't publish yet.

The button on page 1 of "Sales Overview" should filter the chart by the selected year. Write the script and check the story for errors.

The assistant asks for whatever it needs (model, theme, layout) and states every change before it makes it.

Guidance for good stories (layout, naming, minimal look, colours) is in docs/STORY_GUIDELINES.md and is also offered to the client as the MCP prompt story_guidelines.

Skills for scripting and CSS

skills/ contains two Agent Skills — sac-scripting (SAC scripts against the official API reference) and sac-css (story/theme CSS with only the classes and properties SAC accepts). They work with this server and on their own. The scripting reference is built locally from SAP's public documentation; see skills/README.md.

How it works

• Browser session (main path). The server starts its own Chrome profile with Playwright. You sign in once via SSO (npm run login); the session is kept in the profile. Requests are sent from inside the signed-in SAC page, with the CSRF handling the UI uses. No passwords are stored or entered by the server.
• OAuth client (optional). No tool needs one: without it, the documented public REST APIs (/api/v1/... — data import/export, users and teams, transport, schedules, calendar, translation, audit) also go through the browser session. Configure an OAuth client (client credentials) only if those calls should run as a technical user. Everything else always uses the browser session.
• Recorded templates. templates/ contains payloads recorded from the SAC UI with placeholders; the server fills in your tenant, user, folder, model and ids at runtime.

Requirements

• Node.js 22 or newer
• Google Chrome (or set SAC_BROWSER_CHANNEL)
• An SAP Analytics Cloud tenant and a user with the rights you want the assistant to use. For Datasphere models a live connection to SAP Datasphere.

Setup

From npm

npm install -g sac-workbench-mcp
sac-workbench-mcp setup     # asks for the tenant URL and defaults, opens Chrome for the SSO login

The configuration goes to ~/.sac-mcp/config.json, so updates (npm install -g sac-workbench-mcp@latest) keep it. Sign in again later with sac-workbench-mcp login. With Claude Code: claude mcp add sac -- sac-workbench-mcp.

From source

git clone https://github.com/RafaelSachsNL/sac-workbench-mcp.git
cd sac-workbench-mcp
npm install
npm run setup     # asks for the tenant URL and defaults, builds, opens Chrome for the SSO login

Recordings, snapshots and data backups go to captures/ in the repository folder (git-ignored).

npm run setup writes sac-mcp.config.json (see sac-mcp.config.example.json) and can register the server in Claude Desktop. Manual registration:

{
  "mcpServers": {
    "sac": { "command": "node", "args": ["/absolute/path/to/sac-workbench-mcp/build/index.js"] }
  }
}

With Claude Code, one line does the same:

claude mcp add sac -- node /absolute/path/to/sac-workbench-mcp/build/index.js

On Windows use the full path with escaped backslashes in JSON, e.g. "args": ["C:\\Users\\you\\sac-workbench-mcp\\build\\index.js"].

Sign in again later with npm run login.

Tested against SAP Analytics Cloud 2026.21 (Optimized Story Experience). SAC updates can change the internal interfaces; the smoke test (see Development) shows quickly what broke.

Configuration

sac-mcp.config.json (lookup: $SAC_MCP_CONFIG → <repo>/sac-mcp.config.json → ~/.sac-mcp/config.json); environment variables override it.

Troubleshooting

Safety

• Every tool that changes something requires allowalteration: true. Without it the tool refuses, so the assistant has to ask for and state the change.
• readOnly: true blocks all writes, whatever the assistant sends.
• sac_resource_delete moves files to Deleted Files (recoverable). Some public-API tools delete for good (e.g. sac_users_delete, sac_transport_delete_content, sac_import_delete_job) — their descriptions say so.
• sac_story_transform saves a snapshot before every change; sac_story_restore brings it back. sac_model_data_reload writes a CSV backup of the facts before replacing them.
• Planning values are written to the private version first; publishing (sac_plan_publish) is a separate step.
• Red/green series colours are refused until confirmed (confirmSignalColors) — they are signal colours.
• captures/ (recordings, snapshots, backups) can contain business data and is git-ignored. Do not commit it.

Recommendation: give the assistant a test folder (defaults.folderId) and test models first.

Development

npm run build                                              # TypeScript → build/
npm run smoke -- <folderId> [model] [connection] [space]   # end-to-end test against your tenant (creates and deletes test objects)
SMOKE_FORECAST="<seamless model>,<row dimension>,<measure>" npm run smoke -- …   # includes the forecast planning story

Extending the server (expert tools)

Three tools give raw access to the tenant and are meant for developing new tools, not for everyday use. They are off by default and only loaded with "expertTools": true in sac-mcp.config.json or SAC_EXPERT_TOOLS=true:

They bypass the guard rails of the regular tools (snapshots, validation, signal-colour check), so switch them on only in a sandbox tenant while you develop.

To teach the server something new, record it: sac_capture_start, do the action in the SAC UI (sac_browser_ui or by hand), sac_capture_stop, inspect with sac_capture_get, try the changed call with sac_replay, then turn the payload into a tool. For stories, sac_story_snapshot → change in the UI → sac_story_diff shows what SAC stores. See CONTRIBUTING.md. Releases: docs/RELEASE.md.

Acknowledgements

The public REST API tools (data import/export, users, transport, schedules …) started from JumenEngels/sap_analytics_cloud_mcp (MIT). The story, chart, theme, planning and model tools were added on top.

License

MIT. Provided "as is", without warranty of any kind — see the disclaimer above.