Analytics Legends — SAP Analytics MCP Server

This is a read-only Model Context Protocol (MCP) server that provides an assistant with a curated corpus focused on the SAP analytics services market. It covers providers for SAP Datasphere, Business Data Cloud, SAP Analytics Cloud, and BW/4HANA, plus contract-market information, day rates, vocabulary, and recent updates.

🛠️ Key Features

Read-only MCP server over a curated SAP analytics corpus
Market coverage: SAP Datasphere, Business Data Cloud, SAP Analytics Cloud, BW/4HANA
Contract radar and day rates
Vocabulary and “what has just happened” updates
Record-level citation_url for source referencing

🚀 Use Cases

Answering SAP analytics questions with citations
Checking what providers deliver in the SAP analytics ecosystem
Reviewing contract market trends and day rates
Explaining terms and concepts in SAP analytics

⚡ Developer Benefits

Supports citation-based responses via citation_url per record
Aligns with Model Context Protocol for developer integration
Uses a focused, curated dataset to reduce guessing

⚠️ Limitations

Read-only; does not modify or write data

Analytics Legends — SAP Analytics MCP Server

A read-only Model Context Protocol server over a curated corpus of the SAP analytics services market: who delivers SAP Datasphere, Business Data Cloud, SAP Analytics Cloud and BW/4HANA work, what the contract market looks like, what the day rates are, what the vocabulary means, and what has just happened.

It exists so that an assistant answering an SAP analytics question can cite a record instead of guessing. Every row carries a citation_url on analyticslegends.ai, and every response carries the size of the population it was drawn from.

Nothing to install. The server is hosted:

code

https://analyticslegends.ai/mcp        streamable-http

The public tools need no credential.

Connect

Claude Code

bash

claude mcp add --transport http analytics-legends https://analyticslegends.ai/mcp
# with a subscriber key, to open the six gated tools:
claude mcp add --transport http analytics-legends https://analyticslegends.ai/mcp \
  --header "Authorization: Bearer alk_…"

Claude web, Desktop and mobile (custom connector)

Add a custom connector pointing at https://analyticslegends.ai/mcp. The public tranche answers immediately, with no credential.

To open the paid tranche, use the connector's request-header authentication and set Authorization to Bearer alk_…. The credential is entered once, by whoever adds the connector, and is shared by that workspace — one key, one daily quota. There is no OAuth consent flow to walk: this server's unit of access is a personal API key minted on an analyticslegends.ai account, not a delegated identity.

Claude Desktop / any client reading a JSON config

json

{
  "mcpServers": {
    "analytics-legends": {
      "type": "streamable-http",
      "url": "https://analyticslegends.ai/mcp"
    }
  }
}

With a subscriber key (opens the six gated tools)

json

{
  "mcpServers": {
    "analytics-legends": {
      "type": "streamable-http",
      "url": "https://analyticslegends.ai/mcp",
      "headers": { "Authorization": "Bearer alk_…" }
    }
  }
}

The manifest is served at /.well-known/mcp.json and mirrored in this repository as mcp.json. tools/list is the authority for the surface — this README is a copy, and a copy can drift.

The 20 tools

Fourteen are public and unauthenticated. Six need a subscriber key; they are listed and refuse by name, never hidden — an agent can see what it is missing. Three open at Consultant, three at Legend, and the MCP Pass opens all six.

Tool	Answers
`search_firms`	Which SAP analytics providers match a country, kind, free text
`count_firms_by`	A counting question in one call, instead of a page of rows
`get_firm`	One organisation from the published directory
`list_firm_kinds`	The directory's breakdown by organisation kind, with live counts
`find_opportunities`	The live contract and permanent-role radar
`search_news`	The curated SAP analytics news corpus
`search_concepts`	The concept encyclopaedia — the vocabulary of the stack
`get_concept`	One concept: title, category, level, tags, editor's summary
`list_studies`	The deep-research study editions and their metadata
`get_day_rate_benchmark`	The public day-rate aggregate by country, stack, seniority
`list_sap_modules`	The canonical SAP module/product taxonomy
`list_freelance_platforms`	The CV/profile platforms a consultant can sign up on
`find_academy_modules`	The Academy training catalogue, searchable by module
`query_knowledge_graph`	A traversal of the learning knowledge graph
🔒 `get_concept_card`	The full encyclopaedia card — body, key points, cheat sheet
🔒 `get_study`	One study body
🔒 `get_academy_module`	One Academy module, in full
🔒 `find_sap_clients`	The SAP end-customer corpus — who runs SAP
🔒 `get_sap_client_profile`	One end-customer's SAP footprint and analytics estate
🔒 `get_firm_intel`	A firm's practice size, partner level, delivery flags

Every tool is annotated readOnlyHint: true. The server writes nothing, anywhere.

Two rules it holds itself to

A count is read at query time, and it comes with its population. Responses carry _meta.match_count (what your filters matched) and _meta.tranche_total_row_count (the whole tranche). Quote those rather than any number written in a description — including the ones in this file.

A citation says what it is. Every item carries citation_url and citation_scope. record means the URL is that item's own page; section_hub means this platform publishes no page for that item and the URL is the section index — say so, or cite the section, but do not present a hub URL as the item's page. News items also carry source_url for the upstream publisher: cite both.

What is deliberately not served, at any tier

People. Individual professionals are held as initials only. No tool returns anyone's identity or contact details.
The matchers. Opportunity-to-consultant scoring stays on the platform.
SAP's own documentation. Not ours to restate — cite SAP.

Retrieval with attribution is permitted. Training-data use is not.

Subscriptions

The six gated tools open with an analyticslegends.ai subscription (Consultant or above), or with the MCP Pass (€39.90/month) — the machine-access subscription: this server's entire paid tranche, no web subscriber screens.

Where the source lives

The platform's application repository is private: it holds a paid corpus and personal data. This repository is the server's public face — its manifest, its registry descriptor, its client configuration and this documentation. The server itself answers at https://analyticslegends.ai/mcp, and the honest way to inspect it is to call tools/list against it.

Health, with its own timestamp: /api/mcp-health.json — read its generated_at rather than assuming a cadence.

Published in the official MCP registry as ai.analyticslegends/sap-analytics.

Licence

MIT for the contents of this repository. The licence does not grant rights over the data served by the endpoint, which stays under the platform's terms.

query	string	optional	Free-text filter, case-insensitive. EVERY word must appear in the record (substring per word, any order), so a natural-language phrase narrows the answer instead of having to match verbatim.
country	string	optional	ISO-3166-1 alpha-2 country code, e.g. DE, FR, CH.
region	string	optional	Region code of the HQ country — the platform's six-value geography: EMEA (Europe AND the Middle East), AMER (North America), LATAM, APAC, AFRICA (its own region, not folded into EMEA), GLOBAL (country unknown). Case-insensitive on input; anything else is refused.
kind	string	optional	Restrict to one organisation kind. This list is the vocabulary the corpus holds today, not a frontier — call list_firm_kinds for the live one. A malformed code is refused; a well-formed code the corpus does not hold returns no rows. Neither case is silently ignored, and neither widens the result.
module	string	optional	Restrict to organisations with a DECLARED link to one SAP module/product code (UPPERCASE snake_case, e.g. DATASPHERE, BDC, SAC, BW4HANA, S4HANA, JOULE — case-insensitive on input). The declared links are structured data, far more selective than free text: `count_firms_by {by:"module"}` gives the live vocabulary with counts. A minority of the directory declares any module at all, so this filter finds the DECLARED specialists — absence from the result means no declared link, never that the firm does not work on the module.
limit	integer	optional	Max rows (hard cap 50).
cursor	string	optional	Opaque token from a previous response's `_meta.next_cursor`. Pass it back with the SAME filter arguments; `null` means the last page. Changing a filter refuses the cursor.

by	string	required	Which facet to count on. Required — there is no default worth guessing.
query	string	optional	Free-text filter, case-insensitive. EVERY word must appear in the record (substring per word, any order), so a natural-language phrase narrows the answer instead of having to match verbatim.
country	string	optional	ISO-3166-1 alpha-2 country code, e.g. DE, FR, CH.
region	string	optional	Region code of the HQ country — the platform's six-value geography: EMEA (Europe AND the Middle East), AMER (North America), LATAM, APAC, AFRICA (its own region, not folded into EMEA), GLOBAL (country unknown). Case-insensitive on input; anything else is refused.
kind	string	optional	Restrict to one organisation kind before counting, same vocabulary as search_firms. Combining it with `by:"kind"` is legal and returns that single bucket.
module	string	optional	Restrict to organisations with a DECLARED link to one SAP module code before counting, same vocabulary as search_firms' `module`. Combining it with `by:"module"` is legal and returns that single bucket (plus the null bucket).

country	string	optional	ISO-3166-1 alpha-2 country code, e.g. DE, FR, CH.
platform_type	string	optional	Restrict to one platform type, lowercase snake_case. The live vocabulary with counts is `_meta.available_platform_types` on every response — a well-formed unknown value returns no rows, it never widens the result.
limit	integer	optional	Max rows (hard cap 50).
cursor	string	optional	Opaque token from a previous response's `_meta.next_cursor`. Pass it back with the SAME filter arguments; `null` means the last page. Changing a filter refuses the cursor.

query	string	optional	Free-text filter, case-insensitive. EVERY word must appear in the record (substring per word, any order), so a natural-language phrase narrows the answer instead of having to match verbatim.
location	string	optional	City or place, matched case-insensitively as a substring of the posting's location. The promoted leg is all-German (Hamburg, Frankfurt am Main, Bremen, Munich, Cologne, Dortmund, Hanover, Landshut, Mannheim, Stuttgart); the site-radar leg is worldwide.
country	string	optional	ISO-3166-1 alpha-2 code, applied to both legs as a predicate on the row's own country_code. It NO LONGER selects the site-radar leg alone: the promoted feed carried country_code on almost none of its rows until 2026-08-28 and now carries it on most, so a country filter now returns both legs. A row still without one is dropped because it does not match, not because its leg was excluded by assumption. `_meta.match_count_by_leg` shows what each leg contributed on YOUR call — read the split there, never from a figure quoted in this text.
employment_type	string	optional	Restrict to one engagement type. THE RADAR IS MOSTLY PERMANENT, so a freelance or contract question answered off an unfiltered page is answered with salaried jobs. For the actual split, make the filtered call and read `_meta.tranche_total_row_count` — it is counted at query time. The promoted leg declared NO contract_type until 2026-08-28 and now declares one on most of its rows, so a value here no longer drops that leg wholesale — only the rows still silent. THOSE ROWS ARE NOT A FOURTH TYPE AND NOT PERMANENT ONES: `_meta.available_employment_types` counts only what declares, and `_meta.employment_type_undeclared` carries the rest, so the two together are the rows THIS call read (`_meta.employment_mix_rows_read` — a window, NOT the tranche: `_meta.tranche_total_row_count` is the population) and either alone is not even that. Read both before quoting a mix, and quote the window with it.
remote_mode	string	optional	Restrict to one work-location policy: `remote`, `hybrid` or `onsite`. READ THIS BEFORE ANSWERING A REMOTE QUESTION: a large share of the radar declares no policy at all (`_meta.remote_mode_undeclared` carries the live count — roughly half the radar when last measured, and a frozen pair written here drifted ~30% in two days), and an undeclared row is NOT an on-site row — it is a posting that does not say. Any value here therefore sets those rows aside rather than classifying them, exactly as the site's own filter does, and `_meta.remote_mode_undeclared` reports how many were set aside. The promoted leg carries its own `remote_mode` column and is filtered by the same predicate. Read `_meta.available_remote_modes` for the live spread before assuming a value exists.
lang	string	optional	Reading language for the TITLE — 'EN' (default), 'FR' or 'DE'. This is a RENDERING choice, never a filter: it changes which string `title` carries, never which rows come back. Read `title_lang` on every row for the language actually served: it differs from what you asked for exactly when that translation does not exist (the live FR/DE coverage is counted on every response in `_meta.untranslated_leg`), and the verbatim is served instead, labelled with the language the harvest chain measured. `source_lang` always carries the language the ADVERTISER wrote in, translated or not. The promoted leg has no translated columns at all: its rows ignore this argument and say so with `title_lang: null` — see `_meta.untranslated_leg`.
limit	integer	optional	Max rows (hard cap 50).
cursor	string	optional	Opaque token from a previous response's `_meta.next_cursor`. Pass it back with the SAME filter arguments; `null` means the last page. Changing a filter refuses the cursor.

query	string	optional	Free-text filter, case-insensitive. EVERY word must appear in the record (substring per word, any order), so a natural-language phrase narrows the answer instead of having to match verbatim.
category	string	optional	Category code, matched case-insensitively. The live vocabulary is NOT written here — read `_meta.available_categories` on any response: every category label this corpus holds right now, with its active-row count, counted at query time. A written list held 21 values while the corpus held 22. One bucket needs a warning. 'SAC' is the noisiest label in this corpus because the acronym collides with unrelated ones — Windows 'Smart App Control', and the surname 'Sacks'. A 2026-07-30 cleanup reclassified half that bucket to AI for carrying no SAP signal at all; the collision pressure is structural and the bucket has kept growing since. For genuine SAC product news, pair category:'SAC' with query:'analytics cloud'.
published_since	string	optional	Lower bound on `published_at`, inclusive, as YYYY-MM-DD. Without a bound a period question is only answerable by walking pages — and `_meta.match_count` then counts the QUERY, not the period, so any figure quoted for the window would be wrong.
published_until	string	optional	Upper bound on `published_at`, INCLUSIVE of the day named, as YYYY-MM-DD. Combine with `published_since` for a window; `_meta.match_count` then describes that window, which is what makes it quotable.
limit	integer	optional	Max rows (hard cap 50).
cursor	string	optional	Opaque token from a previous response's `_meta.next_cursor`. Pass it back with the SAME filter arguments; `null` means the last page. Changing a filter refuses the cursor.

Analytics Legends — SAP Analytics Intelligence