Siemens MCP Server

This page answers four questions for every tool on https://mcp.siemens.com/mcp: where the data comes from, how much of it a tool can see, how fresh it is, and what it does not cover. Agents and their builders should read it before trusting an empty or a surprising result.

The statements are grounded in the plugin code under src/plugins/ and src/services/siemens-graphql.ts, and in read-only measurements against the live backends on 2026-11-04. Numbers marked as measured will drift; the mechanisms will not unless the code changes.

Summary

All backends except SiePortal OneSearch are called anonymously (authorization: anonymous, or no credential at all for the datasheet service). SiePortal OneSearch uses the OAuth client of the New Global Search federation (see "SiePortal articles" below) and returns the public SiePortal catalog only. The server therefore sees what the public internet sees, never more. There is no connection to the Siemens intranet, SharePoint, Teams, hello.siemens.com, Confluence, Jira or any employee-only system.

Search: siemens_search (and the deprecated siemens_aiSearch, siemens_productSearch)

Source

siemens_search queries the Siemens Search GraphQL API at https://api.dc.siemens.com/search, the same index that powers the search box on siemens.com, and SiePortal OneSearch for articles. The schema is vendored in src/plugins/siemens-search/schema.graphql. When the server runs with any stage other than prod (for example npm run dev), it calls the UAT index at https://uat.api.dc.siemens.com/search instead, so local results can differ from production.

The scope parameter selects what comes back:

total_results has one count per returned list (pages, products, articles, partners).

siemens_aiSearch and siemens_productSearch are deprecated aliases kept for clients that still call them by name: they return the same lists as scope pages and products in their previous shape (total_results a number, total_articles). They are not in tools/list (nor on /docs, in the OpenAPI description or the agent card), so clients that read the tool list use siemens_search; calls by name still work. They will be removed once the telemetry shows no more calls.

SiePortal articles

The products lists hold product family pages on siemens.com, which carry no article number. The articles lists come from SiePortal OneSearch and hold orderable articles: article_number (MLFB), description and the SiePortal detail URL. Free text ("SITOP PSU8200 24 V 20 A", "S7-1500 CPU 1511") and partial article numbers ("6ES7511") both match. Pass an article_number to siemens_product_datasheet for its technical data.

• Source: the OneSearch gateway https://apim.pods.siemens.cloud/onesearch/api/v1/documents, the same federation the New Global Search API runs for siemens.com (SieportalFederalClient in new-global-search-api), with its OAuth client (client_credentials against auth.sieportal.siemens.com, scope mall mulesoft) and its allowlisted User-Agent. Code: src/services/sieportal-client.ts.
• Credentials: Secrets Manager siemens-mcp/<stage>/sieportal-api (client_id, client_secret, token_url; placeholder from src/plugins/siemens-search/secrets.json). Without a populated secret the articles lists are empty and total_results.articles is 0; scope articles then answers with an error. A running container re-reads a missing secret after 1 minute and a valid one after 15 minutes (rotation pickup).
• Region: GLOBAL searches worldwide (ww), a country code searches that country. Language defaults to en.
• Counts: total_results.articles is the gateway hit count, passed through unchanged (paging uses it). Measured 2026-11-05: "SITOP PSU8200" worldwide 16, "6EP3436-8SB00-0AY0" 1. For free text the count also includes articles that match only some words of the query: measured 2026-11-07 worldwide, "SITOP" 241 but "SITOP 24 V 20 A power supply" 198,244 and "power supply" 143,440; the ranking still puts the matching articles first. Above 1,000 hits (LOOSE_ARTICLE_HIT_COUNT) the hint says that the count is a raw hit count and not the number of matching articles, and the search view shows "N best matches" with the hit count as a note instead of "N of 198,244".
• Failure: a SiePortal failure leaves articles empty and names "SiePortal articles" in the hint ("Partial result: ..."); the other lists are still returned.

Xcelerator Marketplace partner offerings

The partners lists hold third-party offerings sold through the Siemens Xcelerator Marketplace: per item title, url (the offering page on siemens.com), description, partner (the seller), industries (the first 5; all with response_format: "detailed") and product_type when the page states one ("Software", "Hardware", "Service", "Consulting", "Training", "Software: SaaS-native", ...). Detailed results add partner_url (the partner's seller page) and seller_regions (the countries and regions the partner sells in) from one extra seller-page lookup. Every item belongs to its partner, not to Siemens; the hint tells the agent to attribute it to that partner and never call it a Siemens product, as siemens_get_web_content does for the same pages (see "Provider attribution").

• Source: the same index as the siemens.com Ecosystem Explorer (/<locale>/products/), queried anonymously like the other searches (code: src/services/marketplace-client.ts). Offering pages are mediatype PRODUCT, tenant ONE_SIEMENS_COM, page type offering; the seller is the page's LABEL. Seller pages are page type seller and carry the seller regions; offering pages carry none.
• Partner, not Siemens: the index also lists Siemens' own offerings with the same page type. Offerings whose seller is Siemens, Siemens Industry Software, Siemens Advanta or Global Business Services (GBS) are Siemens products and never appear in partners; every other seller is a partner (including Siemens Energy and Siemens Healthineers, separate companies). The API has no exclude filter, so the search sends an include filter on the partner seller labels, read from a LABEL facet. Measured 2026-11-06 (EN, region US): 2,271 offering pages, about 1,500 of them by the four Siemens units, 744 partner offerings from 381 partner sellers.
• industry: one value of the INDUSTRY facet, case-insensitive, 109 values (EN/US), for example "Automotive", "Machine building", "Pharma and life science", "Healthcare", "Food and beverages", "Commercial buildings", "Chemicals". An unknown value is a validation error that lists the values the index uses.
• sellerRegion: one value of the SELLER_REGION facet of the partner seller pages, case-insensitive, 194 values (EN/US): countries ("Germany", "USA", "United Kingdom", "India", "China") and regions ("North America", "Central Europe", "Asia-Pacific", "Latin America", "Middle East"). Because offering pages carry no seller region, the filter keeps offerings of partners whose seller page lists the region.
• facets: industries are the top 15 industries of the current query and filters (partner offerings only); seller_regions are the top 15 regions by number of partner sellers, independent of the query. Measured 2026-11-06 (EN/US): Germany 210 partners, USA 160, United Kingdom 144, India 134. Facet counts are approximate index counts and do not add up exactly to total_results.partners.
• Counts: total_results.partners is the hit count of the query. Measured 2026-11-06 (EN, GLOBAL): "digital twin" 428, "predictive maintenance" with industry "Automotive" 69, "access control" with sellerRegion "Germany" 262; DE/DE "Energiemanagement" 84.
• Cache: the partner seller labels, the industry values and the seller-region values are read with two facet calls per language and region and cached for one hour per container; a seller-region filter adds one cached seller lookup per region. The offering search itself is never cached.
• Product lists: products (the catalog search) also contains offering pages of partners. Such a product item carries partner with the seller's name, and its text line says "partner offering by ..., not a Siemens product".
• Not covered: offering pages without a LABEL (11 of 2,271 measured) are missing from partners, prices and availability, and partner-specific data behind the marketplace login.
• Failure: in scope all a marketplace failure leaves partners empty and names "partner offerings" in the hint; scope partners then answers with an error.

Coverage

The index groups content into tenants (see TenantEnum in the schema). The vendored schema lists 21 tenants, among them the main siemens.com site, Siemens Mobility, Press, Advanta, Mendix, SITRAIN, SiePortal, Siemens Jobs, Siemens Ausbildung, Brightly Software, Solid Edge, Siemens Stadt, Siemens News, Siemens Trials and Siemens Software events. Live results also carry tenants missing from the vendored schema (for example BLOGS_SW, the Siemens Software blogs), so the live index is broader than the schema file suggests. Tenants seen in the measured samples: Siemens Mobility, SITRAIN, Siemens Jobs, Siemens Software blogs and SiePortal.

The language and region facets for a broad query ("siemens") returned 32 languages and 78 regions. Coverage is strongest for English: in that sample 69 of 150 hits were English and 10 were German.

The product catalog is SiePortal. Products carry the tenant SIEPORTAL, a product id, a title, a description and a URL to the SiePortal page.

How to read result counts (volume)

The counts in total_results mean different things per source. Do not add them up and do not treat them as index size.

• Web content (total_results.pages): the web index uses semantic top-k retrieval and caps every query at 150 hits. Every broad query measured ("siemens", "automation", "simatic s7", "mendix", "order management service") reported a count of exactly 150, and paging past offset 150 returns no hits. A count of 150 means "at least 150 candidates", not "150 documents in the index".
• Products (total_results.products): the catalog search reports the real number of matching catalog entries. Measured examples: "siemens" 265,198, "automation" 24,611, "simatic s7" 1,140, "order management service" 95, "mendix" 0.

Apart from the facets of scope partners (industries and seller regions of the marketplace), no tool returns facet counts or the size of the index. The API supports facets (SearchFacetsGlobal, fields such as MEDIA_TYPES, LANGUAGES, REGIONS, LABEL), and the client in src/services/siemens-graphql.ts has a searchFacets method, but no tool exposes it.

Relevance scores

relevance_score (returned with response_format: "detailed") is the raw score from the backend. Web scores and product scores come from different rankers and live on different scales (measured: web hits roughly 0.8 to 7, product hits roughly 80 to 240). Compare scores only within one result list and one source. A high web score does not mean the hit answers the question; it means it is semantically close to the query text. Queries for internal concepts (for example "order management service") match public pages that share the words, such as job postings, because the internal documents are not in the index.

Filters and parameters

• region GLOBAL (the default) is sent to the API as region US, because the API has no GLOBAL value. Content that exists only on a country site may be missing; pass the country code (for example DE) to target it.
• language defaults to EN.
• siemens_search accepts query, scope, industry and sellerRegion (scope partners only; other scopes reject them), language, region, summaryLevel, maxResults, offset (ignored for scope all) and response_format.
• siemens_search tolerates partial failure: if some backends fail, their sections are empty, the others are still returned, and hint names the unavailable parts ("Partial result: ..."). An empty products list without that note means "no products".

Freshness

The server does not cache search results: every call goes to the live index with cache-control: no-cache. Freshness is therefore that of the Siemens Search crawler, which this project does not operate. The API exposes a lastCrawled field, but it was empty on every measured hit, and created is present only for some web pages and absent for products. The tools cannot tell you how old an indexed page is.

Fallback behavior

If the search API fails or is unreachable, the tools return a tool execution error (isError: true) that says what failed and what to do. They never return placeholder data.

siemens_assets

Source

The Siemens Assets GraphQL API at https://api.dc.siemens.com/assets, the digital asset management (DAM) system behind assets.new.siemens.com.

Coverage

• Every query is restricted to the Web Asset Pool (metaDataFilter: { wap: true }), the set of assets cleared for public web use. Restricted, internal or embargoed assets are never returned.
• collection selects one DAM source: ASSETS (main collection), PRESS (press media) or ENERGY (Siemens Energy). ALL, the default, currently searches only ASSETS; it does not merge the three. The ENERGY_PRESS source in the schema is not exposed.
• assetType narrows to images, videos, documents, audio or icons.
• usageRights, imageSize and orientation filter by the DAM restriction level, the image resolution group and the orientation. WAP does not mean free to use: many WAP assets are limited or restricted, and some need approval.
• Each asset carries its per-asset usage_rights from the DAM licenceData (restriction level, approval, credit line, expiry, legal notes), a 640 px preview and the DAM download variants with their real sizes (images: small, medium, large, original; videos: sd, hd, full_hd).

Volume and freshness

total_results is the real number of matching assets (measured: "automation" 29,294). Results come from the live DAM with no caching in this server. Asset metadata and rights reflect the DAM at query time.

siemens_get_web_content

Source

Pages on siemens.com and its subdomains are read through the CommsCloud Markdown API (https://markdown-api.c2comms.cloud, MARKDOWN_API_URL, https only, off disables it). The API renders the page in a headless browser (Selenium Grid), accepts cookie consent, converts the DOM to Markdown with YAML frontmatter, converts PDFs, reads SiePortal pages behind Akamai and caches results (memory, filesystem, S3). Measured on 2026-11-05: cache hits 0.09 s, fresh www.siemens.com pages 3.1-3.5 s, support.industry.siemens.com 5.4 s, developer.siemens.com 3.4 s. support.industry.siemens.com answers 403 to direct requests from the Lambda; through the API it returns the page.

backend names the path that produced the content:

Backend selection, in order:

1. The tool's own URL checks run first and stay authoritative: https, port 443, no user info, allowed domains (see Host policy). A URL they reject is never sent anywhere.
2. render: false reads over plain HTTP only: no Markdown API and no browser. The API has no option to skip its browser, so this is the only no-browser path.
3. Otherwise a siemens.com URL goes to GET /convert?format=json&allow_async=true with a timeout of at most 12 s. A slow conversion (PDF, SiePortal) answers 202 with a job id; the tool long-polls GET /jobs/<id>?wait=<s> (at most 20 s per poll) while the request budget leaves at least 2 s to answer. If the job is still running then, the tool returns isError with a retry hint: the job continues in the background and a repeated call is served from the API cache (measured 0.09 s after a 4.3 s job).
4. A timeout, network error, 5xx, failed job, unusable answer or bot-check page falls back to the direct path: plain HTTP, and AgentCore Browser in automatic mode when the HTTP result has fewer than 40 words (the shell of a page built with JavaScript, for example https://www.siemens.com/en-us/products/ with about 20 words), or with render: true. A render starts only when its whole 14 s deadline plus a 1.5 s margin fits the time left in the request (#108), so after a slow API failure the HTTP result is returned instead.
5. A 400 from the API (domain not served), a page the API saw as 404 or 410, an answer above 5 MB, and a final URL outside the allowed domains are reported as errors without a fallback.

On the direct path the main content (<main>, otherwise <body>) is converted to Markdown deterministically with turndown, without a language model. No Bedrock model is called on any path; the Markdown API is called without its LLM features (summarize, question, extract_ctas).

Provider attribution

siemens.com also hosts third-party offerings sold through the Siemens Xcelerator Marketplace. Their pages look first-party (the title ends in "| Siemens", and the page's own JSON-LD names Siemens as producer), but the product belongs to a partner. The Markdown API states the provider; the tool returns it as provider (name, is_siemens, is_marketplace_offering). When is_marketplace_offering is true and is_siemens is false, the hint (outside the untrusted-content block) tells the agent to attribute the product, its capabilities, pricing and support to the partner and not to call it a Siemens product. provider: null means the page states no seller: unknown, not Siemens, so no attribution claim. The provider name is page-derived data; control characters, <, > and backticks are removed and it is cut to 120 characters. datasheet_url is the datasheet PDF of a SiePortal product page, otherwise null. Direct-path results always carry provider: null.

Coverage

• Only https URLs whose host is siemens.com, siemens.de or siemens.net, or a subdomain of them. Other domains are rejected, and redirects must stay on these domains. In the browser, every document request (page, redirect hop, iframe) is checked before it is sent.
• The HTML an anonymous client receives, after JavaScript when rendered. Pages behind a login are not covered.
• On the plain HTTP path the host must resolve to public addresses only (see Host policy below). A host with any private, loopback, link-local, CGNAT, multicast or reserved address is refused before a request is sent, on the first URL and on every redirect hop.
• HTML or Markdown API answers above 5 MB are rejected. A plain HTTP request has an 8 s timeout and a render a 14 s deadline, so both fit the 25 s request budget; at most 6 renders per minute and Lambda instance (WEB_CONTENT_RENDERS_PER_MINUTE). In automatic mode a failed, rate-limited or unaffordable render returns the HTTP result instead. Long pages are returned in chunks of maxLength characters (default 12,000); next_offset continues reading. Markdown API results include the API's YAML frontmatter at the start of the markdown.

Host policy

Decision: the tool keeps the domain wildcard (siemens.com, siemens.de, siemens.net and every subdomain) instead of an explicit host list, and checks the resolved address on every HTTP request.

• Why the wildcard: the tool exists to read pages that siemens_search and siemens_developer_search return, and those results span many Siemens subdomains. Measured over the 30 days up to 2026-11-04 on the production Lambda: 358 calls; the 104 calls whose log line carries the URL went to www.siemens.com (90), support.industry.siemens.com (5), sieportal.siemens.com (4) and single calls to other hosts, while search results in the same period linked to further subdomains such as news.siemens.com, resources.sw.siemens.com, www.mobility.siemens.com, docs.sw.siemens.com and developer.siemens.com. An explicit list would refuse legitimate follow-up reads whenever a business publishes on a new subdomain. The three registered domains are owned by Siemens, so the wildcard is a domain allowlist in the sense of the OWASP SSRF Prevention Cheat Sheet, not an open fetch.
• What protects internal targets: a Siemens name can still point at an internal address (a split-horizon or mistaken DNS record). Before each request, including each redirect hop, the hostname is resolved with all of its addresses, and the request is refused if any of them is in a blocked range. The blocklist is the one the Developer Portal uses for portalUrl (src/utils/public-address.ts).
• DNS rebinding: the connection is pinned to the IPv4 address that was checked. The HTTP client receives the IP literal in the URL and the original hostname in the Host header, so it does not resolve the name again; TLS still sends the hostname as SNI and verifies the certificate for it. A resolver that answers with a public address for the check and an internal one afterwards cannot move the connection. A host with IPv6 addresses only is refused, because the Lambda has IPv4 egress only.
• Behind an HTTP(S) proxy (local development, the Lambda has none) the proxy resolves the name, so the check runs but the request is not pinned.
• Browser path: AgentCore Browser resolves names inside the AWS-managed browser session, not in the Lambda, so this server cannot check or pin those addresses. The browser path is covered by the domain and https checks on every document request (page, redirect hop, iframe) and by the isolation of the AgentCore session, which has no access to the Lambda's network, credentials or instance metadata. In automatic mode the page is fetched over HTTP first, so its host has passed the address check before a render starts; render: true skips that fetch.
• Markdown API path: the API resolves and fetches the page on its own infrastructure, so the address check and pinning apply to the direct path only. The tool sends the API only URLs that passed its domain, https, port and user-info checks, and refuses a result whose final URL leaves the allowed domains. The API output is third-party content and is wrapped as untrusted like every other page.
• Accepted residual risk: the Lambda runs outside a VPC, so even an unblocked internal name has no route into a Siemens network.

Freshness

The page is fetched at call time and cached in memory for 10 minutes per Lambda instance, so reading further chunks does not refetch it. The Markdown API has its own cache (TTL configured by the API, reported by GET /), so a Markdown API result can be older than the call.

siemens_developer_search

The portal has no search API. It publishes its MkDocs search index at https://developer.siemens.com/search/search_index.json (measured: about 7,500 sections in 51 product areas). The tool loads that index, caches it for 6 hours per Lambda instance and ranks sections locally by title, tag and body matches, one result per page. Changelogs rank lower unless the query asks for them. product narrows to one or more product areas. If the index cannot be loaded, the tool returns isError: true. Freshness is that of the published index; read the current page with siemens_get_web_content.

An optional portalUrl selects another portal instance, such as a merge request review deployment. Only https hosts on DEVELOPER_PORTAL_ALLOWED_HOSTS (default developer.siemens.com,developer.code.siemens.io) that resolve to public addresses are accepted, redirects are not followed, and access-controlled instances (a login redirect) return an error. Review instances are cached for 5 minutes.

siemens_product_datasheet

Source

The HTML datasheet of the TED datasheet service at https://apim.industry.siemens.cloud/ted/datasheet (format=HTML, caller=SiePortal), the same table SiePortal shows under "Technical data" and the Industry Mall prints as the PDF datasheet. It is called without any credential; the gateway has a shared limit of 550 requests per minute. The table is parsed deterministically into sections and properties (name, the sub-headings above it as path, value as printed with its unit, and number plus unit for plain "number unit" values), together with the eClass, ETIM and IDEA classes per version, the certificates and the "last modified" date of the data. No language model is involved and no value is derived or estimated. Use of the service for this server was approved on 2026-11-04 (issue #17).

Coverage

• One article number (MLFB) per call, for example 6ES7214-1AG40-0XB0. Languages en, de, fr, es, it and zh; other languages are not offered by the tool.
• Measured on 2026-11-04: full tables for SITOP power supplies (134 properties), SIMATIC S7-1200 CPUs (243) and SIRIUS circuit breakers (164). SIMOTICS S-1FK2 servo motors and configurable SINAMICS S210 drives have no table in the HTML datasheet; the tool returns hasTechnicalData: false and the PDF link (datasheetPdfUrl), and never fills in values.
• An unknown article number returns isError: true ("Product does not exist" from the service).
• Not covered: prices, availability, delivery times, lifecycle and successor data (the SiePortal Mall API ProductInformation/GetProductsAndPrices answers 403 to the federation client), the content of the PDF datasheet, and search across products by property (find an article number with siemens_search and scope articles).

Freshness

The datasheet is fetched at call time and cached in memory for 6 hours per Lambda instance and language (at most 200 entries). lastModified is the date the data was last changed at the source.

Prompts and resources

• Prompts siemens_check_power_budget (sum the 24 V DC current of devices and compare it with a power supply rating) and siemens_compare_products (property table of several articles). Both are plans over siemens_search (scope articles) and siemens_product_datasheet; they add no data source and require every value to come from a datasheet.
• Resource siemens://docs/scope: this page as Markdown.
• Resource templates siemens://product/{articleNumber} and siemens://product/{articleNumber}/section/{section} (resources/templates/list): the English datasheet of an article number, or one section of it by its exact name (case-insensitive), as text/markdown. The content is the same Markdown summary siemens_product_datasheet returns as text, from the same source and cache; the structured JSON stays with the tool. Variables are percent-encoded (6EP3436 8SB00 as 6EP3436%208SB00, a comma in a section name as %2C). An invalid or unknown article number or an unknown section is "resource not found": -32002 for the initialize-based protocol versions, -32602 for 2026-07-28, with the available sections in the message. A failure of the datasheet service is -32603 with a request id.
• Resource links: siemens_product_datasheet and every article of siemens_search add a resource_link content item (uri, name, title, mimeType, description) to the datasheet resource, after the text items. Text content and structuredContent are unchanged. Links are sent for protocol versions 2025-06-18 and later only; for 2025-03-26, 2024-11-05 and requests without MCP-Protocol-Version they are left out.
• Resources ui://siemens/search-view.html and ui://siemens/datasheet-view.html: the interactive views below.

Completions

completion/complete covers prompt arguments and resource template variables (the MCP completion utility has no tool-argument completion). The completions capability is declared in initialize and server/discover. Every completion is best effort: it returns an empty list, never an error, when the input is too short, SiePortal is not configured or the upstream call fails.

• articleNumber of both product templates and the prompt arguments supply (siemens_check_power_budget), loads (siemens_check_power_budget) and products (siemens_compare_products): one SiePortal article search (the source of siemens_search scope articles) for the typed text, from 3 letters or digits on. The values are the article numbers among the top 30 hits that start with the typed text (spaces, -, . and / ignored), sorted, at most 10; total and hasMore describe only that list, not all matching articles. For the comma-separated loads and products, the last entry is completed and each value is the whole argument with a leading quantity such as 4x kept. Measured on 2026-11-06: 6ES7214 returned 10 values, 6ES7 511-1 5 and 6EP3436 8, in 270-800 ms. SiePortal also lists configurable article numbers with placeholders (1FK7011-5AK2.-1..3); they are suggested as found.
• section of the section template: the section names of the datasheet of the articleNumber in context.arguments (names that start with the typed text first, then names that contain it, in datasheet order, at most 10). Without an article number in the context it returns nothing.
• criteria of siemens_compare_products has no completion.

Interactive views

siemens_search and siemens_product_datasheet declare an MCP Apps view (extension io.modelcontextprotocol/ui, specification 2026-01-26) in _meta.ui.resourceUri. Hosts with MCP Apps support render the result in a sandboxed iframe; every other client gets the same content and structuredContent as before.

• Search view: the answer with its sources, articles and products as cards, pages as a list, each linked, with the counts the tool returned.
• Datasheet view: article number, description, links to the PDF datasheet and the SiePortal page, every section as a collapsible table, classifications and certificates. The datasheet service returns no product image, so the view shows none.
• Comparison: a view belongs to one tool call. When the host offers serverTools, the datasheet view adds up to three more article numbers through tools/call of siemens_product_datasheet (proxied by the host to this server) and shows the products side by side, optionally only the differing values. In a narrow view the table scrolls sideways (a keyboard-focusable region, with scroll shadows and a hint while it overflows) and the property column stays in place. Without it, comparing is left to the model (prompt siemens_compare_products).
• The views are self-contained HTML (inline CSS and script) and render only the tool results the host passes. They load nothing from the network (empty CSP domain lists), escape all text and link only absolute https URLs. They follow the host theme variables and light or dark mode.
• The server is stateless and does not read the client's capabilities per request, so the metadata is always present; hosts without MCP Apps support ignore it, as the specification requires.

Not covered by any tool

• Siemens intranet, hello.siemens.com, SharePoint, Teams, OneDrive, Confluence, Jira, code repositories
• Customer-specific data: prices, availability, stock, orders, contracts
• Content of downloadable files (PDF manuals, CAD files): search can return the file URL, but no tool reads the file
• Websites outside siemens.com, siemens.de and siemens.net
• Siemens Healthineers and other separately operated sites, unless they are indexed under one of the tenants above

Keeping this page current

When a plugin changes its backend, filters or limits, update this page in the same pull request. The facts above can be re-measured with read-only GraphQL calls against https://api.dc.siemens.com/search and https://api.dc.siemens.com/assets, or with tools/call against https://mcp.siemens.com/mcp.