Skip to main content

​
Atoti Intelligence SDK

This is part of the Atoti Intelligence SDK offer.
The Atoti MCP Server exposes Atoti’s cube tools to external MCP clients, once it is set up. This page covers the behavior shared by both SDKs; setting the server up is SDK-specific.
To set up the MCP Server, see How to set up the Atoti MCP Server in Java or How to set up the Atoti MCP Server in Python. For what the MCP Server is and why to use it, see What is the Atoti MCP Server?.The Atoti MCP Server requires the Atoti Intelligence Extension tier, which itself requires Essentials.

​
Where is the MCP endpoint?

On a Java application, the MCP Server answers POST /mcp. On a session started from the Atoti Python SDK, it answers f"{session.url}/mcp". Both speak the Streamable HTTP transport. Spring AI’s own default transport is the older server-sent events transport, SSE. Atoti overrides that default: a property contributed at the lowest precedence sets spring.ai.mcp.server.protocol=STREAMABLE. An application that sets the property itself always wins over it. The server-sent events transport at /sse is deprecated for removal in Spring AI 2.0.0. Prefer /mcp everywhere.

​
Which license tier does the MCP Server require?

The Atoti MCP Server requires a license including the Atoti Intelligence Extension tier, which itself requires Essentials. The Extension tier gates the whole MCP surface:
  • The MCP Server itself
  • Its OAuth 2.1 discovery and self-issued authorization-server modes
  • The credentials page
  • Connecting out to other MCP Servers
Essentials alone is enough for the cube tools to exist. Chat and Auto-Explain already call them internally under Essentials. The Extension tier does not add new tools. It exposes the existing ones externally, at the MCP endpoint, to any MCP client.

​
What happens when the license does not include the Extension tier?

When the license lacks the Extension tier, Atoti forces spring.ai.mcp.server.enabled=false at the highest property precedence. The property cannot be overridden, even when set to true the forced value false wins. A warning is logged naming the ai-extension license component. A license without Essentials does not start the server.

​
How does an MCP client authenticate?

The MCP endpoint and the deprecated /sse endpoint sit behind their own security filter chain. It requires an authenticated user holding a user or admin role. Each SDK implements this separately.
  • On a Java application: the required roles are the combined set of configured admin roles and user roles.
  • On a session started from the Atoti Python SDK, the required roles are ROLE_USER or ROLE_ADMIN.
Only an authenticated user holding one of those roles is admitted. OPTIONS requests are permitted without authentication, for CORS preflight. The accepted credentials are the ones Atoti’s machine-to-machine security allows everywhere: HTTP Basic, and a bearer JWT. An unauthenticated request is answered 401 Unauthorized. When OAuth 2.1 discovery is enabled, that response additionally carries WWW-Authenticate: Bearer realm="mcp", resource_metadata="...". A compliant MCP client reads that header to discover the authorization server. It then drives the browser PKCE flow with no manual token handling. See Configure OAuth 2.1 discovery and Configure self-issued OAuth 2.1 for the two modes. The opt-in MCP credentials page is the alternative to embedding a password in a client configuration. A signed-in user mints a long-lived bearer token once, then pastes it into the client.

​
Which tools does the MCP Server expose?

The MCP Server exposes the same built-in cube tools that chat and Auto-Explain use internally:
  • Discovery tools, for cube names, dimensions, hierarchies, levels, members, measures, and KPIs
  • Validation tools, for checking that a measure, level, hierarchy, or dimension exists
  • MDX query tools, including runMdxQuery and getAvailableCubeNames
  • Auto-Explain tools, for running an analysis, listing analyses that already ran, and dropping one
Custom tools add to that set. The tools from other MCP Server can be exposed by the Atoti server they are connected to. This means that: One tool is the exception. getConnectedServers is registered for chat only, never on the MCP endpoint. It puts deployment-internal addresses into the prompt. No LLM is required to expose or call these tools. An LLM is only required for chat.

​
What does each SDK decide?

Server set up follows the same pattern for both Java and Python set up:
  1. Install a package
  2. Supply minimal or no configuration
  3. The server is available.
However, Atoti Java SDK and Atoti Python SDK do not share the same configuration: