MCP

This page explains what the Anjana Data Platform MCP (Model Context Protocol) is, how it works and what role each component plays when an AI assistant works with the platform. It also answers the most frequent questions about AI, privacy and control, such as whether the MCP includes a language model, whether Anjana Data analyzes conversations or whether its behavior is deterministic.

Installation is described in MCP Installation and day-to-day use in MCP User Guide.

What an MCP is

Model Context Protocol is an open standard that defines how an AI assistant connects to external applications. It works as a universal connector: the application publishes a list of operations with a name, description and parameters, and any compatible assistant can discover and invoke them. The standard is not exclusive to any AI provider.

An MCP server is therefore an integration layer, comparable to an API. It does not think, generate text or make decisions. It receives structured requests (for example, “search for assets containing the word customers”), executes them and returns the result.

What the Anjana Data Platform MCP is

The Anjana Data Platform MCP is the server that exposes the platform's governance capabilities to AI assistants. It translates each request from the assistant into a call to the same APIs used by the Data Portal, mainly those of the governance core (Kerno) and those for search and audit (Minerva).

As a result, a user can ask in natural language for tasks such as finding Catalog assets, viewing their details and lineage, cataloging and enriching metadata, submitting for validation, requesting access to a DSA or reviewing the audit log. The MCP does not replace the Data Portal or open a parallel governance channel: it applies the same rules, permissions, approval workflows and audit logging.

How it works

Four components take part in an interaction with the MCP, each with a different owner:

A request always follows this path:

1. The user writes their request in the AI client, for example “Which sales datasets are approved?”.

2. The client's language model decides which MCP operation it needs and with which parameters, for example searchEntities with the text sales and the status APPROVED.

3. The client sends the MCP server only that structured invocation, not the conversation.

4. The MCP server checks the user's identity and calls the Anjana Data Platform API on their behalf.

5. The platform applies the user's permissions, executes the operation and, if it modifies anything, records it in the audit log.

6. The result returns to the AI client, whose model interprets it and writes the answer for the user.

What it includes and what it does not

Frequently asked questions

Does the Anjana MCP include artificial intelligence?

No. The MCP is an integration layer that executes platform operations when asked to. The artificial intelligence is in the AI client that the organization connects to the MCP: that assistant is the one that understands natural language and decides which operations to invoke.

Does it have or use an LLM?

No. The Anjana MCP server does not contain, host or call any language model. The LLM is supplied by the AI client chosen by the organization, under the terms of the provider it has contracted. Anjana Data SL does not impose or recommend any specific provider.

Is it deterministic?

The MCP server is: the same operation, with the same parameters and on the same platform state, always returns the same result. What is not deterministic is the language model of the AI client. Given the same question, the assistant may choose different operations, broaden or narrow a search on its own or word the answer differently.

This variability is bounded by the platform design:

• Every change starts as a draft, any asset or relationship created or modified through the MCP remains in DRAFT status until it is submitted for validation.

• Approval is always human, the move to APPROVED or REJECTED is decided by an authorized person from the platform.

• The Data Portal is the source of truth, when an exact or reproducible figure is needed, it should be obtained in the Catalog with explicit filters.

Does Anjana Data analyze interactions or conversations?

No. The conversation between the user and the assistant takes place in the AI client and does not reach the MCP server. The server only receives the structured invocations of the operations (operation name and parameters), like any API. Anjana Data SL does not receive the prompts or the assistant's answers.

What is recorded are the governance actions executed on the platform, in the same audit log as the actions performed from the Data Portal. That log belongs to the organization and is accessed with the usual audit permissions.

What information leaves the platform for the AI provider?

The results of the invoked operations return to the AI client, and its model processes them to write the answer. Those results are governance metadata: asset names and attributes, relationships, lineage, DSA information or audit entries, always limited to what the authenticated user can see.

The MCP does not offer operations to read the content of governed data (table records or files). Even so, metadata may be sensitive information for the organization. It is therefore advisable to choose the AI client and provider in line with internal policies, for example through enterprise contracts that exclude the use of data for training or through models deployed on the organization's own infrastructure.

Is the data used to train AI models?

Anjana Data SL does not train models with the information that flows through the MCP. Whether the AI client's provider may use that information depends exclusively on the contract and configuration the organization has with that provider.

Can the assistant do something the user could not do from the Data Portal?

No. Each operation is executed with the identity and permissions of the authenticated user. If the user does not have a role, the assistant cannot act with it either, and restricted information is only returned to those who have the corresponding permission. Actions are attributed to the user, not to the assistant.

Can the assistant approve, reject or delete assets?

No. The MCP allows creating, enriching and submitting for validation, and also requesting or withdrawing an adherence to a DSA. It does not expose operations to approve or delete assets. The governance cycle is always closed by an authorized person.

How do users authenticate?

Each user signs in with their own platform account, using the same login as the Data Portal or the configured SSO, through the OAuth 2.0 standard. There are no shared tokens or technical credentials to distribute among users.

Is it mandatory to use it?

No. The MCP is an optional capability that the organization enables if it wishes. Anjana Data SL provides the capability; enabling it, exposing it on the network and choosing the AI client are the organization's responsibility.

Does it work in environments without internet access?

The AI client must be able to reach the platform's MCP endpoint. In environments that require zero internet exposure, the alternative is to use an AI client or agent that runs inside the organization's internal network, preferably with a model also deployed on that network.

This page is intended for data and AI governance leads, security, architecture and compliance teams evaluating the use of the MCP, and any user who wants to understand how it works before using it.