An industrial MCP server can give an AI application a structured way to request information from connected systems. For an OT architect, the important question is not simply whether the connection works. It is whether the application can obtain the specific data needed for an approved task, without exposing unnecessary assets or operations.
That decision requires a clear view of the application, the MCP server, the industrial data platform and the boundaries between them. It also requires checking what the returned data means and how current it is. MCP provides a way to exchange context; it does not by itself establish that the connection is secure or that an AI-generated conclusion is correct.
Start with a bounded industrial question
Consider a hypothetical packaging oven with recurring temperature excursions. An engineer wants to review temperature history and related events for that oven over a defined period. The OT architect is evaluating whether an AI application can help assemble the information for review.
The proposed investigation needs the oven identity, measurement units, timestamps, a bounded history of the relevant temperature, and references to applicable events. It does not need to change a setpoint, acknowledge an alarm or control the oven. The architect can use that distinction to define an initial read-oriented scope before deciding which connections and tools are necessary.
This is an architecture exercise, not a description of a tested deployment. The engineer must still check the source data and the resulting explanation. A temperature trend can support an investigation without establishing why an excursion occurred.
How does an AI application reach industrial data through MCP?
The Model Context Protocol, or MCP, uses a host, client and server arrangement. The host is the AI application coordinating the interaction. An MCP client within the host maintains a connection to an MCP server. The server exposes selected capabilities or context; it may in turn obtain information from an industrial data platform.
For the oven investigation, the conceptual path is:
Engineer → AI host and MCP client → MCP server → authorized industrial data source
The response travels back through the application for the engineer to review. The server may run locally or remotely, depending on the architecture. The data source, server, client and model do not have to reside in the same network zone. Map each connection explicitly rather than assuming that the MCP server is the industrial platform itself.
Tools, Resources and Prompts
An MCP server may expose different kinds of capabilities:
- Tools are operations a client can request, such as retrieving a time-bounded history when the server provides that operation.
- Resources provide context that a client can access, such as available metadata or definitions exposed by the server.
- Prompts are reusable templates for interactions that a client may offer or use.
The actual server determines what it exposes, and the client determines how it presents or invokes those capabilities. Not every server offers all three. For the architect, the practical task is to inventory the capabilities available in the intended deployment and decide which ones the investigation needs.
Check the meaning and freshness of the returned data
A temperature value is not enough on its own. The application and the reviewing engineer need to know which oven and sensor produced it, its unit, its timestamp and the requested time range. Related events need their own time references and enough context to distinguish an event record from a confirmed physical cause.
Check whether the returned values are recent enough for the question. A recent request can still return stale source data, delayed updates or a period with missing readings. If the application summarizes the result, the engineer should be able to return to the relevant source records and identify gaps or conflicting information.
The TDengine industrial data architecture overview discusses the relationship between data storage, industrial context and AI access. Whatever platform is used, the architect should verify which asset relationships, history and events are actually accessible through the selected interface.
Define identity, permissions and network boundaries
A read-oriented design is an intentional choice. It is not a property to assume from the presence of MCP or from a server’s name. Work with the OT and security teams to review each part of the path.
Identify the caller and its authority
Determine which identity the client uses to authenticate to the server, which identity the server uses to reach the data source, and who owns their credentials. Authentication establishes identity; authorization determines what that identity may do. Check both layers, including access to the intended oven and any broader asset hierarchy.
Inventory every exposed tool and its effective permissions. A tool described as useful for investigation may still allow an operation that the oven scenario does not require. Check whether the implementation supports restricting tools and asset scope, and test the restrictions with the actual client and identity. If the proposed read-only boundary cannot be demonstrated, the design has not yet been validated for this use case.
Read access can also expose sensitive production information or place load on the source. Define appropriate time ranges, query scope and operational limits with the data owner, and assess what information may be returned to the application or model.
Map the network path and review sensitive operations
Document the network zones crossed by the client, MCP server and data platform, including direction of traffic and the systems responsible for each boundary. Confirm encryption, credential handling, monitoring and operational ownership against site policy. Do not assume that a protocol, a private network location or a confirmation prompt alone provides an adequate security boundary.
If a later use case requires changing an analysis, acknowledging an event or taking another sensitive action, treat it as a separate design decision. The responsible people should review the actual tool, permissions, consequences and site procedure. AI output remains subject to human engineering review and does not authorize industrial control.
IDMP MCP and TDengine TSDB MCP are different interfaces
TDengine documents two distinct MCP interfaces with different scopes.
The TDengine IDMP MCP interface describes access to industrial elements, time-series attributes, events and analyses, as well as controlled write operations within applicable permissions. Its documentation lists different requirements for individual tools. The available operations and permissions vary by tool and deployment, so an architect should verify the effective authorization and scope of each operation. Do not assume IDMP MCP is read-only by default.
The separate TDengine TSDB MCP server is documented as providing read-only database queries and metadata operations, without data writes, deletion or schema modification. That documented scope does not automatically apply to IDMP MCP, and it should not be taken to mean that TSDB MCP exposes the same industrial asset and event context.
For either interface, verify the deployed version, configuration, credentials and source access. A read-only interface can still return data that the caller should not receive, so data scope and identity remain part of the review.
Review the oven investigation before relying on it
Before testing, the architect and the relevant OT and security owners should be able to answer:
- Scope: Which oven, sensor, events and time range are needed, and which asset access is permitted?
- Identity: Which credentials are used at each connection, and what authority do they carry?
- Capabilities: Which tools are exposed, which can write, and how is the proposed read-oriented boundary enforced and tested?
- Evidence: Are units, timestamps, freshness, source references and data gaps available to the engineer?
- Operations: Who reviews the output, monitors the integration and handles failure or unexpected access?
Use a representative, authorized test to compare the application’s output with the source records. If the response is incomplete or the access boundary cannot be verified, record the limitation and resolve it before relying on the integration. The resulting summary can help an engineer investigate, but it is not a substitute for evidence or a physical root-cause assessment.


