Data governance
7 min

Semantic Layer and MCP: What the Agent Actually Receives

Key takeaways

  • Over MCP, a metrics-layer server returns a metric name, a typed description, compiled SQL or rows, and whatever the signed-in user's warehouse grants allow. A governed-meaning server returns the agreed definition, its named owner, its version, the permitted joins, and the entitlement applied, with the number.
  • The Model Context Protocol defines the envelope only: a tool name, a description, a schema, and a result. Which definition the agent receives, who approved it, and whether it is current are left to the server.
  • dbt, Cube, Snowflake and Databricks all document the same shape in their own words: a discovery tool that lists metrics or models and a query tool that runs them. Each is accurate about what it ships.
  • After the call, the agent still cannot tell which definition it used, who signed it, whether it has expired, which joins were off limits, or whether a concept-level rule applied. A second server that chooses the definition is one answer. One record that holds definition and calculation together is the other.
  • Evaluate a semantic layer's MCP server on the raw payload: definition or description, owner present, version present, joins bounded, entitlement on the concept.
In this article
Share

A platform lead connects the company's first agent to the semantic layer over MCP. The demo works. The agent lists the metrics, picks net revenue, and returns a number by region. Then someone asks which definition it used, who approved it, and whether the caller was allowed to see the refund detail. The semantic layer and MCP question starts there. The protocol carried the request perfectly. What came back is the subject of this article.

MCP, the Model Context Protocol, is the open standard for connecting an AI model to tools and data. A semantic layer MCP server exposes the layer's definitions as tools an agent can call. Every vendor on this query now ships one. They differ in what a tool call returns, and that difference decides what the agent can defend.

MCP standardizes the ask. The server decides the answer. A metrics-layer server returns a metric name, a description, compiled SQL, and rows. A governed-meaning server returns the definition, its owner, its version, the permitted joins, and the entitlement applied, with the number.

What does MCP do for a semantic layer?

The MCP specification (revision 2025-06-18) defines a tool as a name, a description, an input schema, and an optional output schema. A call returns a content array and, optionally, a structured JSON object. That is the whole envelope. The protocol says nothing about what a metric is, who owns it, or whether it is current. Our post What Is a Semantic Layer? lists what a layer has to hold for an agent. MCP is how the agent asks for it.

The envelope, and what it leaves to the server

Part of the protocolWhat the specification definesWhat it leaves open
Tool definitionname, description, inputSchema, and an optional outputSchemaWhether the description is a typed sentence or a governed definition
Tool resultA content array, plus optional structuredContent that "MUST" match the output schemaWhich fields are in it
Resource annotationsaudience, priority, and lastModified, all optionalWho approved the content and whether it is still in force
TrustClients "MUST consider tool annotations to be untrusted unless they come from trusted servers"Where the trust comes from

Source: the specification's tools and resources pages. One catalog vendor's page on this query (published 2026-10-12) reads it the same way: "MCP standardizes how an agent asks a semantic layer for a metric. It leaves open which definition the agent should receive." That is correct. The page's remedy is a second MCP server, from that vendor, which picks the definition while the semantic layer's own server calculates it. In its words, "Neither server takes over the other's job." The open question is whether the definition and the calculation should have been two objects at all.

What does a metrics-layer MCP server return?

The four servers people search for with this term all document their tools. They describe one shape: a discovery tool that lists what exists, and a query tool that runs it.

Four servers, in their own words

ServerDiscovery toolQuery toolWhat comes back
dbt MCP server (docs updated 2026-11-20)list_metrics, which "Retrieves all defined metrics"; get_dimensions; get_entitiesquery_metrics, which "Executes metric queries with filtering and grouping options"; get_metrics_compiled_sqlPer dbt's launch post, "metric names, types, labels, and descriptions", then compiled SQL and results
Cube MCP server (read 2026-11-22)searchDataModel, which returns "compact records (name, title, description, type)"runQuery, which returns "a schema, a page of rows, and hasMore / totalRows"Names, descriptions, rows. "Every tool runs as the authenticated user"
Snowflake-managed MCP server (read 2026-11-22)One tool declared per semantic view; the server "only supports using semantic views"A semantic view tool that "generates SQL from natural language text"The SQL statement. Access needs "SELECT on Semantic View"
Databricks Genie Agent MCP server (Public Preview, 2026-10-23)A Genie Agent of "up to 50 Unity Catalog tables, views, or metric views"A plain-language question; "Genie writes and runs the SQL against that agent's tables and returns the result"The result, "with Unity Catalog permissions enforced on every request"

The open-source Boring Semantic Layer (version 0.3.19, released 2026-11-11) does the same inside a Python package, with tools for "Listing available models", "Getting model metadata", and "Querying models with dimensions, measures, and filters". Every one of these pages is accurate about what it ships. dbt adds that its MCP tools "can change without notice or a deprecation period. Don't treat them as a contract."

So the agent receives a name, a description somebody typed, a type, and either compiled SQL or rows. Our post Metrics Layer vs Semantic Layer: The Difference Is a Noun explains why that is the natural output of a layer built from metrics. A metrics layer can define a calculation. It cannot define a concept.

What does the agent still not know after the call?

Picture the agent after the two tools return. It holds a metric named net_revenue, a description, and a table of numbers. Five things it needs are missing. The catalog vendor's page agrees on most of them: "MCP carries no field for staleness or certification."

Five questions the payload does not answer

  • Which definition? The description is free text. Finance and Sales may each have a net revenue, and the payload cannot say which this is.
  • Who approved it, and when? In dbt, an owner is a free-form meta key or a group setting, and changes to meta "don't trigger state:modified". The GraphQL Metric type carries a description and no owner field.
  • Is it still in force? No tool description above mentions a version or a review date. The protocol's lastModified is an optional annotation on a resource.
  • Which joins are allowed? Entities are join keys. The agent is told what it may group by, never what it may not traverse.
  • What may this caller see? Cube and Snowflake enforce the signed-in user's grants, which is right. Grants sit on tables and views. A rule about a concept, such as compensation wherever it appears, has nowhere to live.

The agent's side of this is our post A Semantic Layer for AI Agents: What the Agent Needs Before It Asks for Revenue. This article stays on the wire.

What does a governed-meaning MCP server return instead?

Here is the same request against a semantic layer built on an ontology, a formal model of the business's concepts and how they relate. The left column is assembled from the vendor tool docs above. The right column is what TopQuadrant's live agent-ready page says is served.

Two payloads for one question

FieldMetrics-layer serverGoverned-meaning server
Namenet_revenueNet revenue, a measure of the Order concept
DescriptionA free-text fieldThe agreed definition, the one its owner approved
OwnerA meta tag or group setting, if someone set itA named person, on the record
Version and dateGit history, outside the payloadVersion, status, and the date it last changed
CalculationCompiled SQLThe same logic, held on the same governed asset as the definition
MappingImplicit in the SQLConcept to table to column, "held as data you can query"
Permitted joinsEntities as join keysOnly modelled relationships; "the join paths available to them come from the ontology"
EntitlementThe warehouse grant of the signed-in userEvaluated per request on the concept, for the identity the agent acts for
Record of the answerServer logs, if kept"the definitions resolved, their versions, the entitlements applied, and the data returned"

The live page's own example: Active account, version 3, current: an account with at least one posted transaction in the trailing 90 days, excluding fee reversals and internal transfers. Owner, Retail Finance. Changed 2026-05-14. A BI pack, a dbt model, and an assistant over MCP all read it. Change it once and all three move.

That is the sentence the two-server architecture cannot write. Where the definition and the calculation are one asset, there is nothing left for a second server to choose. What the approval adds is covered in our post Ground Truth: SME-Approved vs Inferred Metadata. The MCP answer carries it.

Is a refusal enough?

dbt's own benchmark (dbt Developer Blog, 2026-04-07) makes the best case for a metrics-layer server: "the Semantic Layer tells you it can't answer. It never returns invalid data." The same page states the limit: the layer "can only answer questions that fall within the scope of what's been modeled" (dbt Developer Blog). A refusal is safe when the definition it refused on was the agreed one. Inside scope, the payload above is still what the agent gets. Our post Semantic Layer vs Text-to-SQL: What a Refusal Is Worth takes that question in full.

How do you evaluate a semantic layer's MCP server?

Connect a client, call the discovery tool, call the query tool, and read the raw result. Then ask five questions of the payload.

Five checks on the raw payload

  1. Definition or description? An approved definition with a scope, or a free-text field someone filled in?
  2. Owner in the payload? A named person, or a tag that may or may not be set?
  3. Version in the payload? Can the agent tell a current definition from an expired one without opening git?
  4. Joins bounded? Does the server say what the agent may not traverse, or only what it may group by?
  5. Entitlement on the concept? Does a rule about compensation hold across every table it appears in?

A server that passes all five is serving authoritative context: meaning that named people agreed to, in a form a machine can act on. The discipline behind that is our post Context Management: The Complete Guide. A server that passes none is a metrics layer with an MCP endpoint, which is fine for a dashboard.

Where does TopQuadrant fit?

TopQuadrant ships the governed record and serves it over MCP. Our agent-ready semantic layer page says what the agent receives: concepts, permitted relationships, and metric logic. The agent resolves what a term means before composing a query. Every metric and entity carries its business definition, its owner, its calculation, and its mapping to physical columns. Access rules attach to governed concepts and are evaluated per request against the identity the agent is acting for. Every served answer records the definitions resolved, their versions, the entitlements applied, and the data returned. It is also served over API and SPARQL.

AI proposes definitions and mappings, and your experts confirm them. Each definition is validated with SHACL, on open standards TopQuadrant helped author, so the whole record is yours to export. Your catalog stays, and it syncs.

MCP carries the request. TQ Data Foundation carries the record back with the number. That is what the agent actually receives.

Give your AI the context it's been missing

See how the TQ Data Foundation turns your enterprise knowledge into trusted, Al-ready context.