An AI agent that asks a semantic layer for net revenue should get the same number the finance dashboard shows. MCP, the Model Context Protocol, is how more and more of those requests travel. What the protocol does not carry is any judgment about which definition is certified, who owns it, or whether the caller should see it. Atlan is the Context Layer for AI, and its MCP server supplies that governed context beside the semantic-layer response.
This page covers what MCP does for a semantic layer, why vendors ship MCP servers, what those servers leave out, and how a context-layer server fits next to them.
| What MCP gives a semantic layer | A standard interface for agents to discover metrics and request values |
| What each vendor’s server covers | The definitions and permissions inside that vendor’s own environment |
| What the docs we reviewed leave open | Certification status, cross-system ownership, and which of several competing definitions applies |
| Where a context layer fits | Governed definitions, lineage, and permissions that the agent reads alongside the metric result |
What does MCP do, and not do, for a semantic layer?
MCP standardizes how an agent asks for a metric. It leaves open whether the definition that comes back is appropriate for that agent and that request. When an agent calls the Atlan MCP server, Atlan’s documentation says it can search enterprise context, traverse lineage, and retrieve governed definitions and glossaries with certified information, all within the caller’s Atlan permissions.
The Model Context Protocol is the connectivity layer in this architecture. Anthropic introduced MCP on November 25, 2024 as an open standard for two-way connections between data sources and AI tools. The MCP specification defines three server primitives, and each maps onto a semantic layer:
| MCP primitive | What it does | Semantic-layer example |
|---|---|---|
| Tools | Executable functions the model can call | A query tool that returns net revenue by region for a quarter |
| Resources | Structured data that gives the model context | The list of available metrics, dimensions, and entities |
| Prompts | Pre-defined templates that guide an interaction | A template for comparing a metric across two periods |
A semantic-layer server can expose the same capabilities to any compatible client, from Claude to an IDE to a custom agent. That reuse is why MCP matters for AI agents. The implementation trade-offs sit in MCP vs. function calling and when to use MCP vs. an API.
Reuse does not settle what each caller may receive. The specification makes authorization optional. When a server supports it over an HTTP-based transport, it follows OAuth 2.1, and servers on a stdio transport should instead take credentials from the environment (MCP authorization specification). That mechanism tells a server who is connecting. It does not choose which metric, which definition version, or which lineage that caller should receive. The MCP architecture deep dive shows where each responsibility sits, and MCP gateway vs. context layer explains why reaching a server and trusting its answer are separate problems.
The semantic layer defines shared business meaning for metrics such as revenue or churn. It is a broader job than the calculation-focused one in semantic layer vs. metrics layer. Putting those definitions behind MCP lets agents run the same metric logic that BI tools run, which explains why vendors are building the servers.
Why is every semantic-layer vendor shipping an MCP server?
Agents that infer business logic from raw tables can calculate a metric differently from the BI tools people already trust. An MCP server lets the agent use the joins, filters, dimensions, and time rules the semantic layer already maintains, so it stops reconstructing them from schema. That shared calculation path attacks the “dashboard says X, AI says Y” problem directly.
Two recent findings show why semantic context matters for agents:
- Accuracy gains from semantics. Rumiantsau and Fokeev tested three frontier models on 100 natural-language questions. Adding a 4 KB semantic document to the database schema raised accuracy by 17 to 23 percentage points, from a 45.5 to 50.5% baseline to between 67.7 and 68.7% (arXiv, 2026).
- Analyst forecast. Gartner predicts that by 2027, organizations that prioritize semantics in AI-ready data will increase agentic AI accuracy by up to 80% and reduce costs by up to 60% (Gartner, 2026).
As of 2026, MCP servers span the main semantic-layer and warehouse-native options. The list below comes from each vendor’s own documentation, reviewed on 2026-10-12:
- dbt Labs. The dbt MCP server lists semantic-layer tools including
list_metrics,get_dimensions, andquery_metrics, plus discovery tools such asget_lineage(dbt available tools). See the dbt MCP server in more depth and dbt Semantic Layer vs. Cube. - Cube. Cube documents an MCP server with tools such as
searchDataModelandrunQuery, authenticated through OAuth with each tool running as the signed-in user (Cube MCP docs). It builds on Cube’s semantic-layer architecture. - Kyvos. The Kyvos MCP server exposes three tools: one lists semantic models, one returns column metadata including tags and descriptions, and one executes a query. It supports basic and OAuth 2.0 authentication (Kyvos MCP docs).
- AtScale. AtScale describes an MCP server that lets agents explore and query deployed semantic models and serve governed results to any client (AtScale). For architecture context, see AtScale vs. Cube.
- Looker. The Looker-managed MCP server lets agents interact with business data and LookML models, uses OAuth 2.1 with PKCE, and applies the authenticated user’s Looker roles (Looker MCP docs). It raises the question of LookML versus a universal semantic layer.
- Snowflake. The Snowflake-managed MCP server supports Cortex Analyst with semantic views, and access runs through role-based privileges on the server and on each tool (Snowflake MCP docs). See the MCP server for Snowflake and Unity Catalog’s semantic layer for warehouse-native approaches.
Each server brings consistency inside its own semantic model. None of them reconciles definitions across an enterprise that runs several of the best semantic-layer tools while other metrics stay inside BI tools and operational systems. Semantic layers serve BI and AI agents differently, which is why the next question is what these servers leave out.
What does a semantic layer’s own MCP server leave out?
A semantic layer’s MCP server tells an agent what a metric means inside its own model. The documentation we reviewed on 2026-10-12 describes three questions that sit outside it: who is acting and for what purpose, where the number came from before it reached the model, and whether this is the definition the organization has certified.
| Question | What the documented servers cover | What the documentation does not describe |
|---|---|---|
| Who is asking? | Looker, Cube, and Snowflake document OAuth sign-in and permissions for the authenticated user inside their own systems | Which agent acts for which user and purpose across other tools, and which enterprise policies apply across those systems |
| Where did this metric come from? | dbt documents lineage tools over its own project; Kyvos returns column tags and descriptions | Business logic applied before the data reaches the semantic layer, such as opportunity-stage rules in a CRM, and calculations recreated in downstream BI tools |
| Is this the definition to use? | The model and version currently deployed in that environment | Whether another team or system maintains a competing version, and which one is certified for this use case |
None of the server documentation we reviewed lists a certification status for a metric or a decision trace explaining why an agent picked one definition over another. That is a statement about the pages we read, not about every product version, so check the current docs for the server you run.
Competing definitions make the gap concrete. Take pipeline coverage. Sales operations might count all open opportunities, while the BI team excludes those below a confidence threshold. Both definitions can be current and correct for the team that wrote them. An agent that sees two servers, or one server and a dashboard, has no signal for choosing between them unless something supplies the context that says which one applies to this request.
Two comparisons sharpen the boundary. A semantic layer defines and calculates metrics, while a catalog inventories and describes data assets, as laid out in semantic layer vs. data catalog and context layer vs. data catalog vs. semantic layer. Even an MCP-connected catalog does not by itself reconcile competing definitions.
Where does Apache Ossie fit?
Open standards are starting to address portability. Apache Ossie, formerly Open Semantic Interchange, announced its entry into the Apache Incubator on July 10, 2026, and says its participating coalition has grown from 17 launch partners to more than 50 organizations. It defines a vendor-neutral format for business metrics, dimensions, and relationships, as well as broader business concepts and rules. That broader reach touches ontology versus semantic layer and what an operational ontology is.
Ossie and MCP solve different problems. Ossie standardizes how semantic definitions are represented so they can move between tools. MCP standardizes how an agent discovers and requests capabilities from a server. Neither one decides which definition an organization has certified for a given agent, user, and purpose, and the next section covers who does.
How do context-layer and semantic-layer MCP servers work together?
A context-layer MCP server helps the agent choose and apply the right metric definition, and the semantic-layer server calculates it. The semantic-layer server exposes executable metrics and query capabilities from its model. The context-layer server supplies governed business definitions, lineage, and ownership that can span several models and systems.
This is a reference architecture, not a sequence MCP defines. The agent or its host coordinates calls between the two servers in four steps:
- Frame the task. The agent identifies what it is doing, for whom, and what the answer needs: the business concept, reporting period, dimensions, intended use, and whether a certified or current definition is required.
- Identify and validate the metric. The agent searches the semantic-layer server for a matching metric. It then asks the context-layer server about that metric’s governed definition, ownership, lineage, and applicable permissions before anything is calculated.
- Calculate consistently. The agent calls the semantic-layer server with the validated metric, dimensions, filters, and time range. The semantic layer applies its maintained formula, joins, and aggregation rules, then runs the query against the data platform.
- Return a traceable answer. The agent completes the task and attaches the definition, owner, and lineage it retrieved. Its orchestration logic decides which supporting details appear in the response.
Neither server takes over the other’s job. The context layer supplies permission-scoped business context, and the semantic layer applies the selected metric logic. Access to the metric and the data under it stays with the semantic-layer implementation and the data platform’s permissions. That split underlies the context-layer and semantic-layer split and an enterprise context layer that agents reuse across systems. Context architecture for AI agents and what makes an AI agent context-aware go further, and permission-scoped access in a context layer covers how the permission side works.
How does Atlan provide the context layer for this architecture?
Atlan connects context from warehouses, pipelines, BI tools, and business systems into its Enterprise Data Graph. Through Atlan MCP, agents can search that context, follow lineage, and retrieve governed definitions and glossaries. Atlan’s documentation states that every operation respects existing Atlan permissions, and that OAuth ties each session to the signed-in user while an API key acts as a shared identity for automation (Atlan MCP docs).
Each gap from the previous section maps to a part of Atlan:
| Gap | Atlan capability | What it gives the agent |
|---|---|---|
| Servers cover only their own environment | Context Lakehouse | One store that serves governed context through MCP, A2A, SQL, and APIs beside the semantic layer’s own server |
| Definitions need a governed lifecycle | Context Engineering Studio | A place to build, test, and govern definitions as versioned Context Repos |
| Lineage reaches past one semantic model | Data Lineage | Column-level provenance across pipelines, transformation models, and BI tools |
| Business knowledge scattered across documentation | Context Agents | AI teammates that organize documentation into actionable knowledge so data is AI-ready |
Workday describes why that shared business language has to reach AI through MCP:
“Atlan captures Workday’s shared language to be leveraged by AI via its MCP server. As part of Atlan’s AI labs, we’re co-building the semantic layer that AI needs.”
Joe DosSantos, VP, Enterprise Data and Analytics, Workday
To see where your own metric definitions stand, start with what an enterprise context layer has to hold, then list which metrics exist in more than one place. The metrics with two owners are the ones an agent will get wrong first.
FAQs about MCP and the semantic layer
1. What is the difference between MCP and a semantic layer?
MCP standardizes how an AI agent discovers and calls external tools and data. A semantic layer defines business metrics, dimensions, relationships, and calculation logic. When a semantic layer is exposed through MCP, the protocol carries the request and the semantic layer supplies and applies the metric logic.
2. Can an AI agent query a semantic layer without going through MCP?
Yes. Depending on the product, a semantic layer may expose SQL, REST, GraphQL, or product-specific APIs. MCP gives compatible clients a common way to discover and call capabilities, but it is one access path among several.
3. Does an MCP server on a semantic layer solve the “dashboard says X, AI says Y” problem by itself?
Only partly. It can reduce discrepancies inside that semantic layer, because BI tools and agents calculate from the same model. It does not reconcile cases where another team, tool, or operational system maintains a different definition of the same metric.
4. Which semantic layer vendors document an MCP server?
dbt Labs, Cube, Kyvos, and AtScale document MCP servers for their semantic layers. Looker has a managed MCP server for LookML models, and Snowflake’s managed MCP server exposes Cortex Analyst over semantic views. Each server covers the definitions available inside its own environment.
5. Who decides which AI agents can access which metric definitions through MCP?
MCP does not make that decision. Its optional authorization applies to HTTP-based transports and lets a server verify who is connecting. Each server and the systems behind it then apply their own configured permissions. A context layer can supply broader policy context, while enforcement stays with the systems that serve the context and the data.
6. Is Open Semantic Interchange (now Apache Ossie) related to MCP, or a separate standard?
It is a separate standard. Apache Ossie defines a vendor-neutral format for metrics, dimensions, relationships, and broader business concepts. MCP defines how an agent discovers and requests capabilities from a server. Neither one determines which definition an organization has certified for a given agent and purpose.
7. What happens when an agent calls an MCP server for a metric that is out of date?
The agent receives whatever definition is currently deployed in that model. MCP carries no field for staleness or certification. Version history, certification status, and change information from a governed context layer give the agent something to check before it uses the number.
8. Does a semantic layer’s MCP server expose lineage, or only the metric definition?
It depends on the vendor. dbt’s MCP server documents lineage tools over a dbt project, while the Cube and Kyvos MCP documentation we reviewed on 2026-10-12 lists query and model tools without lineage. End-to-end provenance across warehouses, pipelines, and BI tools needs lineage connected across the wider data estate.
Sources
- Model Context Protocol specification, Server overview, modelcontextprotocol.io, 2025-06-18. https://modelcontextprotocol.io/specification/2025-06-18/server
- Model Context Protocol specification, Authorization, modelcontextprotocol.io, 2025-06-18. https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization
- Introducing the Model Context Protocol, Anthropic, November 25, 2024. https://www.anthropic.com/news/model-context-protocol
- Semantic Layers for Reliable LLM-Powered Data Analytics: A Paired Benchmark of Accuracy and Hallucination Across Three Frontier Models, Rumiantsau and Fokeev, arXiv, 2026. https://arxiv.org/abs/2604.25149
- Gartner Says Lack of Semantics Causes Inaccurate AI Agents and Wasted Spending, Gartner, May 11, 2026. https://www.gartner.com/en/newsroom/press-releases/2026-05-11-gartner-says-lack-of-semantics-causes-inaccurate-artificial-intelligence-agents-and-wasted-spending
- Apache Ossie (Incubating): The New Name for Open Semantic Interchange, Apache Software Foundation, July 10, 2026. https://ossie.apache.org/updates/ossie-enters-apache-incubator/
- dbt MCP server available tools, dbt Labs Documentation, reviewed 2026-10-12. https://docs.getdbt.com/docs/dbt-ai/mcp-available-tools
- Cube MCP server, Cube Documentation, reviewed 2026-10-12. https://docs.cube.dev/docs/integrations/mcp-server
- Kyvos Model Context Protocol (MCP) Server 2.1.0, Kyvos Documentation, reviewed 2026-10-12. https://docs.support.kyvosinsights.com/wiki/spaces/KD20265/pages/1157687095
- MCP server for AI semantic models, AtScale, reviewed 2026-10-12. https://www.atscale.com/resource/mcp-server-ai-semantic-models/
- Looker-managed MCP server, Google Cloud Documentation, reviewed 2026-10-12. https://docs.cloud.google.com/looker/docs/mcp
- Snowflake-managed MCP server, Snowflake Documentation, reviewed 2026-10-12. https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-mcp
- Atlan MCP overview and setup, Atlan Documentation, reviewed 2026-10-12. https://docs.atlan.com/product/capabilities/atlan-ai/how-tos/remote-mcp-overview
