All Posts

9 Best Semantic Layer Tools For AI Agents In 2026

by Rick Radewagen12 min read

Our pick for AI agents is Malloy: an agent asks for a measure by name, aggregates stay correct across joins, and the compiler rejects a bad query before the warehouse runs it, all in one open-source (MIT) project. The exceptions are layers you already run and trust: the dbt Semantic Layer (MetricFlow) if your metrics live in dbt, LookML if you report out of Looker, Power BI semantic models in a Microsoft shop, and Cube or AtScale when many BI tools must read the same metrics live. As of October 2026 every tool on this list has an MCP server or an AI API, so "works with agents" no longer separates them. What does: how metrics are defined, which query interfaces exist, whether a user's permissions reach the agent, and who owns the definitions.

A note on bias: GetDot is an AI data analyst. We don't sell a semantic layer and none of the tools below is ours, but we recommend Malloy and GetDot reads it natively (as it does six others), so weigh that accordingly. Product facts and prices are as of October 2026, link to their source, and are list prices.

Semantic layer tools compared

Tool

Best for

Query interfaces

MCP server

Pricing (October 2026)

Malloy

Our pick: one open-source language for the model and the queries

Compiles to warehouse SQL; Publisher REST API

Yes, built into Publisher (no auth by default)

Free (MIT)

dbt Semantic Layer (MetricFlow)

Teams whose models already live in dbt

JDBC, ADBC, GraphQL, Python SDK

Yes, official (local or remote)

MetricFlow free (Apache 2.0); hosted APIs on paid dbt plans, Starter at $100 per developer seat a month

Cube

One layer for many BI tools, apps and agents

REST, GraphQL, SQL; MDX and DAX in Cube Cloud

Yes, in Cube Cloud (hosted, OAuth)

Cube Core free (Apache 2.0); Cube Cloud seats plus metered compute

Looker (LookML)

Google Cloud teams already on Looker

Looker API, Open SQL Interface (JDBC), Conversational Analytics API

Yes, Looker-managed (preview) or MCP Toolbox

Platform edition plus user licenses, annual, through sales

AtScale

Enterprises with many Excel and Power BI users

XMLA (DAX, MDX), SQL over JDBC/ODBC, Python, REST

Yes, incl. Databricks MCP Marketplace

Consumption-based quote; launched at $2,500 a month and up in 2024

Power BI semantic models

Microsoft shops with DAX measures

DAX over XMLA and the REST API

Yes: Fabric IQ MCP (GA), remote Power BI MCP (preview), open-source Authoring MCP

Pro $14, PPU $24 per user a month, billed annually; XMLA needs PPU, Premium or Fabric capacity

Snowflake semantic views

Snowflake-only stacks using Cortex

SQL in Snowflake, Cortex Analyst, Cortex Agents

Through the Snowflake-managed MCP server

No separate price found; warehouse compute plus Cortex AI credits

Databricks metric views

Databricks-only stacks using Genie

SQL with MEASURE(), dashboards, Genie

Through Databricks managed MCP servers

No separate price found; Databricks compute

Steep

Metrics, catalog and exploration UI in one product

Steep API

Yes, hosted (OAuth)

Not published in a form we could verify

What a semantic layer does for AI agents

A semantic layer stores business definitions once and compiles every request into SQL that follows them, turning "write correct SQL against a large schema" into "pick a metric and a few dimensions". It doesn't decide which metric the person meant (the semantic layer solves the wrong half); our semantic layer guide covers the evidence on both sides.

How I compared them

On definitions, query interfaces, MCP support, whether row-level security reaches the agent, pricing and portability, from vendor docs and pricing pages read in October 2026 plus building Dot's connectors for six of them. I ran no head-to-head accuracy benchmark, so Malloy's place at the top is a design recommendation, not an accuracy ranking.

1. Malloy (our pick)

Best for: any team that can choose its semantic layer: one open-source, code-first language for the model and the queries, with no BI platform to buy.

Malloy is an open-source modeling and query language (MIT, on GitHub) co-created by Looker founder Lloyd Tabb (Santa Cruz Works). The startup Credible builds a commercial product on it (VentureBeat).

  • Why it's our pick: measures are names, not formulas; aggregates stay correct across joins; the compiler catches mistakes before the warehouse does; and a restricted mode keeps LLM-written queries inside the model. Our Malloy guide makes the full case.
  • Interfaces: Malloy compiles to SQL for your warehouse. Its README lists nine database engines, including BigQuery, Snowflake, Postgres, DuckDB/MotherDuck and Trino (as of October 2026). Publisher, Malloy's open-source model server, serves models over REST and MCP (malloydata/publisher).
  • MCP and AI: Publisher has a built-in MCP server at localhost:4040/mcp with tools such as get_context, compile_model and execute_query (README). The endpoint has no authentication: the docs recommend binding it to 127.0.0.1 and putting an authenticating gateway in front of anything wider (Malloy docs).
  • Pricing: free and open source. You pay for the warehouse, and for hosting Publisher if you run it.
  • Limitations: a young project with a small ecosystem, so business users need a front end you build or buy, and Publisher and its MCP server are younger than the long-running commercial layers (as Cube's own comparison notes).

2. dbt Semantic Layer (MetricFlow)

Best for: teams whose transformations already live in dbt and who want metrics reviewed and versioned next to the models.

MetricFlow, the engine that turns a metric request into SQL, has been open source under Apache 2.0 since October 2025 (dbt Labs); the dbt Semantic Layer is dbt Labs' hosted service around it. Metrics are YAML in your dbt project (latest spec).

  • Interfaces: JDBC, ADBC and GraphQL APIs, plus a Python SDK (Semantic Layer APIs).
  • MCP and AI: dbt Labs maintains an open-source MCP server (dbt-labs/dbt-mcp) that runs locally or remotely, with Semantic Layer tools such as list_metrics, query_metrics and get_metrics_compiled_sql. The remote server is on every dbt platform plan, but which tools you get depends on the plan (dbt MCP docs).
  • Pricing: MetricFlow is free. The hosted Semantic Layer comes with the Starter, Enterprise and Enterprise+ plans (dbt docs). Starter lists at $100 per developer seat a month and includes 5,000 queried metrics a month; Enterprise is custom-priced and includes 20,000 (dbt pricing, Paradime's 2026 summary).
  • Limitations: an agent can only answer what is modeled as a metric. Query caching and multiple credentials per project are Enterprise features (Semantic Layer FAQ). Setup and the MCP route: our dbt Semantic Layer guide.

3. Cube

Best for: one metric layer that serves BI tools, spreadsheets, embedded analytics and AI agents at once.

Cube Core is the open-source, headless semantic layer (Apache 2.0); Cube Cloud is the managed platform built on it (GitHub, Cube docs). Metrics live in YAML files that declare cubes and views.

  • Interfaces: REST, GraphQL and SQL APIs (Cube docs). An MDX API for Excel on Windows (Enterprise plan, preview; MDX API) and a DAX API for Power BI (public preview; Cube blog) connect over XMLA.
  • MCP and AI: Cube Cloud hosts an MCP server with OAuth on its Premium and Enterprise plans. It exposes 30 tools, nine of them destructive (Cube MCP docs); self-hosted Core relies on community MCP servers.
  • Pricing: Cube Core is free. Cube Cloud has a free tier, and paid plans combine per-user seats with metered compute. Third-party listings put Starter at about $40 and Premium at about $80 per user a month but disagree on the compute rates, so get a quote; Enterprise is custom (TrustRadius, Gartner Peer Insights).
  • Limitations: two meters make the bill harder to predict, and self-hosted Core gets REST, GraphQL and SQL only, run and secured by you.

4. Looker (LookML)

Best for: Google Cloud and BigQuery teams that already run Looker and want agents to reuse its explores.

LookML views define dimensions and measures, explores define how views join, and Looker compiles every query into SQL.

  • Interfaces: the Looker API and its SDKs; the Open SQL Interface, which exposes LookML models to any tool that supports JDBC (Google docs); and the Conversational Analytics API, generally available for BigQuery and Looker since June 23, 2026 (release notes).
  • MCP and AI: the Looker-managed MCP server is in preview and built into Looker-hosted instances, at your instance URL plus /mcp. Every tool is off until an admin enables it, and the client inherits the user's Looker roles (Looker MCP docs). Customer-hosted instances use Google's open-source MCP Toolbox for Databases (Google docs).
  • Pricing: platform editions plus role-based user licenses on an annual commitment, through sales (Looker pricing; estimates in our Looker pricing breakdown).
  • Limitations: the model is only queryable through Looker, so its value depends on keeping a Looker license. The managed MCP server is a preview without OAuth scopes or dynamic client registration, and an agent that ignores an explore's default filters can override them by accident.

5. AtScale

Best for: large enterprises where Excel and Power BI users need live, governed metrics on a cloud warehouse, without extracts.

AtScale is a commercial universal semantic layer. In 2024 it open-sourced its YAML-based modeling language, SML, under the Apache license (BusinessWire), and it joined the Open Semantic Interchange in January 2026 (BusinessWire).

  • Interfaces: Power BI and Excel connect live over XMLA with DAX and MDX, Tableau and Looker over JDBC and ODBC, Python notebooks through AI-Link (AtScale).
  • MCP and AI: AtScale's MCP server has been in the Databricks MCP Marketplace since December 1, 2025 (AtScale). AtScale says row-level security and column masking apply to agents as to BI users (AtScale blog).
  • Pricing: consumption-based, priced on deployed semantic objects. At its 2024 launch AtScale advertised plans "as low as $2,500 per month"; real quotes are custom (BusinessWire).
  • Limitations: built for large BI estates rather than a five-person data team, and the accuracy figures in AtScale's material come from its own tests.

6. Power BI semantic models

Best for: Microsoft shops where DAX measures already define the business and people read them in Power BI, Excel and Teams.

Tables, relationships and DAX measures that reports read live; our Power BI semantic model guide walks through an example.

  • Interfaces: DAX queries over the XMLA endpoint, which only works on workspaces on Fabric capacity, Premium capacity or Premium Per User, and is read-only by default (Microsoft Learn); and DAX through the Power BI REST API.
  • MCP and AI: Microsoft's remote Power BI MCP server (preview) reads a model's schema and generates and runs DAX. The user needs Build permission, and row-level security applies when users sign in but not with a service principal (Microsoft Learn). Microsoft now points question answering to its generally available, read-only Fabric IQ MCP server (our Power BI MCP guide), and the open-source Authoring MCP server (formerly the Modeling MCP server) lets an agent change models and run DAX, capped at 100,000 rows (GitHub).
  • Pricing: Power BI Pro is $14 and Premium Per User $24 per user a month, billed annually (Microsoft). Outside tools that read models over XMLA, Dot's connector included, need PPU, Premium or Fabric capacity. Our Power BI pricing guide has the capacity math.
  • Limitations: outside Microsoft's own tools, only clients that speak XMLA and DAX can use the model, and the capacity requirement raises the floor for small teams.

7. Snowflake semantic views

Best for: teams that keep everything in Snowflake and use Cortex Analyst or Cortex Agents.

A semantic view is a schema-level Snowflake object with tables, relationships, facts, dimensions and metrics, created in SQL or from a YAML spec (CREATE SEMANTIC VIEW, querying).

  • Status: querying semantic views became generally available in August 2025 (release 9.25), and naming a semantic view in a plain FROM clause went GA on March 2, 2026 (release note).
  • MCP and AI: the Snowflake-managed MCP server can expose Cortex Analyst as a tool, and only over semantic views (Snowflake docs); it returns generated SQL, not results (Snowflake guide).
  • Pricing: we found no separate price for the views themselves. Cortex Analyst bills AI credits when called through Cortex Agents or per 1,000 messages through its standalone API, and running the SQL uses warehouse compute (Cortex Analyst docs).
  • Limitations: definitions live inside Snowflake, so outside tools need a bridge such as AtScale.

8. Databricks metric views

Best for: Databricks lakehouse teams that use AI/BI dashboards and Genie.

Metric views are defined in YAML in Unity Catalog. Measures must use aggregate functions and are queried with MEASURE(), and SELECT * isn't supported (Databricks docs).

  • Status: generally available as part of Unity Catalog Business Semantics, which Databricks is also open-sourcing; materialization is still in preview (Databricks blog). External BI tools can query them too (Databricks docs).
  • MCP and AI: Databricks' managed MCP servers include Genie One, generally available since September 2026, a Genie Agent server scoped to one agent, and Databricks SQL (Databricks docs, Databricks blog).
  • Pricing: we found no separate price; you pay for the Databricks compute that runs the queries.
  • Limitations: Databricks only, and every column has to be listed explicitly.

9. Steep

Best for: teams that want metrics, a metric catalog and an exploration UI in one product.

A metrics platform with its own semantic layer that can also sit on top of the dbt Cloud Semantic Layer (Steep docs) or Cube (Steep).

  • Interfaces: the Steep API, launched on January 20, 2026, exposes the metrics and semantic model you define in Steep (Steep).
  • MCP and AI: a hosted MCP server at https://mcp.steep.app/mcp since May 28, 2026, with OAuth; an agent only sees data you can see in Steep (Steep changelog).
  • Pricing: third-party listings disagree, so check Steep's pricing page.
  • Limitations: it is metrics-first, so a question that no defined metric covers is out of scope.

Others worth knowing

  • Honeydew: a semantic layer that runs as a Snowflake Native App (Honeydew). Its text-to-data API is also available through a Honeydew MCP server (Honeydew docs), and it publishes per-user pricing (pricing).
  • GoodData: an analytics-as-code BI platform with its own semantic layer. Its MCP server, announced in January 2026, lets AI create, change and validate semantic models, metrics and dashboards (GoodData). Pricing is through sales.
  • Open Semantic Interchange (OSI): not a tool but a vendor-neutral format for moving definitions between these layers, now Apache Ossie; see our OSI guide for which tools support it.

How to choose a semantic layer for your stack

If you can choose: Malloy

If you can choose, choose Malloy. It costs nothing to license, and GetDot reads it natively (Malloy integration). The cases below are the exceptions, where switching is unrealistic and the layer you run should stay.

Looker shops: keep LookML

If Looker is your BI tool, keep the definitions in LookML so every Looker surface and outside agent computes revenue the same way. Malloy, co-created by Looker's founder, keeps LookML's core idea, so a large LookML estate can move to it gradually, starting with new models. See the Looker integration page and our Gemini in Looker comparison.

Microsoft: Power BI semantic models

If your measures already exist in DAX, don't rebuild them in a second layer; fix descriptions and relationships first, because every AI client reads the same model. Copilot or an outside analyst? See our Power BI Copilot comparison.

dbt shops: the dbt Semantic Layer

If you are committed to dbt, MetricFlow keeps metric definitions in the same pull requests as the models, and the dbt MCP server gives agents query_metrics without extra glue. The catch is that the hosted APIs need a paid dbt plan.

Many BI tools or embedded analytics: Cube or AtScale

When the same metric has to reach Tableau, Power BI, Excel, your own product and an agent, a universal layer earns its cost: Cube for engineering-led teams (open core, APIs, pre-aggregations), AtScale for enterprises whose heaviest users live in Excel and Power BI.

All in on Snowflake or Databricks

Semantic views and metric views are the cheapest place to start if every consumer sits inside that platform. Once a second BI tool or an outside agent needs the definitions, you're back to a bridge or a second layer.

No semantic layer yet

Don't start with a tool. Write down what your top 20 metrics mean and which table is the source of truth, then make the ones people argue about deterministic, in Malloy unless your stack already settles it. Our semantic layer guide covers when you need one, and AI made that second step cheap.

How AI analysts use these layers

  1. The layer's own MCP server. Quick and governed for modeled metrics, but it only sees its own layer, so a question that crosses systems or needs an unmodeled table stalls or gets guessed.
  2. The platform's own assistant (Gemini in Looker, Copilot in Power BI, Cortex Agents, Genie): strong inside one platform, blind outside it.
  3. A governed AI analyst that reads the layer. It queries the layer for defined metrics and the warehouse for everything else, keeps the context the layer can't hold (which metric people mean, known data issues), and answers where people work.

Which layers GetDot reads natively

GetDot is the third kind. As of October 2026 it reads seven semantic layers natively and lets each one compute the number, plus plain dbt Core projects as context:

Layer

What runs when you ask a question

Malloy

Each source is a table. Dot compiles your model with the Malloy engine inside Dot, so there is no Malloy server to host, and runs each query in Malloy's restricted mode, through your sources, joins and measures. Row-level security applies per group, and warehouse secrets stay in Dot, never in your repo.

Looker (LookML)

Each explore is a table to Dot. Dot writes an explore query (fields, filters, sorts), and Looker compiles it from your LookML and runs it. Read-only until an admin allows edits; access filters tied to Looker user attributes apply, and since August 2026 Dot respects each explore's default filters instead of overriding them.

dbt Semantic Layer

Each metric is a table, and its dimensions and entities are columns. Dot sends a semantic_layer.query(metrics=[...], group_by=[...]) request, so MetricFlow writes the SQL.

dbt Core project

Dot parses the project and uses model docs, SQL and lineage as context when it queries the warehouse. Dot can export its dashboards as dbt exposures.

Cube

Each cube or view is a table. Every answer is a Cube query through Cube's REST API, so your measures, joins, pre-aggregations and access rules apply.

Power BI semantic models

Each semantic model is a table, like a Looker explore. Dot runs DAX through the Power BI REST API and reads tables, measures and relationships over XMLA.

Azure Analysis Services

DAX against your tabular models over XMLA, reusing their measures.

Steep

Each metric is a table. Dot asks Steep's API for every number, broken down by your dimensions.

The Cube connector shipped in October 2025, the metric-first dbt integration later that month, Power BI semantic models in December 2025 and Malloy in July 2026. Dot does not read Snowflake semantic views, Databricks metric views, AtScale, Honeydew or GoodData natively today; on Snowflake and Databricks it queries tables directly.

Dot is also an MCP server (Dot for MCP): one connection from Claude, ChatGPT, Cursor or Copilot covers every layer and warehouse Dot reads, with the asker's own Dot permissions.

FAQ

Software that stores business definitions once and compiles every request into SQL that follows them, either standalone (Malloy, dbt, Cube, AtScale), inside a BI tool (LookML, Power BI) or inside the warehouse (Snowflake, Databricks).

Our pick is Malloy: measures are names, aggregates stay correct across joins, and a restricted mode keeps LLM-written queries inside the model. Keep the layer you have where switching is unrealistic, such as a Power BI model or a large LookML estate.

A semantic layer outside any single BI tool, also called headless BI, that serves many consumers through standard interfaces such as SQL, REST, XMLA with DAX or MDX, and MCP, as Cube and AtScale do.

As of October 2026, all nine tools in the table above plus Honeydew and GoodData; Snowflake semantic views and Databricks metric views are reached through each platform's managed MCP servers.

dbt if your models are already in dbt and you want metrics reviewed in the same pull requests; Cube for pre-aggregations, embedded analytics or live Excel and Power BI connections. They aren't mutually exclusive, and if you have neither and need no Excel or Power BI connection, we'd start in Malloy instead; our head-to-head comparison goes further.

If you already run one of these layers, connect it to Dot for free and see how your definitions hold up when people ask real questions in Slack or Teams. The Free plan includes 300 one-time credits with full Pro features, and paid plans are billed on credits, not seats.

Rick Radewagen

Rick is a co-founder of Dot, on a mission to make data accessible to everyone. When he's not building AI-powered analytics, you'll find him obsessing over well-arranged pixels and surprising himself by learning new languages.

Read more