What are the best text-to-SQL tools for complex enterprise databases?

As of November 2, 2026, pick by where the data lives and give the agent a governed tool rather than raw SQL: Snowflake's managed MCP server exposing a Cortex Agent or Cortex Analyst over semantic views; the Databricks Genie Agent MCP server (up to 50 curated Unity Catalog tables, views or metric views, read-only); the dbt MCP server's query_metrics tool when your metrics already live in the dbt Semantic Layer; Microsoft Fabric data agents on F2 or higher capacity; WrenAI ($179 a month billed annually, with an MCP server) across several databases; and Google's open-source MCP Toolbox for Databases when you want to define your own SQL tools.

How accurate is text-to-SQL on enterprise schemas in 2026?

Agent systems now score far higher than early models on the Spider 2.0 benchmark of real enterprise workflows: the top entry on the Spider 2.0 leaderboard page in November 2026 is Genloop's Sentinel Agent v2 Pro at 96.70, with Snowflake AI Research's Arctic-FLEX at 75.14. Those are tuned research systems; production accuracy depends on your semantic model, so test with 20 to 50 verified questions from your own business before rollout.

Do these tools respect database permissions?

The managed ones do. Databricks' Genie Agent MCP server enforces Unity Catalog permissions on every request, Snowflake's managed MCP server uses role-based access control on the server and each tool with Snowflake OAuth by default, and Fabric data agents use the asking user's read permission on the data source. With self-built tools such as MCP Toolbox, you are responsible for scoping credentials and queries.

Quick Answer

Best Text-to-SQL Tools for AI Agents and Complex Databases

Q: How should an AI agent query a database: natural language to SQL or predefined tools?

For governed business questions, through a semantic layer: the agent asks a question, and a tool grounded in defined metrics, joins and verified queries writes the SQL. Snowflake explicitly recommends exposing a Cortex Agent as the only client-facing tool for governed questions and putting raw SQL execution on a separate MCP server with a least-privileged role. Free-form SQL generation against hundreds of raw tables is where complex joins and aggregations go wrong.

Published: November 2, 2026

The short answer

Do not hand an agent a database connection and a schema dump. For complex enterprise databases — hundreds of tables, multi-step joins, business-specific calculations — the tools that work put a semantic layer between the agent and the SQL. As of November 2, 2026:

Data in Snowflake: Snowflake-managed MCP server exposing a Cortex Agent (or Cortex Analyst) grounded in semantic views.
Data in Databricks: the Genie Agent MCP server, one curated Genie Agent per domain.
Metrics already defined in dbt: the dbt MCP server and its query_metrics tool.
Microsoft shop: Fabric data agents (GA, F2 capacity or higher).
Several databases, or embedding in your product: WrenAI with its MCP server.
You want to write the tools yourself: Google’s open-source MCP Toolbox for Databases.

This page is the agent-facing spoke of our best text-to-SQL tools guide, which covers the human analyst side.

The comparison

Tool	How the agent connects	Grounding for complex joins	Permissions	Cost model (Nov 2, 2026)
Snowflake managed MCP server	MCP; tools: Cortex Agents, Cortex Analyst, Cortex Search, `SYSTEM_EXECUTE_SQL`, custom	Semantic views + verified queries	RBAC per server and tool; Snowflake OAuth or External OAuth	AI Credits: $2.00 global / $2.20 regional routing; standalone Analyst API per 1,000 messages; plus warehouse compute
Databricks Genie Agent MCP	MCP, `/api/2.0/mcp/genie/{space_id}`	Genie Agent with up to 50 tables, views or metric views	Unity Catalog on every request	Databricks SQL pricing
dbt MCP server	MCP; `query_metrics`, `list_metrics`, `text_to_sql`, `execute_sql`	dbt Semantic Layer metrics and dimensions	Your dbt Platform and warehouse roles	dbt plan + warehouse
Fabric data agent	Fabric, Microsoft 365 Copilot, multi-agent apps	Instructions, examples, Power BI semantic models	Asking user’s read permission	Fabric capacity, F2 or higher
WrenAI	MCP server, embedded API	Semantic modelling, dbt integration	RBAC; row/column-level on Enterprise	Essential $179/mo, Enterprise $559/mo (annual)
MCP Toolbox for Databases	MCP; prebuilt `list_tables`, `execute_sql` + custom tools	Whatever you encode in custom tools	Your credentials and query design	Open source, self-hosted

1. Snowflake managed MCP server — best for Snowflake

The Snowflake-managed MCP server lets agents retrieve data without you deploying separate infrastructure. It can serve Cortex Analyst, Cortex Search and Cortex Agents as tools, plus custom tools and SQL execution, up to 50 tools per server. The important design guidance is in Snowflake’s own docs: expose a Cortex Agent as the only client-facing tool for governed business questions, because exposing SYSTEM_EXECUTE_SQL on the same server lets the client bypass semantic views and verified queries. If you need raw SQL, put it on a separate server with a least-privileged role.

Cost: Cortex Agents bill AI Credits per million tokens ($2.00 per AI Credit with global routing, $2.20 with regional routing for data residency); Cortex Analyst invoked through an agent uses the same model, while the standalone Analyst API bills per 1,000 messages. Every generated query also runs on your warehouse.

2. Databricks Genie Agent MCP server — best for Databricks

Each Genie Agent MCP endpoint is scoped to one Genie Agent: a curated set of up to 50 Unity Catalog tables, views or metric views. The agent sends a plain-language question; Genie writes and runs the SQL and returns results, with Unity Catalog permissions enforced on every request. Two limits matter: it is read-only, and it passes no conversation history (use Databricks’ AI Bridge from Python for follow-ups). Calls bill at Databricks SQL rates, and a workspace admin must enable Managed MCP Servers on the Previews page.

The 50-object cap is a feature for complex schemas: build one Genie Agent per domain (finance, supply chain) with metric views for the hard calculations, rather than one space over everything.

3. dbt MCP server — best when metrics live in dbt

If revenue, churn and margin are already defined in the dbt Semantic Layer, let the agent call list_metrics, get_dimensions and query_metrics instead of writing SQL. The metric definitions carry the joins and aggregations, so “net revenue by region last quarter” returns the number finance agrees with. text_to_sql and execute_sql exist for the long tail. Works across warehouses that dbt supports.

4. Microsoft Fabric data agents — best for Microsoft shops

Fabric data agents are generally available and run on any paid F2 or higher capacity (or Power BI Premium P1+). The agent routes each question to the right source and uses NL2SQL for lakehouses and warehouses, NL2DAX for Power BI semantic models and NL2KQL for KQL databases. For Power BI models it uses the asking user’s Read permission. One agent can combine up to five data sources, it only generates read queries, and you can add organisation-specific instructions and example question/query pairs (not yet supported for Power BI semantic model sources).

5. WrenAI — best across databases

WrenAI’s MCP server and embedded API give agents the same semantic model human users query. Pricing is per credit with unlimited members: Essential Cloud $179/month (13,200 annual credits) and Enterprise Cloud $559/month (24,000 annual credits, row- and column-level control, audit logs), billed annually; agents and humans draw from the same pool. Self-hosted Enterprise Plus is quoted by concurrent sessions.

6. MCP Toolbox for Databases — best for custom tools

Google’s open-source MCP Toolbox (renamed from genai-toolbox) has two modes: prebuilt generic tools such as list_tables and execute_sql for developer exploration, and a custom tools framework where you define parameterised queries the production agent may call. For complex enterprise logic, the second mode is the safe one: you encode the join once, the agent fills in parameters.

How to handle complex joins and calculations

Encode the hard parts once: metric views (Databricks), semantic views (Snowflake), dbt metrics, or parameterised tools. The model should choose a metric, not invent a five-way join.
Curate scope: fewer, documented tables per agent beat the whole warehouse.
Keep a verified-question set of 20–50 real business questions with known answers and re-run it on every model or schema change.
Separate read paths from SQL execution and give the raw-SQL path the narrowest role.

Last verified: November 2, 2026. Features and prices from vendor documentation and pricing pages read that day.