MCP Natural Language Database AI Agent Plugin for Revive Adserver

The MCP Natural Language Database Agent turns your Revive Adserver database into something anyone on your team can query directly — no SQL, no database client, no waiting on a developer to pull a report. Built on the Model Context Protocol (MCP) and powered by a Gemini- or Groq-hosted AI model, it accepts ordinary business questions, works out whether database access is needed, discovers your actual schema, generates a safe read-only query, executes it, and hands back a clear, human-readable answer.

The Problem It Solves

Business users routinely need answers like:

• How many advertisers were added this month?
• What are the top 5 zones by impressions?
• Show orders by month for this quarter.
• Which publishers had the highest revenue last week?
• What was the average order value last quarter?

Traditionally, getting there means understanding the schema, writing SQL, running it in a database client, and interpreting the result — usually a developer or analyst's job, not a sales or ops team's. This plugin automates the entire chain.

Benefits

• Self-service reporting — anyone on your team gets answers in seconds, without pulling in a developer or waiting on a report request
• No SQL knowledge required — questions are asked the way people actually think about their data, not the way a database expects them
• Faster decisions — campaign and revenue questions get answered in real time instead of after a manual export-and-analyze cycle
• Always grounded in your real setup — schema is discovered automatically, so answers are based on your actual tables and columns, not assumptions
• Safe by design — read-only access, a dedicated low-privilege database account, and no credentials ever reaching the AI model, so this can be rolled out without exposing sensitive data
• Vendor flexibility — works with either Gemini or Groq, so you're not locked into a single AI provider and can choose based on cost or latency needs
• Frees up developer time — routine “can you pull me a number” requests stop landing on your engineering team's plate

How It Works, End to End

1. You ask a question in plain English — e.g. “Show the top 5 customers by total sales.”
2. The AI model receives the question along with the tools available to it.
3. If database access is required, it calls a schema-discovery tool to learn your real tables and columns rather than guessing.
4. It generates a SELECT statement based on that actual schema.
5. It calls the query tool, which validates and executes the SQL against your database.
6. Rows/aggregates are returned to the model.
7. The model interprets the result and replies with a concise, final natural-language answer.

The two tools that make this work:

• get_schema — takes no input, returns a JSON map of your database's tables and columns, so query generation is always grounded in what actually exists (not assumed table/column names).
• query_database — accepts a single read-only SQL string (SELECT … FROM … WHERE … LIMIT …) and returns the resulting rows as JSON/text.

Architecturally, your MCP client (the orchestration layer) handles all AI-model interaction and tool calling, while a separate MCP server owns the actual database connection and enforces what's allowed to run against it — the AI model never touches the database directly.

Security and Safety Model

Database access is treated as the most sensitive part of this system, and the AI model is never trusted as a database administrator. The MCP server enforces security independently of anything the model requests:

• Only read-only SELECT statements are permitted.
• INSERT, UPDATE, DELETE, DROP, ALTER, TRUNCATE, and any other write/DDL operation is rejected outright.
• Runs on a dedicated, SELECT-only database account rather than an admin/root credential.
• A maximum result-row limit is enforced on every query.
• Query execution timeouts and resource limits are supported.
• Database credentials are never exposed to the AI model.
• Sensitive columns are not exposed unless explicitly permitted by your configuration.
• Every tool call and database query is logged for auditing.
• SQL is validated with a proper parser rather than simple keyword filtering, which is not sufficient protection on its own.

Setup

Configuration is handled through environment variables — your Gemini or Groq API key/model choice, and standard database connection settings (host, port, name, and a read-only user/password), kept out of source code entirely. The underlying AI provider is swappable without any change to how the plugin functions, so you're not locked into one model vendor.