The 5 MCP tools

For: developers writing prompts that should drive the editor to use QueryDeck.

QueryDeck's MCP server exposes five tools. The names are stable — your prompts and AI agent rules can reference them directly.

execute_query

Run a SQL query against the active QueryDeck connection.

Parameters:

Returns: A formatted text block with the column headers, the row data (table-rendered for small results, JSON-list for larger ones), the row count, and the execution time.

Example prompt:

"Count active users by country, top 10."

The AI generates a SQL SELECT and calls execute_query with it. The result lands back in the conversation.

Blocked by:

• Trial expired without license → error.
• Write keyword in query and allowWrites: false → error.
• Query exceeds queryTimeoutSeconds → cancelled, error returned.

get_schema

Return the schema for a single table or the entire database.

Parameters:

Returns: Structured text with the table's columns (name, type, nullable, default), indexes (name, columns, unique), foreign keys (source column, referenced table.column, rules), and triggers if requested.

Example prompt:

"What does the orders table look like?"

search_schema

Find tables and columns by name or fragment.

Parameters:

Returns: A list of matching objects with their containing table, type, and a short description if available.

Example prompt:

"Where is the email stored?"

The AI calls search_schema with query: "email" and gets back something like:

users.email          VARCHAR(255) — Primary email
contacts.email_addr  VARCHAR(255)
admins.email         VARCHAR(255) — Admin login email
audit.email_changes  VARCHAR(255)

Useful when you don't remember the exact column name.

explain_query

Run EXPLAIN (or the database's equivalent) on a query and return the plan.

Parameters:

Returns: The query plan as text — tree-formatted for PostgreSQL, table-formatted for MySQL.

Example prompt:

"Why is this query slow?" (paired with the SQL the user just ran)

The AI calls explain_query and gets back the plan, then reasons about indexes, joins, and selectivity.

Notes:

• analyze: true requires write-equivalent permissions on some databases (it executes the query). It's allowed in MCP because it's a read-only execution from the database's perspective — no rows are mutated.
• For INSERT/UPDATE/DELETE queries with analyze: true, the write guard still applies.

check_drift

Run the same Drift Mode comparison the QueryDeck app would, and return the result.

Parameters:

Returns: A summary of how many drift items were found, by severity, plus the list of items in human-readable sentences. If verbose: true, each item includes the ORM value and the DB value side by side.

Example prompt:

"Is my Prisma schema in sync with the database?"

The AI calls check_drift and either reports "in sync" or lists the divergences with their severity.

Notes:

• Requires the connection to have been opened via qdeck (so the ORM is known) or for the project to be discoverable from the current working directory.
• Returns "not supported" for ORMs without a Drift parser (Knex, Sequelize, others).

Error handling

All five tools return errors in MCP's standard error envelope:

The AI editor surfaces the error message in the chat, so you'll see what went wrong.

What's next

• Safety — what's blocked and how to allow writes if you really need to
• Install — if you haven't set up the editor yet
• Uninstall — clean removal