Agent Skills: Postgres query

Postgres query

Purpose

Execute read-only queries against a Postgres database and return structured results. This skill enables other skills to access business data stored in Postgres (e.g. signup metrics, campaign performance, user data).

This skill is read-only by default. It refuses to execute any statement that would modify data.

Instructions

1. Validate credentials

Check that database connection credentials are available from environment variables. Two formats are supported:

• Connection URI: DATABASE_URL (e.g. postgresql://user:pass@host:5432/dbname). If set, this takes precedence.
• Individual variables: PGHOST, PGDATABASE, PGUSER, PGPASSWORD (all required), plus PGPORT (optional, defaults to 5432).

If neither DATABASE_URL nor the required individual variables are set, stop and report:

Postgres credentials not configured. Set the following environment variables: PGHOST, PGDATABASE, PGUSER, PGPASSWORD (or DATABASE_URL as a connection URI).

Never hardcode credentials. Never log or output credential values.

2. Determine the query

If the input is an explicit SQL statement (starts with SELECT, WITH, or EXPLAIN), use it directly. Proceed to step 3.

If the input is a natural-language description, inspect the database schema first:

1. List available tables:

   SELECT table_schema, table_name
   FROM information_schema.tables
   WHERE table_schema NOT IN ('pg_catalog', 'information_schema')
   ORDER BY table_schema, table_name;
   

2. For tables that appear relevant to the request, inspect their columns:

   SELECT column_name, data_type, is_nullable
   FROM information_schema.columns
   WHERE table_name = '{table}'
   ORDER BY ordinal_position;
   

3. Formulate an appropriate SELECT query based on the schema and the natural-language request.

4. Apply a default LIMIT 100 unless the caller specified a different limit or the query already contains a LIMIT clause.

3. Read-only guard

Before executing any SQL, verify the statement is read-only. This check is mandatory and cannot be skipped.

Refuse the query if any of the following are true:

• The statement (ignoring leading whitespace and comments) does not begin with SELECT, WITH, or EXPLAIN
• The statement contains any of these SQL commands: INSERT, UPDATE, DELETE, DROP, ALTER, TRUNCATE, CREATE, GRANT, REVOKE
• The statement contains COPY ... FROM
• The statement contains multiple statements separated by semicolons (excluding a single trailing semicolon)

Important nuance: Check for these as SQL commands, not merely as substrings. A column named update_count or a WHERE clause like WHERE deleted = false should not trigger a rejection. Analyse the statement structure. When in doubt, err on the side of caution and refuse.

If the query fails the read-only check, stop and report:

This skill is read-only. The query contains a write operation ({operation}). Only SELECT queries are permitted.

4. Execute the query

Use one of the following methods, in order of preference:

Option A: Postgres MCP server (preferred if available)

If a Postgres MCP server is configured and accessible, use it to execute the query. This provides structured input/output.

Option B: psql command line (fallback)

Execute the query using psql:

psql -c "{query}" --csv --no-psqlrc -v ON_ERROR_STOP=1

• --csv produces machine-readable CSV output
• --no-psqlrc prevents user configuration from altering output format
• -v ON_ERROR_STOP=1 stops on errors

If DATABASE_URL is set, pass it as the connection parameter. Otherwise, psql reads PGHOST, PGPORT, PGDATABASE, PGUSER, and PGPASSWORD from the environment automatically.

5. Return results

Present the results as structured data:

• rows: The query result rows
• row_count: Number of rows returned
• query_executed: The exact SQL that was run (for transparency and debugging)
• truncated: Whether the result was capped by the row limit

If the query returned no rows, return an empty result set (rows: [], row_count: 0). This is not an error.

Error handling

• Missing credentials: Stop and report which environment variables are needed (see step 1)
• Connection failure: Report the connection error, including host and port. Never include the password in error messages.
• Read-only violation: Refuse the query and explain why (see step 3)
• Query execution error (bad SQL, missing table, permission denied): Report the error from psql or the MCP server, including the SQL that was attempted
• Empty result: Not an error. Return the empty result set normally.

Output format

Return the following to the invoking context:

query_executed: <the SQL that was run>
row_count: <number of rows>
truncated: <true if result was limited>

<results as a formatted table or structured data>

When used as a step in a composite skill, the labelled output contains the full result set for subsequent steps to use.