Run one read-only SQL statement (SELECT, WITH, VALUES, TABLE, SHOW) against PostgreSQL. Runs in a read-only transaction that is always rolled back, with server-side timeouts. Results are bounded: has_more means more rows existed (truncated_by says whether the row cap or byte budget stopped retrieval), and long values are cut with a …[+size] marker. Columns with sensitive-looking names (password, token, api_key, ...) come back as [redacted], counted in columns_redacted and values_redacted.

List tables outside system schemas: schema-qualified name, estimated_rows (planner estimate, null if never analyzed; no scans), and column count. Optional filter: case-insensitive substring matched against table and column names.

Describe several tables in one call: columns (name, type, nullable), primary key, outgoing and incoming foreign keys (constraint, from_table/from_columns, to_table/to_columns, paired by position), and indexes. Each column carries planner statistics (no scans): stats_available, and when true null_frac, n_distinct (negative = minus the distinct fraction of rows, -1 = unique) and, for enum-like non-sensitive columns, common_values with frequencies. Columns matching a redaction pattern get stats_available: false, redacted: true. Unknown tables get a per-table error; the rest still succeed.

Fetch one row by its full primary key plus its depth-1 declared foreign-key neighborhood in one call. Each relation is one FK constraint: direction outgoing (rows the root references) or incoming (rows referencing the root), the related table, constraint, source_columns (referencing) paired by position with target_columns (referenced), status (ok, timeout, error), and compact rows (columns once, rows as arrays). Each relation returns up to N rows (default 5) ordered by the related table's primary key (order_by; ctid without one), never by recency; has_more means more rows existed, with no count. Relations not attempted due to the relation cap or total deadline are listed in relations_not_attempted with reason (max_relations, deadline). Values are truncated and redacted as in query, totalled in values_truncated and values_redacted. Tables without a declared primary key need query instead.