Start here. Summarise an SQLite database file: size, page size, text encoding, journal mode (WAL or not), user_version, counts of tables / views / indexes / triggers, and a row count for every table (from ANALYZE statistics when present, otherwise an exact count that stops at 100,000). Virtual tables (FTS5, rtree) are listed with their module and their internal shadow tables are folded away. Opening is read-only and never modifies the file or creates -wal / -shm files.

path: the .db / .sqlite file (relative paths are resolved inside the --dir folder, if one was set).

Describe one table or view in depth: columns with declared type, NOT NULL, default, primary key and generated-column flags; per-column NULL percentage, distinct count, min, max and the storage types actually seen (computed over a sample of up to stats_sample rows); foreign keys in both directions; indexes; CHECK constraints; triggers; the CREATE statement; and sample rows. BLOBs are shown as size plus a hex prefix, long text is truncated with its length.

table: table or view name (case-insensitive; a close-match hint is given if it does not exist). sample_rows: how many example rows to show (0-50). stats_sample: rows to sample for the statistics (100-200000).

Show how tables relate through foreign keys: every child -> parent edge, tables with no relationships, a parents-first load order and circular references. Also reports foreign-key columns that have no index (with a CREATE INDEX suggestion as text) and checks for orphan rows (child rows whose parent does not exist) over a sample, even when PRAGMA foreign_keys is off.

format: "text" (default) or "mermaid" (an erDiagram you can paste into a renderer). orphan_sample: child rows to sample per foreign key (100-100000).

Run one read-only SELECT / WITH / VALUES statement. Anything that would write (INSERT, UPDATE, DELETE, CREATE, DROP, ATTACH, VACUUM, PRAGMA assignments, load_extension) is refused. Results stop at max_rows rows and about 200 KB, and the query is interrupted after timeout_seconds. BLOB cells show size plus a hex prefix, and text longer than 200 characters is truncated with its length. Prefer aggregates (COUNT, GROUP BY) over pulling raw rows.

format: markdown (default), csv or json. max_rows: 1-5000. timeout_seconds: 0.1-60.

Show the execution plan of a SELECT / WITH query without running it (EXPLAIN QUERY PLAN), as an indented tree, followed by plain-language findings: full table scans (with row counts), temporary B-tree sorts for ORDER BY / GROUP BY / DISTINCT, and automatic indexes SQLite has to build on every run. Missing-index suggestions are returned as CREATE INDEX text only; nothing is ever created.

Find where a string or number appears anywhere in the database: a substring search over every non-BLOB column of every ordinary table (virtual and shadow tables are skipped). Returns, per table, the matching rows (rowid or primary key), the column and a snippet around the match. Each table has a time budget and a match cap, so a huge database cannot hang the call; tables that ran out of budget are listed.

case_sensitive: default false (SQLite folds ASCII letters only, so Turkish or other non-ASCII letters match case-sensitively). max_matches_per_table: 1-50. seconds_per_table: 0.1-30.

Look for data problems, using SQL over a sample of each table: columns that are mostly or always NULL (and NULL versus empty-string mixes); columns holding mixed storage types (SQLite does not enforce column types; the typeof() distribution is reported, with examples); key-like columns (id, email, uuid, ...) that contain duplicates, case/space-only duplicates, and rows identical in every non-key column; dangling foreign keys even though SQLite does not enforce them by default; and date-looking text columns whose formats are inconsistent.

table: check just this table (default: all ordinary tables). sample: rows to examine per table (100-200000).

Generate markdown documentation of the whole schema as text: an entity-relationship diagram (mermaid) when foreign keys exist, then every table with its columns, types, defaults, key and reference notes, indexes, CHECK constraints and triggers, followed by views and trigger definitions. Nothing is written to disk; paste the result where you want it.