DB Snooper

Agent skill: src/db_snooper/skills/SKILL.md

Spec: spec/profiler.md

DB Snooper generates compact, LLM-ready database context for SQL generation, query debugging, and schema exploration. Profiling alone drives state-of-the-art text-to-SQL accuracy (Automatic Metadata Extraction for Text-to-SQL). Supports SQLite, PostgreSQL, MySQL, MariaDB, DuckDB, Google BigQuery, and Amazon RDS for PostgreSQL/MySQL/MariaDB. Requires Python ≥ 3.10.

It inspects an existing database and produces a Markdown profile (<database>/<schema>.md): DDL, row counts, sampled rows, and per-column summaries. Use --per-table for one .md per table.

AI agents and text-to-SQL pipelines can read this context instead of guessing table meanings.

Quick Start

Install with pip:

pip install db-snooper

Or run instantly with uvx (no install needed):

uvx db-snooper profile --db-type mysql --user user --password password --database db --schema sch --port 3306

This creates a profile at db/sch.md.

What The Outputs Contain

The profile .md file contains:

• Metadata (YAML frontmatter) with db-snooper version, UTC generation timestamp, SQL dialect, database name, and schema.
• A top-level Relationships section listing every foreign key as - "parent"."col" ← "child"."col" bullets (composite keys as "table".("c1", "c2")), grouped so a parent referenced by many tables appears once as - "parent"."col" ← "child1"."col", "child2"."col". Lines are sorted by parent.
• One compact block per non-empty table, in this fixed layout:
  • A # "<table>" (rows=<N>) header (the count uses the engine's row estimate — ≈N — when available, otherwise an exact COUNT(*)). Table names are delimited identifiers everywhere they appear — headers, Relationships bullets, and the skipped-table summaries — for the same reason as column names.
  • A merged columns: block — one line per column as "name" type[ flags]: profile, so each column's type, flags, and value distribution sit together and the name is printed once. The name is always a delimited identifier ("Enrollment (K-12)" — double quotes on PostgreSQL/Oracle/SQLite and any other dialect, backticks on MySQL/MariaDB/BigQuery, square brackets on SQL Server): the delimiter marks where the name ends even when it contains spaces or parentheses, and shows the exact quoting to use when referencing the column in SQL. Flags (emitted only when they apply): PK, UNIQ (single-column unique), NOTNULL, FK. Example: "id" bigserial PK: unique identifier, 1..12592, "tick" bigint: 4079 distinct, 1..12592, avg=1944.8. Numeric ranges omit the int/float/numeric qualifier — the type token already carries it.
  • SQLite declared types are affinity hints, so the profile checks them against the actual storage (typeof()): untyped columns resolve to their storage class ("x" int, or "x" int|text when mixed), and declared↔stored contradictions render as "qty" numeric→text. String columns also get a content-shape label — digits (compare as strings: leading zeros, lexicographic ordering), iso-date, bool-like, uuid, numeric — so numeric- or date-looking text is never mistaken for real numbers/dates: "County Code" text: digits, 58 distinct.
  • indexes: — parenthesized column lists with delimited names, multi-column indexes keep their order, partial indexes append WHERE <predicate>. The primary-key index is not repeated. none when there are no non-PK indexes. Foreign keys are not repeated per table — the Relationships section is their single home, and the per-column FK flag points there.
  • Profiles carry distinct counts, full histograms for low-cardinality columns, null fractions, numeric ranges, average/median, and top values on the column's single line. High-cardinality free-text/JSON/blob columns are annotated ← dropped from samples.
  • A samples: block — a transposed markdown table (one row per column, columns latest | sample | sample) showing 1 latest row and 2 random rows for the columns whose concrete values add information.
• Tables with fewer than 10 rows emit all rows: (listing every row) in place of samples:, with bare "name" type[ flags] tokens in their columns: block — the dumped rows already expose every value, so profile text would restate them.
• Views and materialized views emit their CREATE VIEW DDL (their SELECT definition) in place of the indexes: line, followed by the merged columns: block.
• Catalog-derived estimates for very large tables (hundreds of millions of rows or more) from each engine's internal statistics — PostgreSQL pg_stats, MySQL COLUMN_STATISTICS histograms, and MariaDB mysql.column_stats — emitted with ≈ markers and a trailing (from db stats) tag so they are distinguishable from exact values.
• Top-level key frequencies for JSON/JSONB columns and min/avg/max element counts for ARRAY columns (when row counts allow), rendered as trailing annotations on the column's columns: line.
• Redacted values for sensitive column names containing password, passwd, pwd, hash, salt, secret, or token — the column's columns: line reads redacted and it is excluded from samples.
• A skipped_technical_tables entry in the frontmatter naming migration/framework tables excluded from the profile.
• Empty tables are skipped by default. A - Skipped N empty table(s): bullet names them; the Python API's ProfileOptions(include_empty_tables=True) includes them (emitting only the bare per-column type tokens in the columns: block, with no profile text or samples:).
• A <schema>.toc.md sidecar written next to the profile in the same run, indexing every top-level section — Relationships, each table block ("Match" (rows=25979): lines 50-289), and the trailing summary bullets — with its exact line range. Its frontmatter pins the profile by profile_sha256 (plus generator, version, generated_at_utc, and profile_lines) so consumers can fail fast on a stale TOC instead of reading shifted line ranges. Emitted by default; --no-toc disables it, and --per-table profiles get none.

Database Examples

SQLite

db-snooper profile --db-type sqlite --database path/to/app.sqlite

PostgreSQL

Profile

db-snooper profile --db-type postgres --database app_db --schema sch --user readonly_user --host localhost --port 5432 --ask-password

MySQL

db-snooper profile --db-type mysql --database app_db --user readonly_user --host localhost --port 3306 --ask-password

MariaDB

db-snooper profile --db-type mariadb --database app_db --user readonly_user --host localhost --port 3306 --ask-password

Amazon RDS

RDS uses the matching PostgreSQL, MySQL, or MariaDB type. Download the Amazon RDS CA bundle, then connect with a database password:

db-snooper profile --db-type postgres --database app_db --user readonly_user \
  --host mydb.123456789012.eu-west-1.rds.amazonaws.com --ssl-ca global-bundle.pem --ask-password

Or use IAM database authentication with a configured AWS CLI:

db-snooper profile --db-type postgres --database app_db --user readonly_user \
  --host mydb.123456789012.eu-west-1.rds.amazonaws.com --ssl-ca global-bundle.pem \
  --rds-iam --aws-region eu-west-1

DuckDB

db-snooper profile --db-type duckdb --database warehouse.duckdb --schema sch

Google BigQuery

Authenticate with Application Default Credentials, then use the Google Cloud project as the database and the dataset as the schema. BigQuery query charges apply to profiling queries.

gcloud auth application-default login
db-snooper profile --db-type bigquery --database project-id --schema dataset

Files (Parquet, JSON, CSV, ...) via DuckDB

DuckDB reads many file formats directly (parquet, JSON/JSONL, CSV/TSV, Arrow, Iceberg, ...), so profile a file by loading it into a DuckDB database as a table, then profiling that database.

duckdb warehouse.duckdb -c "CREATE TABLE sales AS SELECT * FROM read_parquet('sales.parquet')"
db-snooper profile --db-type duckdb --database warehouse.duckdb

Environment Variables

Connection values can come from environment variables instead of flags:

DB_SNOOPER_DB_TYPE=sqlite \
DB_SNOOPER_DATABASE=eval-dataset/student_club/student_club.sqlite \
db-snooper profile

Supported variables:

• DB_SNOOPER_DB_TYPE
• DB_SNOOPER_DATABASE
• DB_SNOOPER_DB_HOST
• DB_SNOOPER_DB_PORT
• DB_SNOOPER_DB_USER
• DB_SNOOPER_DB_PASSWORD
• DB_SNOOPER_SSL_CA
• DB_SNOOPER_SCHEMA

For server databases, --host defaults to localhost, --port defaults to the database default, and DB Snooper securely prompts for a password when DB_SNOOPER_DB_PASSWORD is not set.

Help

db-snooper -h
db-snooper profile -h

Table filters:

db-snooper profile --db-type sqlite --database app.sqlite --include-tables users,orders,line_items

Schema filter:

db-snooper profile --db-type postgres --database app_db --schema reporting --user readonly_user --port 5432 --ask-password
DB_SNOOPER_SCHEMA=reporting db-snooper profile --db-type postgres --database app_db --user readonly_user --port 5432 --ask-password

Profile options:

• --output path: write profiles to a custom directory.
• --metadata-only: emit schema, relationships, row estimates, and available catalog statistics without scanning table rows.
• --per-table: generate one .md profile for each table instead of a single schema profile.
• --no-toc: skip the <schema>.toc.md sidecar (emitted by default).
• --include-tables table_a,table_b: only profile selected tables.
• --exclude-tables table_c: skip selected tables.
• --query-timeout 10: skip individual PostgreSQL/MySQL/MariaDB profiling queries that exceed this many seconds; 0 disables the timeout.
• --max-bytes-billed 1073741824: cumulative BigQuery scan budget. Each profiling query is dry-run first and skipped if it would exceed the remaining budget; 0 disables the cap.

Sampling thresholds and edge-case inclusion behavior are available through ProfileOptions in the Python API.

Python API

Use the simple helpers when you have a SQLAlchemy URL:

from db_snooper import generate_profile

database_url = "sqlite:///eval-dataset/superhero/superhero.sqlite"

profile_md = generate_profile(database_url)

Use the lower-level API when you already have a SQLAlchemy engine or need options:

from sqlalchemy import create_engine
from db_snooper import ProfileOptions, profile_database

engine = create_engine("sqlite:///eval-dataset/superhero/superhero.sqlite")

profile_md = profile_database(
    engine,
    ProfileOptions(
        small_table_threshold=25,
        include_tables=frozenset({"superhero", "publisher"}),
        include_empty_tables=True,
    ),
)

To also get the <profile>.toc.md sidecar content (write it next to the profile file yourself), use the with-TOC variants:

from db_snooper import generate_profile_with_toc

profile_md, toc_md = generate_profile_with_toc(database_url)

toc_md is None when ProfileOptions(emit_toc=False) is set or the profile has no sections to index. profile_database_with_toc offers the same pair for an existing engine.

Agent Skill

DB Snooper bundles db-snooper-profile, an agent skill for generating schema and data context before writing or debugging SQL. It runs db-snooper profile and produces <database>/<schema>.md.

Install it as a Claude Code plugin:

/plugin marketplace add renatyv/db-snooper
/plugin install db-snooper@db-snooper

The skill ships inside the wheel, so you can inspect or install it without cloning this repository:

uvx db-snooper skills list
uvx db-snooper skills install
uvx db-snooper skills install --target all
uvx db-snooper skills install --dir ./.opencode/skills --force

By default, installation targets ~/.config/opencode/skills. Use --target for Claude or agents-compatible directories, or --dir for a custom location.

License

The DB Snooper source code is licensed under the MIT License. See LICENCE.

Third-party Python dependencies remain under their own upstream licenses. See THIRD_PARTY_NOTICES.md for a dependency license summary.

The dataset files included under eval-dataset/ are derived from birdsql by The BIRD Team, and are used and redistributed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

These files are not covered by the MIT source-code license. They retain their original CC BY-SA 4.0 terms. Any derivative works that include these files must also be distributed under CC BY-SA 4.0.