dbprint

Write down once what your database means, next to your code, and every agent session reads it.

Your coding agent reads your migrations and your models, so it gets the shape of your database right. What it gets wrong is everything the shape does not say: that revenue_cents is gross and every report uses the net column, that status is spelled two ways, that orders joins shipments one-to-many more often than anyone expects, that legacy_uid has been dead since the migration. A colleague would tell it in thirty seconds. Without a database connection the agent cannot find out. With one, it rediscovers the same facts live, every session — queries that cost your warehouse time and budget and cost the agent tokens — and can still read them wrong: a count does not say that a code is dead or that a column is gross.

dbprint turns that knowledge into plain text committed in your repository:

• It measures the database once. Each table's DDL, its joins — declared, and inferred where the schema declares none — and per-column statistics: null shares, complete value lists with every spelling, ranges, distributions, value shapes and more. On any of eight engines.
• An agent writes down what the numbers mean. dbprint annotate hands an agent one table's measurements and the questions they leave open — what a code means, what a sentinel date stands for, which rows a report leaves out — and the agent answers them from your code and read-only queries, into files dbprint never overwrites. dbprint annotate --check flags a note the data contradicts.
• Every later session reads it offline. Through an MCP server, a packaged skill or dbprint context, with no database connection and no credentials at the agent. Because it is text in git, a schema or data change shows up in review, and dbprint check gates it in CI.

With dbprint, agents get more answers right, for less. In our tests across multiple codebases, coding agents reading an annotated print got 80% fewer answers wrong than without one, at 17% lower cost per correct answer — and 90% fewer on questions whose catch only the print stated.

What your agent reads

An excerpt of dbprint context --purpose query on a table from the reference print in this repo — the fragment an agent gets before it writes SQL:

# Table: arboretum.seedbank.germination_trial  (900 rows, 8 columns)

## Joins
- accession_id -> arboretum.seedbank.accession.accession_id (declared)
- collector_id -> arboretum.seedbank.collector.collector_id (inferred)

## Data dictionary

One row per germination attempt run against a batch drawn from an accession.

## Column values

| Column | Values (count) | Coverage | Nulls |
|---|---|---|---|
| medium | 'moist filter paper' (176), 'sterile grit' (176), 'vermiculite mix' (176), 'agar medium' (175), 'sand tray' (176) {'sand tray' (151), 'Sand Tray' (25)}, 'control' (21) = 'No growth medium: a no-treatment control group, its germinated_count compared against ambient conditions rather than a physical medium.' | 100% - the list is the whole domain |  |

Three things here no migration states: medium holds exactly these values and no others, sand tray is stored in two spellings, so a filter on one misses 25 rows, and control is not a medium at all. The first two dbprint measured; the third a human or an agent wrote down once, and every session after it reads.

Run it yourself, with no database and nothing to configure — only uv:

uvx dbprint context arboretum.seedbank.germination_trial --purpose query \
  --project https://github.com/jakubro/dbprint/blob/main/docs/format/v1/examples/production/.dbprint.yaml

Quickstart

Requires Python 3.13+. Install with the driver extra for your database:

uv tool install 'dbprint[postgres]'     # or [mysql], [snowflake], [duckdb], [clickhouse], [redshift], [databricks], [bigquery]

Add the [mcp], [docs] or [dbt] extras in the same install — uv tool install 'dbprint[postgres,mcp]' — since a later install replaces the earlier one rather than adding to it.

The Postgres adapter shells out to pg_dump for DDL, so the PostgreSQL client binaries must be on your PATH.

Scaffold a project in your repo root:

dbprint init

That writes .dbprint.yaml (commit it), creates prints/, and writes a credentials template to ~/.dbprint/connections.yaml — outside the repo, so secrets never sit next to the config.

Put your connection details in the credentials file:

# ~/.dbprint/connections.yaml
primary:
  host: localhost
  port: 5432
  database: my_db
  user: dbprint_ro
  password: change_me

and tell .dbprint.yaml which adapter to use and which tables to read:

# .dbprint.yaml
connections:
  primary:
    adapter: postgres
    auto: true
    tables:
      include:
        - "*.public.*"

Credentials resolve from DBPRINT_<CONNECTION>_<KEY> environment variables first, then ~/.dbprint/connections.yaml, then a project-root .env. Every key of both files is in docs/CONFIG.md.

Profile it:

dbprint generate

Commit the result alongside your code. Then inspect it with no database in reach:

dbprint list

Then have an agent annotate a table — the brief carries the annotation skill's steps, so any agent CLI can follow it:

dbprint annotate my_db.public.orders | claude -p

and connect your agent to the result.

What it does to your database

• It reads. A role that can SELECT the tables you profile is enough; each engine's adapter page, linked from your first print, lists the exact grants. The one write is a temporary copy of a sampled table's drawn rows, gone when the session ends (on BigQuery, when its expiration passes), and only when you sample — read.tables.materialize turns it off.
• It publishes real values. Value lists, ranges and percentiles are cell values, and they land in your repository. Nothing is withheld by default: redaction masks, drops or hashes them by column name, by detected sensitivity (personal_name, national_id, ...) or by shape, keeping every count.
• It costs one profiling run, then what you choose to re-run. A table is re-read when its print is older than its freshness threshold or its settings changed; sampling, row filters and catalog-only reads bound the cost on a warehouse — see choosing what to profile.

What a print looks like

Under the context an agent reads are plain YAML files. An excerpt of one table's statistics.yaml, some fields elided, from the same reference print:

row_count: 2500
columns:
  catalogue_url:
    sql_type: character varying(200)
    nullable: false
    null_rate: 0.0
    classification: text
    cardinality: 2500
    cardinality_ratio: 1.0
    inferred:
      looks_like: url
      sampled: 1000
      matched: 1000
      candidate_key: true
    values:
    - value: https://specimens.example.org/accession/1
      count: 1
    - value: https://specimens.example.org/accession/10
      count: 1
    # ... 28 more
    values_coverage: 0.012      # top 30 of 2500 - a value list, not a claim of completeness
    distribution: long_tail

  provenance_country:
    sql_type: character(2)
    classification: categorical
    cardinality: 10
    inferred:
      looks_like: country_code
      sampled: 10
      matched: 10
    values:
    - value: AU
      count: 250
    - value: CA
      count: 250
    - value: DE
      count: 250
    - value: FR
      count: 250
    - value: GB
      count: 250
    # ... 5 more
    values_coverage: 1.0        # the list is the whole domain, not a sample
    distribution: uniform

  received_at:
    sql_type: timestamp(0) with time zone
    classification: temporal
    range:
      min: '2018-01-08T00:00:00Z'
      max: '2024-11-28T00:00:00Z'
      span_days: 2516
    percentiles:
      p01: '2018-02-09T00:00:00Z'
      p25: '2019-10-02T00:00:00Z'
      p50: '2021-06-19T00:00:00Z'
      p75: '2023-03-05T00:00:00Z'
      p99: '2024-10-25T00:00:00Z'

values_coverage: 1.0 is the load-bearing field: those ten values are the entire domain, not the top ten. An eleventh value can be ruled out without asking the database. (When a table is profiled from a sample rather than in full, coverage is measured over the rows that were actually read — see Scoped profiling below.)

inferred is what dbprint concluded rather than read: looks_like: country_code is a verdict, and sampled/matched are the evidence for it, so a shape confirmed by ten values never looks like one confirmed by ten thousand.

Alongside it, relationships.yaml carries the join graph:

refers_to:
- column:
  - collector_id
  target_table: arboretum.seedbank.collector
  target_column:
  - collector_id
  on_delete: RESTRICT
  detection: declared      # or `inferred`, when dbprint guessed the edge from naming
  observed:
    fanout_avg: 6.25
    target_coverage: 1.0

That half is measured. The other half you write, in the same directory, in the same commit — statistics.annotations.yaml for a column:

columns:
  collector_id:
    note: >-
      FK to collector.collector_id. Field collectors are affiliated with a
      partner institution; collector.institution names it.

and description.md for table-level narrative that no statistic can express.

dbprint never writes to either file, and both are read alongside the statistics — but on anything the statistics can answer, the statistics win; the narrative is authoritative only on what a number can't.

What lands in your repo

prints/<connection>/
  manifest.yaml                    # inventory, freshness, per-table thresholds
  reading.md                       # generated guide - how to read the rest of the print
  diff.yaml                        # machine-readable drift from the last run
  manifest.annotations.yaml        # optional, connection-level notes you write
  <schema>/<table>/
    ddl.sql                        # native DDL, normalized for stable diffs
    statistics.yaml                # column distributions
    relationships.yaml             # declared + inferred foreign keys
    description.md                 # optional, you write it
    statistics.annotations.yaml    # optional, per-column notes you write
    relationships.annotations.yaml # optional, per-edge notes you write

reading.md is written fresh on every run: it teaches a consumer which fields to trust for what, so a print explains itself to an agent that has never seen the format.

The four .md / .annotations.yaml files are yours — dbprint never overwrites them, and all four are folded into what agents read.

Commands

Command	What it does
dbprint init	Scaffold .dbprint.yaml + prints/ + a credentials template
dbprint config explain	Print the profile one table resolves to, each value commented with the default, connection or rule that set it
dbprint generate	Profile the database; write the print and diff.yaml
dbprint list	Offline summary of the committed print
dbprint diff	Compare the live database against the committed print
dbprint check	CI gate: structure, freshness and statistic assertions; --online adds drift and SQL assertions
dbprint context	Emit a prompt-ready Markdown view of one or more tables
dbprint annotate	Print an authoring brief for one table's annotations, or --check them
dbprint docs	Browse the print as an HTML site — serve it live or build it static (the [docs] extra)
dbprint serve	Read-only MCP server over the committed print (the [mcp] extra)

Full reference: docs/CLI.md.

Connecting an agent

An MCP client launches the server itself, so it — not you — picks the working directory. Name the project explicitly:

{
  "mcpServers": {
    "dbprint": {
      "command": "uvx",
      "args": ["--from", "dbprint[mcp]", "dbprint", "serve", "--project", "/srv/analytics"]
    }
  }
}

The server reads the committed print and never connects to a database, so the client needs no credentials.

Reading a remote project

--project also accepts a git address, so a print committed to someone else's repository reads without cloning it by hand first:

dbprint list --project https://github.com/jakubro/dbprint/blob/main/docs/format/v1/examples/production/.dbprint.yaml
dbprint context arboretum.seedbank.accession --project https://github.com/jakubro/dbprint#main:docs/format/v1/examples/production

dbprint clones it for you and asks for no credentials of its own — your existing SSH agent or credential helper covers a private repository. It's read-only: commands that write or connect live to a database (generate, diff, check --online) refuse a remote locator; every other command works against it exactly as it does locally.

Beyond the basics

• Semantic inference. Every column gets one of 8 classifications. Text, categorical and foreign-key columns get one of 38 shape patterns (uuid, email, phone, ip, country_code, ...), published with the sample size and match count behind the verdict. Columns that hold data that must not leave the database get a sensitivity axis (personal_name, postal_address, national_id, credential, ...). Columns effectively unique (cardinality ratio >= 0.9999) are marked candidate_key, with a candidate_key_exception alongside whenever that threshold admits measured or estimated duplicates rather than true uniqueness. Foreign keys your schema never declared are inferred from naming and stamped detection: inferred.
• Redaction. Withhold cell values with mask, drop, or hash (salted), targeting columns by glob or by inferred sensitivity or by shape — so a personal_name column added next quarter is covered by a rule written today. Counts, cardinality and distributions survive intact; only literals are withheld. The salt lives with your credentials, never in the committed config.
• Assertions. Write data-quality checks in YAML against any statistic, or as raw SQL. Statistic assertions are evaluated against the committed print by plain dbprint check, with no database in reach; SQL assertions need dbprint check --online. See docs/ASSERTIONS.md.
• CI exit codes. 0 ok, 1 malformed print or unusable config, 2 stale, 3 drift, 4 connection failure, 5 partial run, 6 assertion failure, 7 total failure. diff, check and context all speak --format json|yaml.
• Scoped profiling. Per-table rules can sample (seeded from the table name, where the engine supports seeding), apply a row filter, or change what's measured — gated on table size, so one rule covers every fact table you haven't named yet. A table read this way carries a scope block and a per-column rows_scanned, and every count in the file is denominated in the rows actually read, never silently in the whole table.
• MCP server. dbprint serve exposes 7 tools — get_table_context, list_tables, search_columns, resolve_value, get_manifest, get_diff, get_reference — plus every artifact as a resource. It reads the committed print only and never opens a database connection. See docs/MCP.md.
• Markdown skills. For clients that take markdown rules or custom instructions, a drop-in skill file tells an agent what a committed print answers, which tool answers it, and what each number covers — no process to run. A second skill has an agent write the annotation files from the code that writes each table and read-only queries, checked with dbprint annotate --check. Both sit at docs/examples/skill/ so you can read them before installing anything.
• dbt lineage. Point a connection's dbt_manifest at a compiled dbt manifest.json and generate records each model column's source columns as derived_from, so a note written once on a source column reaches every model built from it (the [dbt] extra). See docs/CONFIG.md.
• Token-budgeted context. dbprint context arboretum.seedbank.accession --budget 4000 fits the Markdown view into a token budget, dropping the lowest-priority sections first.

Closest neighbours

Project	What it does	How dbprint differs
dryrun	Offline Postgres schema snapshot + MCP server + linting	Nearest thing to this. Its snapshot is a compressed binary bundle you mark binary in .gitattributes; dbprint's artifact is line-diffable YAML with a published spec, across eight databases
dbt-profiler	Column profiles committed as dbt YAML / Markdown	dbt projects only. No DDL, no relationship graph, no semantic inference or redaction
Live-connection MCP servers (dbhub, postgres-mcp, Supabase, Neon)	Query and introspect a database at agent runtime	They need credentials at the agent, every session. dbprint profiles once, commits the result, and serves it with no database in reach
tbls, SchemaSpy	Schema documentation and ER diagrams for humans	Structure only, no data distributions. dbprint docs serve renders the same browsable pages and FK diagrams over a print's measured columns

What dbprint is not

• Not a migration tool. It doesn't author, apply or version DDL changes. Use Alembic, Flyway, Liquibase, sqldef, Atlas.
• Not a schema-management or ORM layer. A print is generated from your database; it is never the source of truth you build it from.
• Not a live query interface. Nothing in a print executes SQL for your agent.
• Not a data-quality platform. The assertion DSL is a CI gate over a print, not monitoring, alerting or lineage. Use Soda, Great Expectations, Elementary.
• Not a data catalog. No hosted service, no shared search index, no business glossary. dbprint docs serves a local read-only view of a committed print, bound to loopback, with nothing to administer.

The format

A print is a portable, git-committed snapshot of a database's structure and column-level data distributions, consumable offline by humans, AI coding agents, and CI. The on-disk format — not the CLI — is the durable artifact here: specified, versioned independently of the tool, and designed for other producers to emit.

The format is specified in docs/format/v1/SPEC.md.

Seven JSON Schemas ship inside the installed package at src/dbprint/spec/v1/ — four for the generated artifacts, three for the annotation files you write — and the conformance validator is a public API with no database, config or CLI required:

from dbprint.conformance import validate_print

issues = validate_print("prints/production")  # conforms iff no error-severity issues

The reference print at docs/format/v1/examples/ is real dbprint output, not a hand-written illustration — what you read there is what you get.

Configuration

.dbprint.yaml holds project config (committed); ~/.dbprint/connections.yaml holds credentials (never committed). Every key of both, with types, defaults and resolution order, is in docs/CONFIG.md. Internals are documented in docs/ARCHITECTURE.md.

License

Copyright 2026 Jakub Roman. Distributed under the Apache License 2.0.