datadiffer MCP server

<!-- mcp-name: io.github.gauthierpiarrette/datadiffer -->

datadiffer

Diff two tables. See what changed, and which slice of your data it's concentrated in.

98.6% of modified rows have country = 'DE' (8.87x over-represented)

datadiffer compares two tables (dev vs. prod, PR vs. main, source vs. destination) and tells you what changed and where it's concentrated: rows added / removed / modified, per-column change rates, and segment attribution.

Built for the era when agents write your pipelines and someone has to check their work. A maintained, MIT-licensed successor to the archived data-diff.

Try it in 60 seconds

No credentials, no config, no account. It generates its own data:

uvx datadiffer demo

datadiffer-demo/orders_prod.parquet vs datadiffer-demo/orders_dev.parquet
key: order_id (inferred from *_id naming)
rows: +800 added  -250 removed  ~432 modified  49318 unchanged
┏━━━━━━━┳━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃column ┃ changed rows ┃ % of matched┃
┡━━━━━━━╇━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│amount │          432 │        0.87%│
└───────┴──────────────┴─────────────┘
-> 98.6% of modified rows have country = 'DE' (8.87x over-represented)
   93.2% of added rows have plan = NULL (63.19x over-represented)
sample changes:
  [order_id=0]  amount: 5.0 -> 7.1
  [order_id=1]  amount: 6.37 -> 6.38
  [order_id=117]  amount: 32.4 -> 34.5
  [order_id=234]  amount: 59.8 -> 61.9
  [order_id=351]  amount: 87.2 -> 89.3
DIFF FOUND

That last block is the point. A diff tells you 432 rows changed; datadiffer tells you they're almost all German orders, which is the difference between "something moved" and "the VAT change landed."

Install

pip install datadiffer                      # local files: Parquet, CSV, DuckDB
pip install 'datadiffer[snowflake]'         # + Snowflake
pip install 'datadiffer[postgres]'          # + Postgres
pip install 'datadiffer[mcp]'               # + MCP server for AI agents

Use it

# local files
datadiffer diff prod.parquet dev.parquet

# DuckDB
datadiffer diff prod.duckdb:orders dev.duckdb:orders

# Postgres (attached read-only)
datadiffer diff orders orders_v2 --source "postgresql://user@host:5432/db"

# Snowflake, or anything in datadiffer.toml (run `datadiffer init` first)
datadiffer diff wh::analytics.orders wh::analytics_dev.orders

# cross-source: warehouse vs. a local file
datadiffer diff wh::orders orders_snapshot.parquet

# machine-readable, for scripts and CI
datadiffer diff a.parquet b.parquet --format json

Exit codes follow the GNU diff convention: 0 no differences, 1 differences found, 2 operational error.

Useful flags: --key (inferred when omitted), --where, --columns / --exclude-columns, --sample-limit, --no-attribution, -q.

In your pull requests

datadiffer-action posts the report as a sticky PR comment:

### datadiffer: ANALYTICS.ORDERS vs ANALYTICSPR482.ORDERS: differences found ❌

+800 added · −250 removed · ~432 modified · 49,318 unchanged (2.96% of base rows affected)

Policy: max-changed-rows-pct exceeded: 2.96% > 0.5%

| Column | Changed rows | % of matched |

|---|---:|---:|

| amount | 432 | 0.87% |

> Where it's concentrated: 98.6% of modified rows have country = 'DE' (8.87× over-represented)

> 93.2% of added rows have plan = NULL (63.19× over-represented)

Shortened. The full README is on GitHub.