datapact

Lint an ODCS contract and validate a CSV against it, offline.

Offline data-contract runtime: parse and lint an Open Data Contract Standard (ODCS) contract, then validate a CSV dataset against it, in the browser or in CI. No warehouse connection, no upload.

by svx · MIT Licensed

Live site: datapact · Source: github.com/srivtx/datapact

Docs: Rules · Usage · CI · FAQ

Why

A data contract is usually a document: written once, filed, and never enforced. The producer starts emitting nulls, renames a column, or changes a currency code, and nothing fails until a dashboard breaks or a downstream job reads garbage. The contract was never wrong; it was just never checked.

The main open-source engine for this is Python and connection-centric: it wants a warehouse, credentials, and a running server. Most teams do not need that. They need to answer one question in a pull request: does this file match this contract?

datapact answers exactly that, and nothing else. It reads a contract and a data file from disk, checks one against the other, and exits non-zero when they disagree. It runs in the browser, in a pre-commit hook, or in CI, with no connection string and no upload.

Related MCP server: io.github.elwsls/verifiable-claim-seed

What it does

• Parses ODCS — YAML or JSON, reading the schema array and the properties in it. The rest of the document is accepted and ignored.

• Lints the contract — duplicate or missing names, unknown logical types, invalid regular expressions, bad numeric bounds, empty enums, unknown formats, and required/unique on object/array properties (DP-101–DP-109).

• Validates a CSV against it — missing required columns, undeclared columns, type mismatches, empty required values, uniqueness, bounds, patterns, enums, and formats (DP-201–DP-210), with a per-rule suppression cap (DP-250).

• Stays offline — no network code, no warehouse driver, no telemetry. Your data never leaves the machine.

• Speaks to CI and agents — human-readable text, stable JSON, and SARIF 2.1.0, with a documented exit-code scheme.

Install

datapact is not published to npm. Install it from GitHub with the one-line script (requires Bun):

# One-line install (installs the `datapact` binary)
curl -fsSL https://raw.githubusercontent.com/srivtx/datapact/main/install.sh | sh

# Or run once, without installing
bunx github:srivtx/datapact#main --help

# Install globally
bun add -g github:srivtx/datapact
datapact --help

# Add to a project as a dev dependency
bun add -d github:srivtx/datapact

Usage (CLI)

# Lint a contract on its own
datapact lint --contract spec/example-contract.yaml

# Lint the contract and validate a dataset against it
datapact check --contract spec/example-contract.yaml --data spec/example-data.csv

# Machine-readable output for a program
datapact check --contract contract.yaml --data orders.csv --json

# Write a SARIF report and fail on warnings as well as errors
datapact check --contract contract.yaml --data orders.csv --sarif datapact.sarif --fail-on warning

Every option that takes a value accepts both --flag value and --flag=value. The failure threshold is one of error (the default), warning, info, or none; it affects only the exit code, and every finding is still reported.

Library

The same engine is exported for use in your own code. It is browser-safe: no node: imports, so it runs in a page as well as in Bun.

import { parseContract, parseCsv, lintContract, validateDataset } from "datapact";

const contract = parseContract(contractSource);
const dataset = parseCsv(csvSource);

const issues = [...lintContract(contract), ...validateDataset(contract, dataset)];
for (const issue of issues) {
  console.log(issue.code, issue.severity, issue.message);
}

Exit codes

An unreadable or structurally invalid contract is never reported as clean: it produces a DP-PARSE-000 error and exit code 2, so a CI gate cannot pass on input the tool did not understand.

For agents

Every surface is built to be read by a program: stable rule codes, stable JSON, SARIF 2.1.0, and a documented exit-code scheme, so an agent can consume findings without scraping a screen.

• Docs index: the site serves a machine-readable index at srivtx.github.io/datapact/llms.txt.

• Agent guide: AGENTS.md covers build, test, layout, and the hard rules.

• MCP server: expose the runtime to any MCP-capable agent over stdio.

  { "mcpServers": { "datapact": { "command": "bunx", "args": ["github:srivtx/datapact#main", "mcp"] } } }

License

MIT.