Version control for CSV files, without git

Git diffs lines, and a CSV re-exported from a database or a spreadsheet changes the order of rows, the quoting or the float formatting without changing a single fact. The result is a 50,000-line diff nobody reads, a repo that grows by the full file every commit, and git LFS once the file passes 100 MB.

Here a version is a file, not a patch. Pushing the same bytes twice adds nothing and answers unchanged: true. Pushing a changed file adds one version and the comparison is done on rows, not lines.

curl -T data.csv https://rows.page creates version 1 and prints the public link plus an owner link with a token. Next time, curl -T data.csv -H "Authorization: Bearer rpt_..." https://rows.page/<id> makes version 2. The JSON response carries a changes block with the previous version, the row delta, columns added and removed, and a link to the comparison.

Name a key column on the push, ?id_column=sku, and the comparison can match rows across versions. Without a key you still get new and gone rows; with it you also get changed rows.

Open the link and the banner reads something like +120 new, 8 gone, 15 changed vs v1, matched by sku, plus +2 columns or 1 column gone when the schema moved. Rows are compared by hashing the columns both versions share, so a reordered export with the same content shows 0 changes.

If no key was named, the page suggests columns that look unique in both versions (ids, codes, emails) and you pick one with a click. Previous versions up to 64 MB are compared automatically when the page opens; bigger ones compare when you ask. Any older version opens from the version menu.

Without an account a link holds up to 20 versions and is deleted 7 days after creation. A free account keeps 10 datasets for good with the last 3 versions open; older versions stay stored for 90 days but locked. Pro ($12/month) opens 90 days of versions, 100 datasets, files up to 500 MB and private datasets; Max ($39/month) opens 365 days, 1,000 datasets, 2 GB files.

Pushes are rate limited to 10 per minute without an account and 120 per minute with an API key, and anonymous pushes are capped at 50 and 1 GB per day per IP. Nothing is deleted when a limit hits: the push answers 402 and names the plan that fits.

On a plan, an API key (rpk_...) and POST /api/v1/push?key=nightly-report upsert the same dataset every run, so a cron job or a GitHub Actions step always writes the next version of one link instead of minting new ones. Add ?source=, ?run= and ?commit= and every version records where it came from.

Agents do the same through the MCP server at rows.page/mcp: push returns the version and the change summary, get_dataset reads rows back with filters, so a model can see what changed since its last run without diffing anything itself.