An MCP server for data: let AI agents publish tables you can open

Model Context Protocol (MCP) is an open standard that lets AI assistants call external tools. The rows.page server runs at https://rows.page/mcp over Streamable HTTP (JSON-RPC over POST) and is listed in the official MCP registry as page.rows/rows.

When your agent calls push with CSV, TSV, JSON or JSONL text, or with a public file URL, rows.page returns a https://rows.page/<id> link. The link opens the data in a browser table where you can filter, sort, run DuckDB SQL against the data table and check each column's histogram and distinct count. With pin_finding, the agent saves the view it found (a filter, sort, column list or SQL query) on the dataset, so you open straight to it.

The server has six tools. Most work without an account, while the ones that change a dataset need its owner token or the account that owns it.

The reply to a new push includes an rpt_ owner token. Your agent needs it later to pin findings, push new versions or delete the dataset.

For Claude Code, run:

claude mcp add --transport http rows https://rows.page/mcp

Claude Desktop connects through the mcp-remote bridge. Add it to claude_desktop_config.json and restart the app:

{"mcpServers": {"rows": {"command": "npx", "args": ["mcp-remote", "https://rows.page/mcp"]}}}

For Cursor, add the server URL to mcp.json:

{"mcpServers": {"rows": {"url": "https://rows.page/mcp"}}}

This works without an account, and pushed links then expire after 7 days. To push into your account instead, sign in with Google on rows.page (the $0 Free plan works) and create an rpk_ API key, up to 20 per account. Send it as Authorization: Bearer rpk_... or add ?apiKey=rpk_... to the server URL.

1. The agent runs a query or a script, then calls push with the result as CSV or JSONL, or with a public url for files over 16 MB.
2. It gets back the link, an owner link and the rpt_ token, then calls pin_finding with that token and a query such as SELECT * FROM data WHERE price > 20 ORDER BY __rid. It can also pass up to 20 findings with the push itself.
3. You open the link on the pinned view and can export it to CSV, JSON, JSONL, Parquet or Markdown (Markdown keeps the first 1,000 rows).
4. On the next run, the agent calls push with dataset and token, or with the same key name and an API key, to add v2 to the same link.
5. If the data has not changed, the reply says unchanged: true and no new version is added.

When the previous version is 64 MB or less, the dataset page shows a banner like "+N new, N gone, N changed vs vN". Rows are matched on the id_column the agent passes with push, or on a column you pick in the banner.

• 60 requests per minute per IP address. Above that, the server answers HTTP 429.
• Pushes have their own limit of 10 per minute without an API key.
• Inline data: the whole request is capped at 16 MB, and a bigger one gets HTTP 413.
• url pushes: up to 100 MB, fetched by rows.page, and the remote server must respond within 30 seconds.
• Without an account: 50 pushes and 1 GB per IP address per UTC day, up to 20 versions per link, and links expire after 7 days.
• Free ($0): 10 hosted datasets that don't expire, last 3 versions visible.
• Pro ($12 per month): 100 datasets, 90 days of versions, private datasets.
• Max ($39 per month): 1,000 datasets, 1 year of versions, private datasets.

MCP pushes stay at 100 MB on every plan. Pro and Max upload bigger files, up to 500 MB and 2 GB, through multipart uploads at POST /api/v1/uploads.

No. The server runs on rows.page, so it only sees what your agent sends. An agent on your machine, like Claude Code or Cursor, reads the file with its own file tools and sends up to 16 MB inline. For bigger files up to 100 MB, run curl -T data.csv https://rows.page in a terminal.

If you only want an agent to query local files without publishing a link, a local filesystem or SQLite MCP server fits better, though you lose the browser table and the link to send a teammate.

Convert other formats before pushing. rows.page reads CSV, TSV, JSON, JSONL and Parquet, and does not read Excel, XML, YAML, PDF, SQLite, Avro, ORC, fixed-width text or zip files.