Parquet Viewer — DBConvert

Open a .parquet file in the editor: schema, row groups, codec, and SQL across several files at once. Nothing leaves your machine.

What the file tells you before a row is read

Parquet keeps its schema and its statistics in the footer, so the shape of the file is known without scanning it. All of this appears the moment you open it:

• Row groups, and the codec that compressed them
• Size on disk against uncompressed size
• Every column with its type
• Per column: minimum, maximum, how many nulls, bytes in the file
• The writer that produced the file

Several files at once

The editor opens one file per tab. Add more .parquet files to that tab with Choose files and each becomes a named view, so the SQL reads like SQL:

SELECT c.country, count(*) AS orders, round(sum(o.amount), 2) AS total
FROM orders o JOIN customers c USING (customer_id)
GROUP BY 1 ORDER BY 3 DESC

Same format only. Joining a Parquet file to a JSONL one is a different job, and that one is DBConvert Streams.

Every table pages to its last row. Export a result as CSV, JSON or Parquet: the first 10,000 rows free, the whole result with PRO.

What it does not do

It is read-only — SELECT and WITH … SELECT, nothing that writes — and it needs a file that already exists.

Where the file goes

Nowhere. The editor reads it from disk and hands it to the panel; the panel parses it with WebAssembly in the editor itself. There is no account and no remote server. The one port the extension opens is the local one AI agents use (below), bound to 127.0.0.1 and closed by a token.

Ask AI about the open file

The file tab has an Ask AI about this file button. It opens the editor's own agent chat with the question already typed — "What is in orders.parquet? Use the DBConvert viewer tools to look at its tables and data." — and the agent answers from the file, not from a guess.

The agent gets three read-only tools: list the tables of the files open in the viewer, describe a table, and run one SELECT. Each query it runs shows in the viewer, and a result comes back as at most 100 rows with the full row count.

Where each agent gets the tools from:

• VS Code with Copilot Chat — nothing to set up. The tools are part of the chat.
• Cursor — nothing to set up. The extension adds its server to Cursor's MCP list. Cursor asks before each tool call unless you allow it.
• Claude Code and Codex — run Connect agent to open files from the Command Palette once. It adds the server to the agent's config and prints the first question to ask. Quit Codex before you connect it: a running Codex rewrites its config and drops the entry.
• Windsurf, Cline, Roo, Kilo, Gemini CLI, Continue and others — the same command shows the config to paste: a JSON block, or the address and header.

In an editor with no agent chat to open, the button copies the question and offers Connect.

One server serves every open editor window and every installed DBConvert viewer, so an agent sees all the files you have open. Reset MCP token issues a new token; connected agents need Connect again after it.

The viewer never talks to a model itself. What the agent reads — table names, columns, the rows its queries return — goes to whichever model that agent uses.

The other viewers

Same viewer, one per format — install only the one you need, so a SQLite user never downloads the Parquet engine.

• SQLite Viewer — VS Code · Open VSX · Chrome
• JSONL Viewer — VS Code · Open VSX
• MDB and ACCDB Viewer — VS Code · Open VSX
• PostgreSQL Dump Viewer — VS Code · Open VSX

Built by DBConvert. The same viewer runs in a browser tab at streams.dbconvert.com/parquet-viewer if you would rather not install anything.