Claude file upload limits in 2026: size, count and file types

Anthropic sets different caps depending on where you add the file. These numbers were checked against the Claude Help Center and the Claude API docs in October 2026.

Passing the size check does not mean Claude can read the whole file. The text it extracts still has to fit in the context window.

Where you add a file changes the ceiling from 500 MB down to 30 MB:

• Chat uploads: up to 20 files per chat and 500 MB per file. From documents other than PDFs, Claude reads the text only.
• Project files: 30 MB each. You can add any number, but their total content must fit in the context window, which makes 30 MB a common blocker for big CSV or JSON exports.
• Code execution: 30 MB per file. This is the path when Claude analyzes a dataset with Python, and it is required for XLSX uploads.
• Files API: 500 MB per file and 1 TB per organization, for developers who upload once and reference the file across API calls.

The Claude Help Center lists these file types:

• Documents: PDF, DOCX, TXT, HTML, ODT, RTF, EPUB
• Data: CSV, JSON, XLSX
• Images: JPEG, PNG, GIF, WebP, up to 8000 x 8000 pixels

XLSX uploads only work with code execution and file creation turned on in your settings. Parquet is not on the list.

A 500 MB cap sounds huge, but text files fill the context window long before that. On paid plans, Claude's chat context window ranges from 200K to 1M tokens depending on the model, and a dense CSV or nested JSON file far below 500 MB can use all of it.

With code execution on, Claude can work through a data file with Python instead of reading it all as text, but the file then has to be 30 MB or less. Either way, raw exports from databases, logs or analytics tools usually need cutting down first.

rows.page runs DuckDB-WASM in your browser tab, so your file stays on your machine unless you press Share, and you don't need an account. It opens CSV, TSV, JSON and JSONL up to 1.5 GB per tab, and Parquet without that check.

1. Drop your .csv, .tsv, .json, .jsonl or .parquet file onto rows.page.
2. Filter rows in the search box with country:NL, price:>20 or -name:* (empty values), or click histogram bars and top values in the Columns panel.
3. For bigger cuts, type a DuckDB SQL query that starts with SELECT, WITH or FROM. The table is called data, and results return up to 1,000,000 rows.
4. Export the filtered view or SQL result as CSV, JSON, JSONL, Parquet or Markdown (first 1,000 rows).

Nested JSON fields become dot-path columns, which need double quotes in SQL:

SELECT "user.city", count(*) AS orders, sum(total) AS revenue
FROM data
WHERE status = 'complete'
GROUP BY "user.city"
ORDER BY revenue DESC

Upload the exported result instead of the raw file. rows.page does not open Excel files, so save a large .xlsx as CSV first.

Instead of downloading and re-uploading files, connect Claude to the rows.page MCP server at https://rows.page/mcp, listed in the official MCP registry as page.rows/rows. Claude can then push result tables and hand you a link, or inspect a hosted dataset without pulling the whole file into context.

In Claude Code, add the server from your terminal:

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

In Claude Desktop, run the remote bridge with npx:

npx mcp-remote https://rows.page/mcp

The tools are push (inline content up to 16 MB, or a url), pin_finding, get_dataset (last 20 versions and 20 preview rows), list_datasets (needs an API key), delete_dataset and get_limits. The server allows 60 requests per minute, and anonymous pushes work without setup.

You can also push a file with curl -T data.csv https://rows.page and ask Claude to read it through get_dataset, so only 20 preview rows enter the context. Anonymous links expire after 7 days unless you sign in with Google and keep the dataset on your account.