Give Claude Code access to your local data files
Claude Code can answer questions about your data — but only if it can actually see it. This guide shows you how to connect Pipetable so Claude queries your real local files instead of guessing.
The problem
When you paste CSV data into Claude, it reads it and forgets it. When you ask "what's the total revenue by region?", Claude generates a plausible-sounding answer — not a real one. Pipetable fixes this by giving Claude a live SQL connection to your files via the Model Context Protocol (MCP).
Install Pipetable
curl -fsSL https://pipetable.com/install | sh
Or with Rust:
cargo install pipetable
Connect to Claude Code
One command registers Pipetable as an MCP server:
claude mcp add pipetable pipetable mcp
That's it. Restart Claude Code and it will have access to Pipetable's four tools.
Query your data
In Claude Code, ask it to scan your data folder:
"scan my ~/data folder and show me what's there"
Claude will call scan_folder, then list_datasets, and you'll see your files. Then ask anything:
"what's our total revenue by region this quarter?" "which customers churned last month?" "show me the top 10 products by units sold"
How it works
Pipetable exposes four MCP tools to Claude:
- scan_folder — registers all CSV, Excel, Parquet, JSON files in a folder
- list_datasets — shows schemas and column types
- get_schema — full schema + sample rows for one table
- execute_sql — runs real DuckDB SQL, returns ground-truth results
Claude writes the SQL. DuckDB runs it. Results are real data, not generated.
Cursor, Copilot, and RooCode
The same Pipetable binary works as an MCP server for any MCP-compatible AI tool:
// ~/.cursor/mcp.json or .roo/mcp.json
{
"mcpServers": {
"pipetable": {
"command": "pipetable",
"args": ["mcp"]
}
}
}
// VS Code (Copilot)
{
"servers": {
"pipetable": {
"type": "stdio",
"command": "pipetable",
"args": ["mcp"]
}
}
}
MIT licensed. 5 MB binary. Files never leave your machine.
Download Free →