MCP server

IntoSQL has two MCP servers. Both let an AI assistant convert spreadsheets, CSV and JSON into SQL for PostgreSQL, MySQL, MariaDB, SQLite, SQL Server or Oracle, or into clean CSV or JSON.

Use the hosted server when the assistant runs in the browser and the file is already online. Use the local server when the file is on your disk, or when it must not leave your computer.

Hosted server#

Pro: The hosted server runs conversions on IntoSQL's servers, which is part of IntoSQL Pro. Without Pro, convert_file answers with "Remote conversions are part of IntoSQL Pro."

The server speaks MCP over streamable HTTP at:

https://intosql.com/mcp

Connect Claude#

1. In Claude, open Settings, then Connectors.
2. Choose Add custom connector.
3. Paste https://intosql.com/mcp and add it.
4. Claude opens an IntoSQL page. Sign in and approve the access.

Connect ChatGPT#

1. In ChatGPT, open Settings, then Apps and connectors.
2. Choose Create.
3. Paste https://intosql.com/mcp and choose OAuth as the authentication.
4. Sign in to IntoSQL when asked and approve the access.

Connect Cursor#

Add the server to mcp.json:

{
    "mcpServers": {
        "intosql": {
            "url": "https://intosql.com/mcp"
        }
    }
}

Cursor asks you to sign in to IntoSQL the first time it uses the server.

Sign-in#

The server accepts two kinds of credentials:

• OAuth. Claude, ChatGPT and Cursor use the authorization code flow. You sign in to IntoSQL in the browser and approve the client. The access token lasts one day and the client renews it with a refresh token, which lasts 90 days.
• An API key. A key from Settings, API keys works as a bearer token, for clients that send headers instead of signing in. See API for how to create one.

{
    "mcpServers": {
        "intosql": {
            "url": "https://intosql.com/mcp",
            "headers": {
                "Authorization": "Bearer YOUR_API_KEY"
            }
        }
    }
}

Either way, the assistant sees the conversions of the account it signed in with, and nothing else. A suspended account is refused.

How files reach the server#

convert_file takes a public HTTPS address, such as a shared download link. IntoSQL's worker downloads the file from there. The hosted server doesn't accept uploads through MCP, and it can't read files on your computer. For those, use the local server.

The address must be HTTPS and reachable from the internet. Addresses on private networks or localhost are refused with "The file address must be HTTPS and reachable from the internet."

When the address doesn't end in a file name with an extension, pass name (for example ventas.xlsx) so the engine knows what it is reading.

Tools#

convert_file

Converts the file at a public HTTPS address. It waits up to 20 seconds for the result. If the conversion finishes in that time, the answer includes the download address. Otherwise it answers with the conversion's id and status, and the assistant calls get_conversion later.

The conversion uses the defaults: the first visible sheet, the header row and delimiter detected, CREATE TABLE included, snake_case column names, INSERT statements of 500 rows, dates read in the order the file proves, and leading zeros kept.

get_conversion

Reads one conversion.

A failed conversion carries errorCode, the same stable code as error.code in the API (Errors), and error, the explanation.

list_conversions

Lists the account's latest 20 conversions, newest first, under conversions. It takes no inputs.

cancel_conversion

Stops a conversion that hasn't ended and deletes its files. It stays listed as canceled and doesn't count toward the month. A conversion that already ended can't be canceled.

delete_conversion

Deletes a conversion's input and output files now, instead of after 24 hours. It answers "Deleted."

The record stays, so the month's usage still counts it.

get_usage

Answers how many bytes of conversions the plan includes this month, how many are used and how many are left:

{
    "limit": 5368709120,
    "used": 18874368,
    "left": 5349834752
}

What a conversion looks like#

convert_file, get_conversion and each item of list_conversions answer with the same fields:

{
    "id": "01jf3k8v2m9q4r7t1x5y6z0abc",
    "name": "ventas.xlsx",
    "status": "done",
    "dialect": "postgresql",
    "bytes": 1843200,
    "rows": 12480,
    "outputBytes": 2310144,
    "error": null,
    "createdAt": "2026-10-16T14:02:11+00:00",
    "finishedAt": "2026-10-16T14:02:19+00:00",
    "expiresAt": "2026-10-17T14:02:11+00:00",
    "download_url": "https://..."
}

Limits#

When a limit is reached, the tool answers with an error the assistant can read, for example "5 conversions are already queued or running on this account. Start this one when one of them ends."

Server card#

The server describes itself at a well-known address, for clients and registries that discover MCP servers:

https://intosql.com/.well-known/mcp/server-card.json

{
    "$schema": "https://static.modelcontextprotocol.io/schemas/v1/server-card.schema.json",
    "name": "com.intosql/mcp",
    "title": "IntoSQL",
    "description": "Converts spreadsheets, CSV and JSON into SQL a database accepts on the first try.",
    "version": "1.0.0",
    "websiteUrl": "https://intosql.com/developers",
    "icons": [
        {
            "src": "https://intosql.com/icon-512.png",
            "mimeType": "image/png",
            "sizes": ["512x512"]
        }
    ],
    "remotes": [{ "type": "streamable-http", "url": "https://intosql.com/mcp" }]
}

The same card is also served at https://intosql.com/mcp/server-card.

Local server#

intosql mcp runs the IntoSQL engine as an MCP server over stdio. The assistant starts it, and it reads and writes files on your computer. Files are never uploaded: conversions are written to your disk, and loads go only to the database you name.

It comes with the command line. Install it first and check that it runs:

intosql version

You don't start the server yourself. The MCP client runs intosql mcp and talks to it. Run in a terminal on its own, it waits for a client on standard input.

Connect Claude Desktop#

Add the server to claude_desktop_config.json. On macOS the file is in ~/Library/Application Support/Claude/; on Windows, in %APPDATA%\Claude\.

{
    "mcpServers": {
        "intosql": {
            "command": "intosql",
            "args": ["mcp"]
        }
    }
}

Restart Claude Desktop. If it can't find intosql, use the full path that which intosql prints, such as /opt/homebrew/bin/intosql.

Connect Cursor#

Add the server to mcp.json, either ~/.cursor/mcp.json for every project or .cursor/mcp.json in one project:

{
    "mcpServers": {
        "intosql": {
            "command": "intosql",
            "args": ["mcp"]
        }
    }
}

Local tools#

Paths are absolute, or relative to the folder the server runs in. Every tool also takes the reading options in Reading options.

inspect_file

Reads a file without converting it. Answers with its format and encoding, where the table is, each column's name and SQL type for the database, every correction the conversion would make, and a preview.

convert_file

Converts a file and writes the output to disk. Answers with output_path, output_bytes and report: rows converted and left out, column types, and every correction.

load_file

Loads a file straight into a database: PostgreSQL with parallel COPY, MySQL, MariaDB or SQLite. It creates the table with the inferred types unless told the table exists, loads every row and answers with report. A failed load drops the table it created.

Pro: Loading into SQLite is free. Loading into PostgreSQL, MySQL or MariaDB needs IntoSQL Pro. Without a license, the first load starts a 14-day trial on this computer, and the answer carries a notice with the days left. Activate a license with intosql license activate <license>; see Command line.

The connection URL, password included, is passed to the tool by the assistant. Keep that in mind before you paste a production password into a chat.

Reading options#

All three local tools take these optional inputs:

For example, to rename the third column and keep the first as text:

{
    "path": "/Users/ana/Downloads/clientes_2024.xlsx",
    "dialect": "mysql",
    "columns": [
        { "index": 2, "name": "signup_date" },
        { "index": 0, "as_text": true }
    ]
}

Example prompts#

Hosted server

• "Convert https://example.com/exports/ventas.csv to PostgreSQL." The assistant calls convert_file with the address and dialect: postgresql, then gives you the download address, or calls get_conversion until the conversion is done.
• "What did I convert this week, and how much of my allowance is left?" The assistant calls list_conversions and get_usage.
• "Stop the ventas conversion, I sent the wrong file." The assistant calls cancel_conversion.
• "Delete the ventas conversion now." The assistant finds its id with list_conversions and calls delete_conversion.

Local server

• "Look at ~/Downloads/clientes_2024.xlsx and tell me what IntoSQL would change." The assistant calls inspect_file and summarizes the column types and corrections.
• "Convert pedidos.json to a SQL Server script with the table name pedidos." The assistant calls convert_file with dialect: sql_server and table: pedidos, and tells you where pedidos.sql_server.sql was written and how many rows went in.
• "Load ventas.xlsx into postgres://ana@localhost:5432/tienda as ventas_2024." The assistant calls load_file and reports the rows loaded.

Troubleshooting#

"Remote conversions are part of IntoSQL Pro." The hosted server needs an active Pro plan on the account the assistant signed in with. See Pricing.

"The file address must be HTTPS and reachable from the internet." The address is http://, points to a private network, or can't be resolved. Share the file with a public HTTPS link, or use the local server.

"The file couldn't be downloaded: ..." IntoSQL's worker couldn't fetch the file. Some sharing links open a web page instead of the file; use the direct download link.

"This month's remote conversions are used up." The month's allowance is spent. get_usage shows the numbers. The local server has no allowance.

The download address stopped working. It lasts ten minutes. Ask the assistant to call get_conversion again for a new one. After 24 hours the files are gone; convert the file again.

Claude Desktop doesn't show the local tools. Check that claude_desktop_config.json is valid JSON, use the full path to intosql, and restart Claude Desktop.

"The MCP server stopped." intosql mcp exits when its client closes the connection. It isn't meant to run on its own in a terminal.