MCP Server: Let ChatGPT and Claude Query Your Data Directly

Your data, meet your AI assistant

AI assistants like ChatGPT and Claude are great at analyzing data — but getting data into them is clumsy. You copy rows from a spreadsheet, paste them into a prompt, hit token limits, lose context, and repeat. Or you export a CSV, upload it, and hope the AI remembers the schema three messages later.

Starting today, csv-api supports MCP (Model Context Protocol) — an open standard that lets AI assistants connect directly to external data sources. Enable MCP on any dataset and your AI assistant can query it live: filtering, sorting, aggregating, and running SQL, all without you pasting a single row.

What is MCP?

MCP (Model Context Protocol) is an open standard created by Anthropic and now supported by OpenAI, Cursor, Windsurf, and a growing list of AI tools. It defines a way for AI assistants to discover and call external tools — similar to how a browser discovers APIs, but designed for AI agents.

When you enable MCP on a csv-api dataset, we expose a standards-compliant MCP server endpoint. AI assistants discover it, authenticate via OAuth, and get access to four tools they can call during a conversation:

Tool	What it does
`get_schema`	Returns column names, data types, and row count
`query_data`	Filter, sort, search, and paginate records
`aggregate_data`	Count, sum, average, min, max — grouped by any column
`sql_query`	Run read-only SQL SELECT queries (up to 1,000 rows)

What it looks like in practice

Once connected, you just talk to your AI assistant in plain English. The AI decides which tools to call behind the scenes.

You:   What are the top 5 cities by number of customers?

AI:    I'll query your customers dataset to find out.
       [Calls get_schema to understand columns]
       [Calls aggregate_data with metric=count, group_by=city]

       Here are the top 5 cities by customer count:
       1. Portland — 142
       2. Seattle — 98
       3. Denver — 87
       4. Austin — 76
       5. Chicago — 71

You:   Now show me the average order value for each of those cities.

AI:    [Calls sql_query: SELECT city, AVG(order_value) ... WHERE city IN (...)]

       Average order values:
       Portland: $84.20 | Seattle: $91.50 | Denver: $72.30 ...

No prompt engineering. No schema descriptions. No token limits from pasted data. The AI discovers everything it needs through MCP.

Setup: ChatGPT

Connecting a csv-api dataset to ChatGPT takes about 30 seconds:

1

Enable MCP on your dataset

Open your dataset page and click the MCP toggle. You'll see the MCP server URL appear.
2

Add to ChatGPT

In ChatGPT, go to Settings → Connected Apps → Add. Paste your MCP server URL.
3

Authorize

ChatGPT opens a browser window. Sign in to csv-api and click Authorize Access. Done.

Note: ChatGPT connects at the dataset level. To give ChatGPT access to multiple datasets, add each MCP URL separately.

Setup: Claude

Claude supports MCP through both claude.ai and Claude Desktop.

claude.ai (web)

Go to Settings → Integrations → Add MCP Server. Paste your MCP server URL. Claude opens a browser window for authorization — sign in and approve.

Claude Desktop (macOS/Windows)

Add your MCP server to claude_desktop_config.json:

{
  "mcpServers": {
    "my-dataset": {
      "url": "https://csv-api.com/api/v1/datasets/YOUR_ID/mcp"
    }
  }
}

Restart Claude Desktop. It will prompt you to authorize via your browser on first use.

Security and authorization

MCP connections use OAuth 2.0 with PKCE — the same standard used by Google, GitHub, and Slack. Here's what that means in practice:

Your AI assistant never sees your password. Authorization happens in your browser, and the AI only receives a scoped access token.
Per-dataset access. Each MCP session is scoped to a single dataset. The AI can't see your other datasets unless you explicitly authorize them.
Read-only. All four MCP tools are read-only. AI assistants cannot create, update, or delete your data.
Revokable anytime. You can revoke any MCP session from your dataset page. The AI immediately loses access.
Tokens are hashed. Access tokens are stored as SHA-256 digests. Even if someone accessed the database, they couldn't extract a usable token.

Rate limits and plans

MCP tool calls count against your plan's hourly quota, the same as REST API requests. Only tools/call requests count — initialization and schema discovery are free.

Plan	MCP queries/hr
Free	— (MCP not available)
Starter ($9/mo)	500/hr
Pro ($29/mo)	2,500/hr
Scale ($79/mo)	10,000/hr

Ideas to get you started

Upload a sales spreadsheet and ask Claude to find seasonal trends
Connect a customer list to ChatGPT and ask "which customers haven't ordered in 90 days?"
Let your team explore survey results by chatting with an AI instead of building a dashboard
Build an AI agent that queries live inventory data to answer support tickets

Try it now

If you're on a paid plan, you can enable MCP on any ready dataset right now. Upload a file (or use one you already have), flip the MCP toggle, and connect your favorite AI assistant.

Your MCP server URL:
https://csv-api.com/api/v1/datasets/YOUR_ID/mcp

For the full feature breakdown, see the MCP Server feature page. For common questions, check the FAQ. And if you're new to csv-api, start with Getting Started — you can have a dataset and MCP connection running in under two minutes.