datatomcp MCP sunucusu

---

datatomcp is one governed place for a team's data, that every AI agent can ask: Google Sheets, CSV and Excel files, PostgreSQL and MySQL, Stripe, HubSpot and Airtable. Connect your data once. Agents do the setup work: they propose the definitions of what each number means, the links between datasets and the business brief, and a person approves them in the app. Every question an agent asks is recorded. Agents (Claude, ChatGPT, Cursor, n8n and others) query it over MCP or REST, and a database computes the answer, so totals are exact. News: datatomcp is now an AI-native data layer (4 Nov 2026).

• Hosted. Nothing to install or run. One URL works in every client that supports remote MCP servers.
• OAuth sign-in. No API key to paste into your agent. You can revoke access at any time.
• Exact, versioned results. Counts, sums and averages come from DuckDB. Every result names the dataset version it came from and links to the rows behind it.
• Shared definitions. Approved definitions like paid_revenue and a glossary of your team's words. For a business number, an agent first uses the approved definition and checks that it ties; when none exists, it proposes one and a person approves it. The answer says who approved the definition.
• Agents propose, people approve. Definitions, links, glossary terms and the business brief that agents propose wait under To approve in the app. Only approved ones reach agents. Each agent connection is set to Build (can propose) or Analyze (reads only).
• Audit trail. Every agent question is kept: who asked, from which agent, the query, the data version. Repeated questions are compared. Export as CSV.
• Linked datasets. Orders, customers and regions in one question, with unmatched rows counted.
• Charts and dashboards. The chart tool draws the answer in the chat (52 chart types, as an MCP Apps view). Agents make dashboards from your questions and refresh them later; dashboards never change your data.
• Workspaces. Owners, editors and viewers; each agent connection can be limited to some workspaces.
• Private Google Sheets. You pick the sheet in Google's file picker, and datatomcp can read that file only. Pro syncs within 5 minutes of an edit.
• Import report. Bad lines, numbers stored as text and mixed date formats are reported before any agent sees the data.
• Read-only data. Agents can query your data, but they can't change it or run raw SQL. The only tools that write create dashboards, send your team a note, or propose something a person must approve.
• Hosted in the EU. Your data is never used for training. AI features inside datatomcp (cleaning suggestions, an assistant on your own key) are off until you turn them on.

This repository holds the public documentation and client configuration. The service itself runs at app.datatomcp.com.

Server

URL	https://app.datatomcp.com/mcp
Transport	Streamable HTTP
Auth	OAuth 2.1 (dynamic client registration and client ID metadata documents)
Registry name	com.datatomcp/datatomcp

Quick start

1. Sign up at app.datatomcp.com (free, no card).
2. Add a dataset: upload a CSV or Excel file, choose a Google Sheet, or connect PostgreSQL, MySQL, Stripe, HubSpot or Airtable. New accounts start with an example dataset.
3. Add the server to your client (below), sign in when asked, and ask a question about your data.

Tools

Tool	What it does
workspace_overview	Every dataset with what its columns mean, the approved definitions, the links between datasets and example questions, in one call
list_datasets	Your datasets, with row counts and freshness
describe_dataset	Columns, types, example values, approved definitions, glossary, links to other datasets, data warnings and today's date
query_rows	Filter, sort, select columns and paginate (up to 1,000 rows per call)
aggregate	Group by columns; count, distinct count, sum, average, min and max, or a saved definition by name
distinct_values	The real values in a column, so filters match exactly
chart	Draws a query's answer as a chart in the conversation, with the takeaway as the title
create_dashboard	Makes a dashboard (a one-page board or a short story of slides) in your workspace
list_dashboards	The workspace's dashboards
refresh_dashboard	Re-runs a dashboard on the current data
report_issue	Sends your team a note about a data problem
propose_definition	Proposes a metric definition (e.g. revenue); it waits until an owner or editor approves it
propose_link	Proposes a link between two datasets, checked against the data first
propose_term	Proposes a glossary term for the team's words
propose_brief	Proposes the workspace's business brief (what the company sells, to whom, currency, fiscal year)
connect_source	Starts a connection to a database, app or file (imports a public link at once)
activity	What agents asked and changed in the workspace recently

Filters support eq, neq, gt, gte, lt, lte, in, not_in, contains, starts_with, between, is_null and not_null.

Client setup

Claude (Claude.ai and Claude Desktop)

Settings → Connectors → Add custom connector. Paste https://app.datatomcp.com/mcp, name it datatomcp, connect, and allow access.

Claude Code

claude mcp add --transport http datatomcp https://app.datatomcp.com/mcp

Then run /mcp, choose datatomcp and sign in.

ChatGPT

Settings → Apps & Connectors → Advanced, and turn on developer mode. Create a connector with the URL above and OAuth authentication, then sign in.

Cursor

Settings → MCP → Add new server. Choose type HTTP and use the URL above. Cursor asks you to log in the first time.

VS Code

.vscode/mcp.json:

{
  "servers": {
    "datatomcp": { "type": "http", "url": "https://app.datatomcp.com/mcp" }
  }
}

Windsurf

~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "datatomcp": { "serverUrl": "https://app.datatomcp.com/mcp" }
  }
}

Cline

Open the MCP Servers panel → Remote Servers, name it datatomcp, and use the URL above. Or add it to cline_mcp_settings.json:

{
  "mcpServers": {
    "datatomcp": { "type": "streamableHttp", "url": "https://app.datatomcp.com/mcp" }
  }
}

Antigravity

Customizations → MCP servers → View raw config (mcp_config.json):

{
  "mcpServers": {
    "datatomcp": { "serverUrl": "https://app.datatomcp.com/mcp" }
  }
}

Click Refresh, then Authenticate.

OpenCode

opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "datatomcp": { "type": "remote", "url": "https://app.datatomcp.com/mcp" }
  }
}

Codex

codex mcp add datatomcp --url https://app.datatomcp.com/mcp
codex mcp login datatomcp

Gemini CLI

gemini mcp add --transport http datatomcp https://app.datatomcp.com/mcp

Then run /mcp auth datatomcp.

Clients with local (stdio) servers only

Use mcp-remote as a bridge. It handles the sign-in:

{
  "mcpServers": {
    "datatomcp": { "command": "npx", "args": ["-y", "mcp-remote", "https://app.datatomcp.com/mcp"] }
  }
}

Example

How many orders came from France in September, and what was the revenue?

The agent calls describe_dataset to learn the columns, then one aggregate call:

{
  "dataset_id": "ds_orders",
  "where": [
    { "column": "country", "op": "eq", "value": "France" },
    { "column": "order_date", "op": "between", "value": ["2026-09-01", "2026-09-30"] }
  ],
  "metrics": [
    { "fn": "count", "as": "orders" },
    { "fn": "sum", "column": "total", "as": "revenue" }
  ]
}

On our 50,000-row test sheet, it answers with 277 orders and €50,279.50 in revenue. The tool result is 32 bytes, not the whole sheet.

REST API

The same datasets, definitions and query engine are available over REST with an API key, for apps and automations such as n8n, Make or Zapier. Reference: datatomcp.com/docs/api. OpenAPI 3.1: datatomcp.com/openapi.json.

Plans

	Free	Pro ($12 or €12 a month)
Datasets	3	25
Rows per dataset	100,000	2,000,000
Queries per day	1,000	50,000
Workspaces	1, 2 people	3, 5 people each
Audit history	30 days	90 days
Google Sheets sync	On demand	Within 5 minutes

Guides

• Connect Google Sheets to Claude
• Google Sheets MCP server
• CSV MCP server
• Analyze large CSVs with AI
• Shared definitions for AI agents
• Audit what your AI agents asked

Support

Questions, bugs or a file that doesn't import the way you expect: hello@datatomcp.com, or open an issue in this repository.

License

The documentation and configuration in this repository are MIT licensed. The datatomcp service is proprietary.