Model Context Protocol

Point your assistant
at your warehouse.

Dataki runs a remote MCP server. Add it as a connector in Claude or ChatGPT, sign in once, and they can query the sources you have connected — writing the SQL, saving dashboards and publishing embeds as you, and never seeing a database credential.

endpoint streamable http
https://dataki.ai/api/mcp

Nothing to install · no API key to paste · scopes, team restriction and spend ceiling all still apply · every query logged.

01 Connecting

Add the connector. Sign in. Ask.

Nothing to install, no config file, no API key to paste. Claude and ChatGPT take the endpoint on its own and sign you in to Dataki, where you approve what they may do and can change it later.

  1. 01

    Add a connector in your assistant and give it this URL.

    https://dataki.ai/api/mcp

    In Claude: Add custom connector.

  2. 02

    Sign in when it asks, and approve what it may reach.

    Your assistant gets its own credential, which you can revoke. Nothing is typed into a config file.

  3. 03

    Ask a question.

    If you have not connected any data yet, it will hand you a link and wait. That takes about a minute.

Step by step, for the assistant you use:

Under the hood that is OAuth 2.1 with PKCE and dynamic client registration: the client discovers where to send you, you sign in and approve the scopes, and it gets its own revocable credential bound to your account. The server holds no credential of its own and acts only as you, so anything you cannot reach, it cannot reach either.

Using a client that still wants a config file?

Claude Code, Cursor and VS Code read the same endpoint from a file. Create an API key, on any plan, under API keys in the app's account menu and paste one of these.

Claude Code · Cursor mcp.json 
{
  "mcpServers": {
    "dataki": {
      "type": "http",
      "url": "https://dataki.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_DATAKI_API_KEY"
      }
    }
  }
}
VS Code .vscode/mcp.json 
{
  "servers": {
    "dataki": {
      "type": "http",
      "url": "https://dataki.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer ${input:dataki_key}"
      }
    }
  }
}

02 In the conversation

The answer arrives as a chart you can keep.

In Claude, and in any host that draws MCP Apps, Dataki answers with a live view beside the chat: charts, tables and headline numbers drawn by the same components as a Dataki dashboard. Hosts without MCP Apps get the same answer as text.

Drawn from the warehouse
Each chart carries its query. The server runs it and draws the rows that come back, so the figures are your warehouse's, not ones the model retyped.
Checked before you see it
A field the rows do not have, text on a numeric axis, a chart that would draw nothing: the server refuses it and tells the assistant what to fix.
One click to a dashboard
Add to dashboard saves the chart as a live Dataki widget, on a new dashboard or one you already have.
Claude · dataki (MCP) render_ui

How has revenue moved this year, and which categories carry it?

Dataki's view inside Claude: a line of monthly revenue rising from about $169K to $359K over twelve months, and bars of the top six categories led by Outerwear & Coats and Jeans, each with an Add this view to a dashboard button

The view as Claude draws it, from Google's public e-commerce sample dataset.

03 Tools

20 tools, in the order you would use them.

An assistant finds the sources, reads what the columns actually mean, then queries. The rest exist because an answer nobody keeps is a message that scrolls away.

  • list_sources List data sources

    Start here — the ids every other tool expects.

  • get_data_context Read a source's tables and what they mean

    Tables, columns and types, with metric definitions, joins and column warnings.

  • record_data_context Write back what it learned

    So the next session starts where this one ended.

  • query Run a query

    Read-only SQL against one or more sources, inside your spend boundary, with a chart and table of the result where the host draws views.

  • render_ui Show something in the conversation

    Charts, tables and scorecards drawn from their queries, checked before you see them.

  • save_dashboard Save a dashboard

    Turn an answer into something that persists.

  • save_widget Add a widget

    Put one chart on a board that already exists.

  • update_widget Change a widget

    Fix a chart's query, title or look where it is.

  • remove_widget Remove a widget

    Take one off a board.

  • list_dashboards List dashboards

    The boards you can add to.

  • get_dashboard Read a dashboard

    What is on a board right now.

  • update_dashboard Update a dashboard

    Its title, theme or who can see it.

  • delete_dashboard Delete a dashboard

    Yours, when it is done.

  • publish_embed Publish an embed

    A public link, and the snippet to put it in a page.

  • watch_metric Watch a metric

    Get told when a number moves.

  • list_schedules List watches

    What is already running.

  • delete_schedule Stop a watch

    When you no longer need telling.

  • upload_file Add a file as a data source

    A CSV or Excel export, queried and charted like any other source.

  • replace_file Replace a file's contents

    Today's export in, the same table names kept: every dashboard on it shows the new rows.

  • create_file_upload Upload a large file

    A signed URL for a file too big to send in the call.

04 Verifying

Check it from a terminal first.

Before wiring it into an assistant, open a session yourself. A valid key answers with the server's name and what it offers; no key returns a 401 that tells your client where to authenticate rather than just refusing it.

More on the assistant
initialize shell 
curl -sX POST https://dataki.ai/api/mcp \
  -H "Authorization: Bearer YOUR_DATAKI_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1"}}}'
what the agent then sends query 
SELECT plan, SUM(mrr) AS mrr
FROM   billing.subscriptions
WHERE  status = 'active'
GROUP BY plan
ORDER BY mrr DESC

Free while we are in beta.

Connect a source, create a key, and point your assistant at it.

Start building for free See it in action