Data Studio Agent

Let your AI coding agent securely access all your databases, in plain language.

Local-first. Enterprise-grade security. Open source.

📖 Product Page · npm · dockit · sqlkit · Releases

English · 简体中文

This repository contains the Data Studio MCP Server, a Model Context Protocol server that gives AI coding agents direct access to your databases through the dockit and sqlkit desktop apps.

• SQL (via sqlkit): 70+ databases (PostgreSQL, MySQL, SQL Server, Oracle, SQLite, DuckDB, ClickHouse, Snowflake, BigQuery, and more)
• NoSQL (via dockit): Elasticsearch, OpenSearch, MongoDB, DynamoDB

Features

• Any AI coding agent. Claude Code, Cursor, Windsurf, OpenCode, Codex, Cline, Pi, Qoder, GitHub Copilot, or any MCP client.
• Any OS. macOS, Windows, Linux.
• Any LLM model. Bring your own provider. No lock-in.
• One MCP server, one config. Routes to both SqlKit (SQL) and DocKit (NoSQL) bridges over localhost.
• Enterprise-grade security. See below.

Quick start

1. Prerequisites

Install and launch dockit and/or sqlkit, add a database connection, and make sure Settings → MCP Bridge → Auto-start is enabled (it is by default). Install both apps for the full SQL + NoSQL tool set.

2. Install the MCP server

npm install -g @geek-fun/data-studio-mcp

Or run it without installing (npx downloads it on first run):

npx -y @geek-fun/data-studio-mcp

3. Add it to your AI tool

OpenAI Codex, one command:

codex mcp add data-studio -- npx -y @geek-fun/data-studio-mcp

Claude Code, one command:

claude mcp add --transport stdio data-studio -- npx -y @geek-fun/data-studio-mcp

Cursor. Create .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "data-studio": {
      "command": "npx",
      "args": ["-y", "@geek-fun/data-studio-mcp"]
    }
  }
}

Windsurf. Create ~/.codeium/windsurf/mcp_config.json (global only):

{
  "mcpServers": {
    "data-studio": {
      "command": "npx",
      "args": ["-y", "@geek-fun/data-studio-mcp"]
    }
  }
}

OpenCode. Add to opencode.json (project) or ~/.config/opencode/opencode.json (global):

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "data-studio": {
      "type": "local",
      "command": ["npx", "-y", "@geek-fun/data-studio-mcp"],
      "enabled": true
    }
  }
}

Any other MCP client. Register a stdio server with command npx and args -y @geek-fun/data-studio-mcp.

4. Tune permissions (optional)

Open Settings → MCP Bridge in dockit/sqlkit to control what the agent can do:

5. Start asking

Use plain language. The agent queries your databases for you:

• "List all tables in my PostgreSQL database"
• "Show me the last 10 orders from the Elasticsearch index orders*"
• "Find all users older than 30 in MongoDB"
• "Run this query and explain the results"

The agent reads schemas, runs queries, and explores your data, then shows you every query it executed.

Enterprise-grade security

The LLM gets broad access to your data, but it never sees your credentials. The policy model gates every capability by risk level.

• Credentials never leave the apps. The LLM only ever sees an opaque connection_id. Real credentials are resolved inside dockit/sqlkit and never cross the MCP boundary. Your passwords and keys stay on your machine, in your app.
• ID-based resource access. Agents access databases strictly by connection ID. Credentials never appear in prompts or tool arguments, so there is no path for the model to obtain or exfiltrate connection secrets.
• Three-tier permission model. Read Only / Data Read-Write / Full Access modes gate every capability by risk level, with per-connection overrides. You can mark any connection read-only or allowlist specific actions.
• Explicit user confirmation. Destructive operations (DELETE, DROP, TRUNCATE) surface as Ask in the policy. The client prompts the user for explicit confirmation before anything destructive runs.
• Action-level statement classification. SQL is parsed and classified by statement kind (Read / Write / Delete / DDL) before execution. Write-only tools reject DELETE statements; delete tools reject DDL.
• Local-only bridge. The bridge binds to 127.0.0.1 exclusively. It is unreachable from other machines, with no server to host and no API keys to manage.

Tools

All tools follow the data_studio__{backend}__{action} convention. The User confirmation column shows which operations surface an explicit confirmation prompt in your AI client before they run.

How it works

code agent (Claude Code / Cursor / OpenCode ...)
    |
    | MCP stdio protocol
    v
@geek-fun/data-studio-mcp   ← npm package (pure TypeScript)
    |
    | HTTP (localhost)
    +----------------+----------------+
    v                v                |
dockit:9120    sqlkit:9121            |
(NoSQL bridge)  (SQL bridge)          |
    |                |                |
    v                v                |
Elasticsearch    PostgreSQL           |
MongoDB          MySQL                |
DynamoDB         SQL Server           |
OpenSearch       SQLite               |

The MCP server is a thin routing layer. All database drivers, SSH tunnels, and connection management live in the desktop apps, which expose a local HTTP bridge (127.0.0.1 only). The MCP server auto-discovers running backends via each app's port file.

For developers

This repository also contains the data-studio-agent Rust framework, the shared AI agent loop (provider adapters, streaming, tool calling, context compaction) that powers the built-in assistants in dockit and sqlkit.

See crates/data-studio-agent/README.md for installation, architecture, and integration guides.

License

Apache 2.0. See LICENSE.