Baserow has a built-in native Model Context Protocol (MCP) server for connecting large language models to Baserow workspaces. It is aimed at users who want to manage databases using natural language without writing SQL, formulas, or API calls.
In terms of main capabilities, the MCP server supports performing standard CRUD operations through natural language prompts, and the official recommendation is to use high-parameter models for the best performance. The typical uses officially listed include: querying databases without writing SQL or formulas; batch updating records through conversational commands; generating reports by asking questions in natural language; and automating data entry processes with AI assistance. It can be used with LLM clients such as Claude, Cursor, and Windsurf.
In terms of how it works, each MCP server endpoint generates a unique URL, and that URL grants the selected LLM secure access to the workspace. The LLM converts natural language requests into Baserow operations to perform database operations on the user's behalf. Unlike cloud-based solutions, it does not pass any private data to third-party servers, streams only the requested database rows to reduce token consumption in the context window, and runs entirely behind the local firewall through standard JSON configuration. The official documentation states that the currently compatible LLM clients are Claude Desktop, Cursor, and Windsurf, and that any client supporting the MCP specification can theoretically also connect through this endpoint.
Prerequisites before use include: an active Baserow workspace; a supported LLM client installed (Claude Desktop, Cursor, or Windsurf); and administrator or corresponding workspace permissions.
Setup steps: Step one, access the MCP settings, click the workspace name in the top navigation bar, select My Settings from the dropdown menu, and then go to the MCP Server tab. Step two, create an endpoint, click Create Endpoint, enter a descriptive name (for example, "Marketing Database Assistant"), select the workspace to connect, and create the endpoint. Step three, copy the MCP URL. Baserow will generate a unique endpoint URL, and the official guidance warns that it should be treated like a password because it can fully modify workspace data. Step four, configure the LLM client. After creating the endpoint, you can click More Details View the configuration options and select from the available capabilities (such as Get table schema, List table rows); then proceed according to the chosen client: In Claude Desktop, open Settings (⌘+, on Mac), go to the Develop tab, click Edit Config to open claude_desktop_config.json and add the JSON configuration; In Cursor, open Settings (⇧+⌘+J on Mac), go to the MCP tab, click Add MCP Server and paste the JSON configuration; In Windsurf, open Settings → Advanced Settings (or open the settings page via the command palette), scroll to the Cascade section, add a new server or view the raw mcp_config.json configuration file and add the JSON configuration; in the configuration, replace YOUR_MCP_URL_HERE with the MCP URL copied from Baserow settings.
Regarding security best practices: the MCP URL contains credentials that provide full access to workspace data and should be kept as securely as a password; if it is ever leaked, it should be revoked and regenerated immediately; use descriptive endpoint names to track their purpose; create different endpoints for different use cases. The official guidance also lists behaviors to avoid: publicly publishing or sharing the MCP URL in version control, storing it in an insecure location, having multiple team members share the same endpoint, and posting screenshots that contain the full URL.
Regarding applicable scenarios, the official list includes operating a database in natural language without programming, batch updates, generating reports by asking questions, and automating data entry. Regarding limitations, the official guidance states that the AI's interpretation depends on model quality and prompt clarity, and that critical operations should always be verified, especially batch updates or deletions; it is recommended to use a high-parameter model and review operations before confirming destructive changes; in addition, the currently explicitly compatible clients are Claude Desktop, Cursor, and Windsurf. Regarding pricing, charges, and more feature details, the materials do not currently provide this. If users need further help, the official guidance is to visit the Baserow community or contact support.
