Connect your data visually. Ask AI through MCP.
Bufflehead now has native MCP support. Connect to your databases visually, save the connections, and let your AI assistant use them to answer questions.
You configure the connection in Bufflehead. Your assistant can then discover those saved connections, inspect their tables and columns, and run SQL through the Model Context Protocol. The same connection setup serves both the visual app and the assistant.
Connect visually. Save once. Ask questions.
Start in Bufflehead: add a supported database connection, check that you can browse its tables, and save it as a bookmark. Give it a recognizable label, such as Reporting Postgres, so you and your assistant can identify the right source.
Next, open Help → Connect an AI Assistant…. The setup dialog helps you connect Bufflehead to your MCP client:
- Claude Desktop: on supported macOS and Windows installations, preview and approve the configuration entry with Add Automatically…, then quit and reopen Claude Desktop.
- Claude Code: use Copy Command and run the supplied command in your terminal.
- Other clients that support local stdio MCP servers: use Copy JSON for the launch configuration.
The automatic Claude Desktop configuration option is unavailable on Linux and in the Mac App Store build; the dialog also provides launch details for manual configuration.
Once connected, start with a simple question: “What Bufflehead bookmarks do I have?” Your assistant can list the saved database connections before choosing one for an investigation.
From a question to the data behind it
Imagine you have saved a reporting database containing orders. You could ask:
Use Reporting Postgres. Which products had the most completed orders last month? Inspect the schema first, confirm how completed orders are represented, and show me the SQL and date range you use.
This is an example prompt, not a recorded demo. The assistant has three tools available to work through a question like this:
list_bookmarksdiscovers saved database connections by label and type.get_schemalists tables and views, then inspects the columns of a selected table.run_sqlexecutes a SELECT query against a selected bookmark and returns columns, rows and the total result count.
The assistant can use the schema to find relevant fields, query the selected source, and explain the result. If “completed” or “last month” is ambiguous, it should clarify the definition before presenting an answer.
Several sources, one assistant connection
You can save multiple supported database connections in Bufflehead and have your assistant choose among them by bookmark name. The shared connection pipeline covers direct PostgreSQL and MySQL connections, optional SSH tunnels, BigQuery, and AWS gateway connections.
For example, you might keep an operational database and an analytics warehouse as separate bookmarks. Tell the assistant which one to use, or ask it to list the available sources first. Each SQL call targets one bookmark; this does not create a cross-database SQL join.
MCP currently exposes saved database bookmarks. Opening a local file in the visual viewer does not automatically make that file an MCP source.
The window can be closed
Your MCP client launches Bufflehead in headless mode using the installed app's executable. It reads your saved bookmarks and opens database connections when a tool needs them. The Bufflehead window does not need to stay open.
This is a local stdio MCP server: the client and Bufflehead communicate through the launched process. You do not need to deploy a separate shared MCP service or manage a listening port for this connection.
Keep access scoped to the question
The MCP query tool is designed for SELECT queries. PostgreSQL connections open with a read-only session default; use appropriately restricted database credentials as well. Returned results default to 100 rows and are capped at 1,000 per call. That is a response limit, not a limit on database work or query cost.
The bookmark-listing tool does not return passwords or keys. Query results do go to your AI client and may be sent to its model provider, so choose credentials, sources and tools that match your organization's data policy.
This is an early release. The documented client test is Claude Desktop on macOS; Claude Code, Windows, Linux and the Mac App Store MCP workflow still need end-to-end validation.
Download Bufflehead, save a database connection, and open Help → Connect an AI Assistant…. Then ask your assistant what data is available—and start with a question you can verify.