Skip to main content
November 17, 2026 Every MCP server’s tool definitions occupy the agent’s context window, leaving less room for your schema, query results, and code. DBHub is built to keep them small. This post repeats our December 2025 benchmark, almost a year later, with current versions.

Results

We added each server to Claude Code in its default configuration and read the token count of its tool definitions from the /context command. MCP Toolbox is the other popular multi-database server, and Supabase MCP is the most popular vendor-specific one.

Why DBHub uses fewer tokens

DBHub’s individual tool definitions are not shorter. The average tool costs about the same in all three servers: The difference is the default tool count. DBHub has two:
  • execute_sql runs SQL statements.
  • search_objects finds schemas, tables, columns, indexes, and procedures. It uses progressive disclosure: you pass a name pattern and a detail level, so one tool covers every object type.
The other two servers expose one tool per operation. MCP Toolbox has a tool per object type (list_tables, list_views, list_indexes, list_triggers) plus database administration tools. Supabase MCP covers the whole Supabase platform: projects, branches, edge functions, and notebooks. Its create_notebook tool alone uses 1,700 tokens, more than three times DBHub’s total. Both servers do more than DBHub, and both let you disable tools. The numbers above are for the default configuration.

Why all three numbers are lower than last year

DBHub itself barely changed: its two tool definitions shrank 12%, from 1,664 characters to 1,468, while the reported token count fell by two thirds. Most of the drop comes from two changes in Claude Code. The first is token accounting. In December 2025, /context reported 607 tokens for a 496-character execute_sql definition. A definition cannot have more tokens than characters, so that figure included per-tool overhead. Today the same tool is reported at 119 tokens, and MCP Toolbox dropped by a similar amount, about 400 tokens per tool. The second is MCP tool search, now enabled by default. It applies deferred tool loading: only tool names and server instructions load at session start, and a definition enters the context window when the agent discovers the tool through search. Deferral postpones the cost; it does not remove it. Once the agent works with the database, the definitions load: 468 tokens for all of DBHub, thousands for the other two if the agent uses all their tools. With ENABLE_TOOL_SEARCH=false, or in an MCP client without tool search, every definition loads upfront. DBHub’s lead has also grown, from about 14 times for both servers in December 2025 to 16 and 25 times today.

How to reproduce

We used Claude Code 2.1.326. These commands add the three servers and print the context report:
MCP Toolbox connects to PostgreSQL at startup, so it needs a reachable database. For Supabase MCP we used the local npm server, not the hosted one. To try DBHub, see the installation guide.

References