A familiar starfish, with more personality
The original starfish expressed the hub in DBHub: several arms meeting at one center, like databases connected through one server.ClickHouse joins the hub
Until now, DBHub focused on operational databases: PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, and SQLite. ClickHouse extends that support to analytical workloads. Useful questions often need both. Suppose you want to know which paying customers stopped using a feature after a release. PostgreSQL might hold accounts and subscription status, while ClickHouse holds the product events. An assistant can look up paying accounts in PostgreSQL, query usage trends in ClickHouse, and combine the results in one conversation. With multiple sources configured, one DBHub process gives the assistant access to both. Each SQL query runs against its selected source; the assistant connects the findings across queries. You can investigate customer activity without setting up a separate MCP server for each engine.PostgreSQL parsing, coming next
Alongside this release, we have replaced PostgreSQL’s keyword-based SQL classifier with PostgreSQL’s parser throughlibpg-query.
The parser lets DBHub inspect the structure of a statement, including writes nested inside a WITH query, instead of inferring its behavior from keywords. Statements it cannot parse are rejected. This makes classification more reliable as SQL gets more complex. PostgreSQL’s read-only transaction guard remains a second layer of protection.
We will evaluate whether suitable parsers can bring the same approach to the other database engines. See read-only mode for how the protections work.
More databases, still minimal
Our design goal is an out-of-the-box experience with little setup and little context overhead. The default interface still has just two tools:search_objects to discover database objects and execute_sql to query them.
DBHub still has the smallest tool-definition footprint among the three database MCP servers in our benchmark:
All counts use the
cl100k_base tokenizer on tool names, descriptions, and input schemas. They compare the configurations documented in the benchmark, whose capabilities differ. They measure tool definitions, not total session context or cost.
Since 1.0.0
From the 1.0.0 tag to our November 21 snapshot:- 187 commits on
main, including 139 through the 2.0.0 tag. - 49 issues closed.
- 7 hours 18 minutes median time to close, measured from issue creation to closure across those 49 issues.
