The same connection that answers questions can insert, update, delete and create tables, and the agent writes the SQL itself after reading the schema. Which account sits in the config file is therefore the decision that matters.
A Python MCP server for Microsoft SQL Server, adapted from the reference SQLite server. It reads with SELECT, writes with INSERT, UPDATE and DELETE, creates tables, lists the tables that exist and describes a table's schema — so an agent can explore a database it has never seen and then change it. A sixth tool appends findings to a running business-insight memo, so an analysis session leaves something behind besides the transcript.
- SELECT queries run against the connected database and returned to the client
- INSERT, UPDATE and DELETE over the same connection
- New tables created from the agent's own DDL
- A list of every table in the database
- The schema of any single table, column by column
- A business-insight memo the agent appends findings to as it works
A SQL Server the machine can reach, and the Microsoft ODBC driver for SQL Server installed locally. Connection details sit in a config.json beside the server: driver, host, database, and either a username and password or a trusted connection. The account in that file is the account the agent gets.