The MCP server work started with a CEO ask about automating a manual client-import process. Somewhere in the middle of testing extraction approaches, the idea stopped being about PDF import and became a way to give AI tools access to systems a plain agent can't reach. I built two for my own workflow.
Questions about the database
The first lets me ask about our database directly: the tables, how they relate, whether a table is missing the history table and trigger scripts our conventions require, whether a column is missing where it shouldn't be. I've stopped opening SSMS for the small questions. A regular agent handles this badly because the schema context is token-intensive and hits context limits fast. The MCP server queries the database directly and builds a schema graph of the relationships instead of holding the whole schema in context. Next step I haven't built: wiring it into the SQL convention agent so it fixes what's missing, not just flags it.
Questions about the codebase
The second uses SVN to answer questions about the code. On a support ticket, I can ask who last edited the affected file, how recent the commit was, and whether it's related to the ticket. I can have it review recent commits and flag ones that might have introduced a bug. I can hand it my working files and have it group them into logical commits with messages that describe what changed.
What it proved
I was surprised how much of this I'd been missing. AI with real access to the systems around the code is as useful as it sounds. Both servers are now accessible to every developer and have been demoed to the whole team; the limit on heavier day-to-day use is token budget, not access.
Update: the demo is what happened next. See the follow-up on scaling these servers with AWS AgentCore, Bedrock, and Lambda.