I built a tool similar to this end of last year with the idea of making it easier for agents (and humans) to work with postgres and do things like getting summaries when they want to. But then, I realized that these models are trained and really know how to use all parts pg and it was easier to let them do that. There can be cases for this, but the cost in extra tokens and steps using pg vs a tool are so little, it wasn't worth it to keep going.
Hi HN, I’m Alex, a longtime PostgreSQL and Rails/Go/Rust engineer.
I built pgbot because I often needed a quick explanation of what was happening inside a database without first deploying a dashboard, collector, telemetry stack, or separate database.
pgbot is an open-source 5.9 MB Go binary that connects directly to PostgreSQL and returns useful findings in seconds:
pgbot inspect — database health
pgbot queries — expensive queries
pgbot indexes — risky and unused indexes
pgbot ask — answers for humans and agents
MCP — lets coding agents inspect PostgreSQL directly
The diagnostic findings are computed deterministically in Go. They are not invented by an LLM.
pgbot is also read-only by design: it inspects the database but never applies changes. An agent can explain and prioritize the evidence, but it cannot silently modify production through pgbot.
I built pgbot because I often needed a quick explanation of what was happening inside a database without first deploying a dashboard, collector, telemetry stack, or separate database.
pgbot is an open-source 5.9 MB Go binary that connects directly to PostgreSQL and returns useful findings in seconds:
pgbot inspect — database health pgbot queries — expensive queries pgbot indexes — risky and unused indexes pgbot ask — answers for humans and agents MCP — lets coding agents inspect PostgreSQL directly
The diagnostic findings are computed deterministically in Go. They are not invented by an LLM.
pgbot is also read-only by design: it inspects the database but never applies changes. An agent can explain and prioritize the evidence, but it cannot silently modify production through pgbot.
Source: https://github.com/pgrundev/pgbot
I’d especially appreciate feedback about missing diagnostic checks, false positives, and the read-only safety model for coding agents.