Five steps from a problem to its solution.

Help me design and build a first working prototype for this problem in Anywhere.

Problem: Developers can't give a coding agent read-only access to a server or database without approving every command by hand.
Who it hurts: People who use an AI coding agent to read logs, inspect files and query databases on real systems.
The moment it hurts: During routine debugging, when the agent asks permission for the tenth repetitive read command in a row.
What people do today: People either approve each command, or rely on lists of allowed command names, which an agent can get around with other tools.
Why it is still unsolved: Filtering command text is fragile. The safer answer, separate accounts and database roles, takes setup most people skip.
What solved looks like: A small, documented setup (script or container image) that gives an agent a shell and a PostgreSQL connection where reads run without prompts and every write is refused by account permissions, not by text matching. Checkable with a test list of 30 commands and queries (15 reads, 15 writes, including tricky ones like find -delete or SELECT ... INTO) where every write fails and every read works.
Where: Anywhere

First, list what you would need to learn from the people affected before building, as interview questions.
Then propose the simplest version that would help, the screens or messages it needs, and the data it would use.
Keep it cheap to run, accessible, and respectful of people's privacy. Flag anything you are unsure about instead of guessing.