An open-data portal agent must query the city’s dataset metadata catalog through a controlled interface. Which tool approach is most appropriate?

An open-data portal agent that must query the city's dataset metadata catalog through a controlled interface should expose custom functions as tools that query that catalog. Tool-mediated catalog access gives the model a governed contract (search, filter, describe datasets) without granting raw infrastructure privileges, and answers stay tied to official metadata rather than scraped guesses. That is the appropriate Foundry-style tool approach for municipal open-data discovery. Registering a generic web browsing tool for the agent to navigate the portal still bypasses the structured, authoritative metadata interface and can return loosely rendered page content instead of the governed catalog records. Connecting the agent to the catalog using a file search tool over exported CSV dumps bypasses the controlled query interface and couples the agent to stale schema details and snapshot data a governed metadata function is meant to abstract away. Configuring a retrieval tool that searches dataset contents for metadata queries cannot scale for catalog discovery, may return content-level matches rather than governed metadata records, and never performs live catalog queries. On the exam, controlled catalog tools beat generic browsing, file-search-over-CSV, or content-retrieval-tool approaches; prefer custom functions over the catalog API/metadata service. Caveat: tool results still need permission filtering—some internal dataset metadata may not be public even on an open-data portal agent. Operational check: register catalog query tools, ask for datasets by topic, verify responses come from the city catalog function (not a generic web browser, file-search tool over CSV dumps, or content retrieval tool), and confirm the agent has no database-server shell permissions.