An MCP server that serves a Structured Data Extraction agent should match each responsibility to the primitive designed for it. Browsing a contract catalog is readable content selection with no side effects — that belongs in an MCP resource such as contracts://open, which the agent reads once before deciding what to extract. Pulling a specific contract's full text is still read-mostly but is parameterized per contract and returns a bounded response, which makes it a natural tool (get_contract(id)) with a schema'd response the agent can pipe into the extraction step. Writing a JSON record into the ERP system is an action with side effects and must be a tool (create_erp_record(payload)) ideally with a strict schema to guarantee the extracted payload matches ERP's expected shape. Option A partitions the responsibilities correctly. Option B abandons resources entirely and reintroduces speculative catalog browsing through tool_use loops. Option C miscategorizes a side-effectful ERP write as a resource. Option D assigns browsing to prompts, which are reusable message templates rather than catalog surfaces.

Plain-Language Explanation:

Designing an MCP server is like designing a warehouse workflow. The shelf labels (resource catalog) let workers see what is on each aisle without opening every box. The 'fetch box N' request (tool) physically retrieves one labelled box. The 'send completed crate to dispatch' action (tool with side effect) changes the warehouse state. Mixing these up — labels that require a box to be opened, dispatches done by reading a shelf tag — creates chaos.

Option Analysis:1622

A：Wrong because tool-only designs lose the cheap catalog-browsing benefit of resources and reintroduce speculative tool calls.
B：Wrong because prompts are reusable message templates for client insertion, not URI-addressed browsable catalogs; resources own that role.
C：Correct because catalog-as-resource plus action-as-tool with strict schemas is the documented best-practice partition for extraction agents.
D：Wrong because ERP write is an action with side effects, which the MCP spec explicitly routes through the tool primitive, not resources.