MCP tool inputSchema and JSON Schema 2020-12: Sume's tool schemas

The MCP 2026-07-28 changelog loosens inputSchema and outputSchema to allow any JSON Schema 2020-12 keywords, and adds $ref resolution requirements. For a Sume client that changes little: Sume's tool inputs come from one objectSchema helper that emits type: "object", properties, required and additionalProperties: false.

Spec text is from the MCP changelog; Sume details are from the MCP server source and the tools and gates docs, read 2026-10-01.

What did the spec loosen?

Per the changelog: inputSchema and outputSchema may use any JSON Schema 2020-12 keywords, structuredContent may be any JSON value, and $ref resolution requirements and resource bounds for composition keywords were added. A client should therefore be ready to see keywords beyond plain type and properties from servers that use them.

What shape do Sume's tool schemas have?

How do I read one tool's contract?

Call tools_schema with the tool name. The docs describe it as fetching one tool contract by name, and say to always discover the live contract with tools_list and tools_schema instead of assuming parity with the HTTP API. For big tool sets see progressive discovery.

What should my client do?

Validate with a full 2020-12 validator rather than a hand-rolled type and properties reader, so a keyword like anyOf is handled the same on every server. Do not hard-code a schema copy: it can change, and additionalProperties: false means a stale extra key is rejected rather than ignored. Paid and write tools also need idempotency_key, which is part of the contract you fetch.