Interface ToolDefinition

Extension metadata advertised on this tool's tools/list descriptor.

Use vendor-namespaced keys for custom data. This definition metadata is distinct from a tool callback result's _meta: it describes the tool, while result _meta belongs to one invocation. For view-bound or visibility-restricted tools, framework-owned MCP Apps keys are derived from ToolDefinition.view and ToolDefinition.visibility; those derived values take precedence over colliding entries here.

Behavioral hints for clients (readOnlyHint, destructiveHint, …).

LLM-facing description of what the tool does.

Object schema for input validation — any Standard Schema library with JSON Schema conversion (StandardSchemaWithJSON): zod v4, ArkType, Valibot, …. Field descriptions become LLM hints. Input is validated by the SDK before the callback runs. Emitted on the wire as inputSchema.

Unique tool identifier, e.g. "fetch-weather".

Schema for structured output (StandardSchemaWithJSON). Any JSON root is allowed — object, array, or primitive (2026-07-28 protocol). When set, the callback must return a result carrying matching structuredContent or an explicit isError result — enforced at compile time and validated by the SDK at runtime.

Alias for ToolDefinition.inputSchema. Prefer inputSchema in new code — it matches the MCP wire field name.

Human-readable title shown in UIs (falls back to name).

Bind this tool to a view for MCP Apps rendering. Requires ToolDefinition.outputSchema — the view reads the result's structuredContent typed by this schema.

Declares who may call or see the tool. Emitted as _meta.ui.visibility on tools/list. Omitted = host default (callable by the model, visible to the app). The server always lists every registered tool — hosts filter by this declaration; the server never omits tools from tools/list. "app" marks app-private helper tools callable from views via useCallTool while the host hides them from the model.