MCP tool schema and result contract

tools/list returns each operation’s current schema. For example, native mode describes wp_get_post as:

{
  "tools": [
    {
      "name": "wp_get_post",
      "description": "…",
      "inputSchema": {
        "type": "object",
        "properties": {"id": {"type": "integer"}},
        "required": ["id"]
      }
    }
  ]
}

The public list exposes the MCP name, description and input schema. Native mode also supplies read-only, destructive and idempotence hints. These are descriptive hints, never permission grants; the server enforces scopes, capabilities and write gates independently.

Call a tool with:

{
  "method": "tools/call",
  "params": {
    "name": "wp_get_post",
    "arguments": {
      "id": 123
    }
  }
}

Required input fields are validated before the callback. The standard transport normalizes common string forms of dry_run; native mode requires a JSON boolean.

On success, the JSON-RPC result contains content with a JSON text item and isError: false. The standard transport pretty-prints that JSON. Native mode also returns structuredContent with the same tool-specific data; prefer that object when present. Do not depend on JSON whitespace.

In standard mode, domain errors are JSON-RPC errors. Native tool execution errors use isError: true; protocol errors remain JSON-RPC errors. Do not parse an error message as a successful tool result.

Tool availability is runtime-dependent. Use exact discovered names and do not assume similarly named aliases.

Rediscover input and output schemas after reconnecting. A familiar tool name alone does not establish compatibility with a different connection implementation. If a connection returns MCP structuredContent, read that object directly; otherwise decode the documented text result. Check isError before treating either representation as successful data.

Related guides: error contract, read versus write, and tool reference chapters.

Native site identity and validation

pom_get_site_context returns contract_version: 2, the WordPress version and a site object containing site_id, site_url and blog_id. Existing tool calls do not require a new identifier. They optionally accept site_id to confirm they target the discovered site; a mismatch is rejected. The identifier is not a credential, and execution always remains scoped to the connection’s site.

Native mode validates described property types and required arguments through WordPress Abilities API. Undescribed extra fields remain accepted for compatibility; operation callbacks only use their supported fields. Batch tools retain their domain validation and partial-result behavior. Inspect nested results as well as the top-level isError; a successful batch envelope does not establish that every row succeeded.

Output schemas are not imposed on the variable legacy payloads. Output validation cannot undo an already completed mutation. After a timeout or ambiguous result, read the target before retrying a write.