Typed MCP Results

Typed MCP results

Plans: same as each tool’s plan (see the Tool Catalog) Where: MCP server and the TypeScript and Python SDKs

Eleven MCP tools now return their result as structuredContent in addition to the text block, and both SDKs turn it into typed values, so you stop hand-writing types for uar_status or evidence_log.

Who benefits: integrators and engineers who call AutoCISO from scripts, agents or the SDKs.

What’s new

• Successful calls of the eleven typed tools carry structuredContent with the same masked JSON as the text block. The server advertises outputSchema on tools/list after a seven-day soak, and never for employee_resolve.
• TypeScript: typed ToolResult<XResult> per typed tool, and a new ResultShapeError. Casting .data directly to an unrelated type now needs as unknown as.
• Python: TypedToolResult[XResult] with a lazy .result that raises ResultSchemaError; the package ships py.typed.
• A typed tool that returns nothing now gives INTERNAL_ERROR; results over 256 KiB give RESULT_TOO_LARGE.

How to use it

1. Update the SDK to the next release.
2. Read res.data (TypeScript) or res.result (Python) and let your editor complete the fields.
3. Check the Result column in the Tool Catalog to see which tools are typed. The other 54 are unchanged. See TypeScript SDK and Python SDK.