This SEP introduces negotiated tool result variants. Tools advertise supported output representations, clients select the outputs needed for each call, and hosts route each result by audience.

Motivation and Context

The current relationship between content and structuredContent is ambiguous. Clients handle them differently, which can duplicate model context, discard structured results, or route data to the wrong consumer.

This proposal adds explicit audiences and request-time selection for assistant, user, programmatic, and extension outputs. It also gives each structured variant its own schema.

How Has This Been Tested?

This PR is a specification proposal only. The SEP includes key details of how a reference implementation and conformance testing plan covering selection, routing, schema validation, errors, MCP Apps, and legacy compatibility could work. Based on feedback, a full implementation and testing will follow.

Breaking Changes

Yes, for protocol the next version (assumed to be 2026-12-01) and later:

• Tool.outputVariants replaces the legacy outputSchema.
• tools/call gains requestedOutputs.
• CallToolResult.outputs replaces content and structuredContent.
• Earlier protocol versions retain the legacy wire shape.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update

Checklist

• I have read the MCP Documentation
• My code follows the repository's style guidelines
• New and existing tests pass locally
• I have added appropriate error handling
• I have added or updated documentation as needed

Additional context

This follows maintainer feedback on SEP-2200 to replace implicit result interpretation with client-selected polymorphic results. The proposal targets the 2026-12-01 protocol revision and is seeking a sponsor.

This SEP was drafted with GitHub Copilot and Codex based with several rounds of human-led feedback and iteration grounded in various pieces of feedback and community discourse.