Servers can now publish agent skillA reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.Full definition → folders over MCP and clients can list and verify them. The release also changes validation behavior: number-typed x-mcp-header is rejected and outputSchema results are no longer wrapped in a result key, which can break existing servers on upgrade.
Key takeaways · AI-distilled
In MCPThe Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.Full definition → TypeScript SDK 2.4.0, a tool that puts x-mcp-header on a number parameter fails on both sides: the client leaves it out of the tool list and the server answers calls with error -32020. The notes say to use z.int() or drop the annotation.
On 2025-11-25 connections, an outputSchema built from z.object(...).meta({ id }) no longer wraps results in { result }, so clients have to read structuredContent directly rather than structuredContent.result.
The bundled Ajv validator now uses fast-uri 3.1.8, which rejects a $id or $ref with an invalid URI scheme such as my_tool:output, so schemas named that way are refused after the upgrade.
Skills over MCP (SEP-2640) works on both the 2025-11-25 and 2026-07-28 protocol revisions, but the release marks skills.attach as experimental and says it may change in a minor release.
Terms in this piece · Glossary
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
