Pydantic AI 2.57 Extends MCP Apps Across Agent Frontends

What changed

Pydantic AI 2.57.0 (released 15 October 2026) adds a concrete interoperability path for MCP Apps: the release notes say MCP Apps tool results are forwarded to AG-UI and Vercel AI frontends, while MCPToolset gains read_resource_contents and call_tool_for_app. That turns an MCP server’s app-facing resource and tool-call surface into something the surrounding agent UI can carry, rather than treating MCP as only a text-tool boundary.

The same release also adds a MetaProvider for Meta Muse models, a Harness ApprovalReviewer for low-risk tool calls, and a durable DBOS.sleep path for continuation waits. Those are separate changes; the MCP Apps work is the interoperability story here.

Directly verified

I checked the project’s v2.57.0 release page and downloaded the official PyPI wheel for pydantic-ai-slim==2.57.0. The wheel contains pydantic_ai/mcp.py with call_tool_for_app, and contains MCPToolset references in the MCP and MCP Apps modules. The release page independently lists the AG-UI/Vercel forwarding, resource-reading, and app-directed call changes.

This is source/package verification, not a live MCP Apps round trip: no MCP server, AG-UI client, or Vercel AI frontend was connected in this check. The release therefore establishes the shipped API surface, not that every host combination renders or preserves the app payload correctly. Treat the app result as untrusted data at the host boundary, and keep authorization outside the model-visible result.

Why it matters

Agent interoperability is often lost at the last adapter: a tool works in the MCP protocol, but its UI/resource context disappears when the result crosses into an agent frontend. A named bridge for app resources and calls gives integrators a narrower contract to test. The useful acceptance test is not “the import exists”; it is a fixture that exercises an MCP App resource, forwards the result through the selected frontend adapter, and checks that approval and identity boundaries remain host-controlled.

Bottom line: 2.57.0 makes MCP Apps a first-class Pydantic AI integration surface. It improves the contract between protocol and frontend, but it does not by itself prove safe rendering, authorization, or cross-host compatibility.