PIDE MCP: Connecting AI Agents to Isabelle

Isabelle/PIDE is Isabelle's asynchronous model of document-oriented proof processing. Its frontends include the Isabelle/jEdit and Isabelle/VSCode editors. PIDE MCP is a new frontend for AI agents: a customisable Model Context Protocol server managing a headless PIDE session. We describe PIDE MCP's architecture, propose desiderata for programs connecting AI agents to proof assistants, and evaluate PIDE MCP and alternative approaches against them.