Pi.dev: You Said No MCP

If you accessed pi.dev previously, you'd see a firm statement that Pi does not support MCP. Throughout various podcasts discussing Pi, dismissive remarks about MCP were made, including one by Mario. However, upgrading to Pi now reveals that MCP is indeed a supported feature. This change occurred because the concept of MCP has evolved and Pi decided to integrate it into the core system.

This decision wasn't solely based on MCP's advancements; it was also influenced by the broader improvements MCP's changes would enable, such as making Jev more accessible within Pi. While numerous improvements have been made to MCP, some aspects still remain problematic, primarily its difficulty in composition. Despite these challenges, it's now more akin to OpenAPI with intelligent tool discovery, meaning tools should return structured data and be discoverable via documentation.

The use of CLIs is facilitated by this approach, as they allow for efficient tool orchestration. However, unlike CLIs, MCP in Pi is built to expose tools to a JavaScript sandbox, similar to other harnesses like Codex. This enables more flexible tool execution and integration within the model's reasoning process. The decision to include Codemode alongside MCP was driven by the need to accommodate new model capabilities like deferred tool loading and mid-conversation system messages.

While it's possible to use Codemode without MCP, the latter provides a more comprehensive solution that better addresses the issues MCP traditionally faced. By embracing MCP and Codemode, Pi aims to contribute to ongoing improvements in their implementation, particularly focusing on enhancing their utility within small harnesses.