Earendil, the team behind the Pi AI assistant, had been a vocal critic of the Model Context Protocol (MCP). Their website at pi.dev once featured a proud declaration that Pi does not support MCP, and team members, including Mario, had repeatedly dismissed MCP in podcasts and posts. This stance reflected a broader skepticism in some corners of the AI assistant ecosystem about MCP’s utility and design.
In a September 29, 2026 blog post titled 'You Said No MCP!', Earendil announced a reversal: MCP is now a supported piece of functionality in Pi’s core. The team explained that their change of heart came after paying close attention to MCP’s evolution over the past year. They realized that while the world is not static and MCP itself has improved, the more compelling reason for integration was the side benefits of the technical work required. Specifically, the changes made to accommodate MCP — such as enhancing the sandbox for tool calls and improving tool discovery through documentation and description — also made it significantly easier to integrate other tools like Jev, a classifier for analyzing text tone.
This image illustrates the blog post where Earendil announced their support for Model Context Protocol after previous opposition, as discussed in the article.
Earendil’s renewed engagement with MCP is not merely about adopting a protocol but about shaping its future for small-scale AI harnesses. They advocate for MCP to evolve closer to OpenAPI, where tools return structured data and are discoverable via metadata, rather than relying on text-only responses optimized for token efficiency. To illustrate the practical use of MCP in Pi, the post highlights Codemode, a feature that runs tool orchestration on the harness side (in a trusted environment) using JavaScript sandboxing. An example demonstrates using Codemode with the typesafe/jev model to analyze frustration in open Linear issues: fetching issues, retrieving comments, classifying emotional tone, and storing results — all without wasting context.