Pi.dev: You Said No MCP

Pi originally rejected MCP but now includes it in core because MCP evolved and the platform changes required to support it turned out to be broadly useful. Earendil engineers found MCP’s needs overlap with Pi’s: both benefit from a sandboxed interpreter, richer tool metadata, deferred tool loading, mid-conversation system messages and adjustable reasoning levels. Many existing MCP servers still favor token-efficient text dumps and poor composability, so Pi treats modern MCP as needing OpenAPI-like structured outputs and intelligent tool discovery. Rather than keep MCP as an optional extension, integrating it into core enabled improvements that also make model-driven tools like Jev easier to use and helps influence better harness patterns.

Codemode is a JavaScript sandbox that runs on the trusted harness side to orchestrate and combine tool calls, keeping state in the session transcript instead of the file system; small JS/WASM runtimes provide reasonable protection. In Pi, codemode auto-loads with MCP and lets agents stitch together services - example: combining Linear MCP and the Jev classifier to analyze 167 open issues, with Jev labeling 156 neutral and 11 mildly frustrated and flagging specific issues (e.g., PI-6907, PI-10031, PI-4714, PI-7730). Integrating MCP into Pi’s core was a deliberate choice to improve composability for small harnesses and help shape better server and tooling patterns.