Pi reverses course, adds MCP support to core
Original: You Said No MCP
Why This Matters
A notable reversal signals MCP's growing de-facto status as AI tool integration standard.
Earendil's Pi product, which previously and publicly rejected MCP support, has reversed course and integrated MCP into its core. The team cites MCP's evolution, internal architectural changes enabling better tool handling, and a desire to actively shape MCP's direction rather than criticize from the outside.
Earendil's Pi had been openly dismissive of MCP — the company's website declared it unsupported, and co-founder Mario published a post criticizing it. That stance is now officially retired. As of the latest Pi update, MCP is built into the core product.
The team's explanation is two-part. First, MCP has matured. Second, and more importantly, the internal work required to integrate MCP properly turned out to solve other problems too — specifically, it made Jev easier to use within Pi and forced improvements to how Pi handles tool loading, deferred tool access, mid-conversation system messages, and reasoning-level changes.
MCP as an extension was technically possible and existed briefly in that form, but the Earendil team concluded that tight core integration was the right call. Their model: MCP should behave more like OpenAPI with intelligent tool discovery, where tools return structured data and are discoverable via documentation — not just token-optimized text dumps aimed at context windows.
The biggest remaining complaint stands: MCP is still hard to compose. The team notes that many MCP servers are still built for simple harnesses that flood the context with tools, limiting composability. Pi's approach runs MCP tools through a JavaScript sandbox — similar to how Codex handles it — which the team sees as closer to how CLI agents wire things together efficiently.
The closing rationale: influence requires participation. Earendil would rather help shape how MCP works in small harnesses than watch from the sidelines.