Earendil Brings MCP to Pi Core After Embracing Modern Standards

Overview

Earendil Engineering has confirmed that the Model Context Protocol (MCP) is now a supported, core functionality within its Pi platform. This marks a significant pivot from the company’s previous public stance, which explicitly rejected MCP integration in favor of proprietary extension architectures. The decision to embrace the standard reflects a broader industry shift toward interoperable tooling for large language models and signals Earendil’s commitment to adapting its technology to meet evolving developer needs.

For years, Pi users visiting the company’s developer portal encountered a clear declaration that the platform did not support MCP. This position was reinforced in various podcast appearances and technical blog posts, including a notable article by Mario. However, with the latest software updates, developers can now access MCP directly within the Pi core environment. This reversal has prompted questions about the rationale behind the change and how it impacts the existing ecosystem.

Rethinking the Architecture

The primary driver behind this strategic shift is the rapid evolution of the MCP specification itself. According to Earendil Engineering, the version of MCP available today is fundamentally different from the one evaluated a year ago. While the protocol could have remained a third-party extension, the team determined that integrating it into the core offered broader utility than maintaining it on the periphery.

Initially, MCP was indeed offered as an endorsed extension. However, internal reviews revealed that the changes required to make MCP function effectively within Pi were generally beneficial for the platform as a whole. The integration process has unlocked easier access to Jev, a typesafe model tooling system, within the Pi environment. This synergy highlights a shared architectural need: both Pi and modern MCP implementations require a robust sandbox interpreter to manage complex agent interactions safely.

The reason we brought MCP into the core is not just about how MCP has changed, but also because we found that the changes it would require were generally useful.

Technical Improvements and Codemode

A critical component of this update is the introduction of Codemode, a mechanism designed to address longstanding issues with tool composition in MCP environments. Historically, MCP has struggled with composability, particularly when relying on traditional harnesses that dump tools into context and optimize for token efficiency through text returns. Earendil argues that modern MCP should align more closely with OpenAPI standards, emphasizing intelligent tool discovery and structured data returns over raw text.

Codemode operates by running tool orchestration within the trusted harness environment rather than in the untrusted sandbox where most agent tools execute. This distinction allows agents to coordinate tool calls with greater flexibility, using JavaScript to combine operations efficiently. Because Codemode runs on the harness side, its state is maintained as part of the session transcript rather than the file system, ensuring consistency and security.

Feature	Traditional MCP	Pi with Codemode
Execution Environment	Untrusted sandbox	Trusted harness loop
Tool Composition	Limited, text-heavy	JavaScript-based, structured
State Management	File system dependent	Session transcript based
Tool Discovery	Context dumping	Intelligent, OpenAPI-like

This architecture allows Pi to support new model capabilities, such as deferred tool loading and mid-conversation system messages, which were previously difficult to manage with older tool loadouts. By configuring tools to be deferred or specific to Codemode, developers can better scale their applications across different LLM providers.

Shaping the Ecosystem

Earendil’s decision to integrate MCP stems from a philosophy of active participation rather than passive observation. The company acknowledges that while the protocol has improved, many existing MCP servers still rely on outdated patterns that are suboptimal for small, focused harnesses. By embedding MCP into the core, Earendil aims to influence the direction of the standard, encouraging server developers to adopt structured data practices and better metadata handling.

The integration is automatic for users who configure MCP, and Codemode can also be added as a default tool. This setup enables powerful use cases, such as analyzing issue tracker sentiment by combining Linear MCP data with Jev’s classification capabilities. For example, users can now ask Pi to identify frustrated commenters on open issues, with the system seamlessly orchestrating data retrieval and analysis without wasting context window space.

We believe the best way to positively influence something is to embrace it. And while we think that modern MCP is in a much better spot than MCP ever was, the servers and patterns still leave room for improvement.

This move positions Pi at the forefront of the evolving LLM tooling landscape. As the industry continues to grapple with how best to integrate external tools and data sources, Earendil’s embrace of modernized MCP and Codemode offers a blueprint for other platforms looking to balance security, performance, and interoperability.