Pi has added Model Context Protocol support to its core, reversing its earlier public resistance to MCP and tying the integration to a broader redesign of how the agent harness loads and coordinates tools. The developers said the change reflects both improvements in MCP and capabilities Pi needed independently.
MCP had previously been available to Pi through an extension, but the new implementation makes it a supported part of the main product. The team argued that an extension alone did not expose enough information to manage newer model capabilities, including deferred tool loading and changes to reasoning settings during a conversation. Bringing the work into the core allowed Pi to distinguish between tools directly available to a model and tools intended to be used through its orchestration layer.
That layer is called Codemode. Pi describes it as a JavaScript sandbox in which an agent can discover, order and combine tool calls. Instead of presenting every available operation to the model at once, the system can expose tools through a programmable environment and preserve the orchestration state in the session transcript. Pi automatically loads Codemode when MCP is configured, and users can also enable it separately as a default tool.
The design is meant to address one of the team's continuing criticisms of MCP: composition across different tools can remain awkward. Pi's developers said many MCP servers are designed for systems that place large tool lists directly into model context and return primarily textual results. Their preferred direction is closer to an API ecosystem with structured output and intelligent discovery based on tool descriptions and documentation.
In that model, MCP supplies a standard way to expose capabilities, while Codemode supplies the execution space for joining those capabilities together. Pi says JavaScript is useful because compact runtimes can be delivered as WebAssembly and constrained with sandboxing. The team also points to the familiar composability of command-line utilities as a benchmark for what structured MCP tools should eventually provide.
The developers did not claim that the integration resolves every limitation. They said server conventions and differing harness approaches still impede reliable composition. The decision to support MCP is therefore also an attempt to participate in shaping those conventions for smaller agent systems, rather than treating the protocol as a finished solution.
For users, the immediate change is straightforward: an upgraded Pi can now connect to MCP tools as a supported feature. The larger significance is the accompanying tool architecture. Pi is betting that discovery, structured data and sandboxed orchestration will let agents use broad tool collections without filling their working context with every tool definition or every intermediate result.
