Pi 1.0 will support the Model Context Protocol (MCP). Developer Earendil is also adding Codemode to control tools and introducing Pi Durable, an experimental component for AI agents that need to remain active for extended periods.
Pi is a relatively minimalist coding agent that can connect various AI models to tools. With version 1.0, Earendil is expanding that foundation, without intending to turn Pi itself into a comprehensive agent platform. The most notable addition is native support for MCP via Codemode.
MCP is a protocol that allows AI models to gain standardized access to external tools and data sources. Earendil was initially hesitant about the protocol. The company felt the added complexity did not outweigh the benefits. It has since changed its mind.
This support is integrated with Codemode, allowing Pi to write scripts to execute various tool calls. As a result, not all available tools need to be constantly made available directly to the AI model. In addition to MCP, Codemode can also be used for other types of models, including image models and decision models such as Jev.
Pi 1.0 also gains support for virtual models via extensions and deferred tool loading, where tools are loaded only when needed. Cache warming has been added for Anthropic models. Additionally, system messages can be inserted during an ongoing conversation, allowing prompts and available tools to be adjusted based on the existing conversation.
Pi durable for long-running agents
Not all new functionality is integrated into Pi itself. For applications where agents need to remain active for longer periods, Earendil is releasing the separate Pi Durable. That package is still in an experimental phase.
Pi Durable is intended as an underlying layer for agentic applications that go beyond a coding agent in a terminal. While Pi connects the AI model to tools and handles individual conversations, Durable manages multiple conversations and provides persistent storage. According to The Register, SQLite and JSONL are used for this purpose.
This allows the agent layer to be used for applications that remain active for extended periods and are accessed through various interfaces. By keeping this functionality separate, Earendil aims to prevent the original Pi agent from becoming increasingly complex.
Pi Durable does share components and design principles with Pi, but is intended as a foundation upon which developers can build broader agent applications. Pi 1.0 and the experimental Pi Durable are available under the MIT license.