Earendil released Pi 1.0, the official version of its open-source AI coding tool, under the MIT license on October 1, 2026 (local time). The release retains a minimal, extensible design while adding built-in support for the Model Context Protocol (MCP) and a way to switch AI models according to the task.

Pi accepts natural-language instructions in a terminal and can read and edit files and run commands. It can also be embedded in other applications as an agent harness, which manages model requests, tool execution, and conversation history. Subagents and planning-only mode are not built in; users can add functions through extensions and packages. Extensions, skills, prompt templates, and display themes can be shared through npm or Git.

Pi supports model providers including Anthropic, OpenAI, and Google, and users can switch models during a conversation or configure it to use local models. The release also adds Codemode, which combines multiple tool calls in JavaScript. Codemode can run tools sequentially or in parallel, filter large amounts of retrieved data in a script, and pass only the needed results to the model.

Earendil had previously chosen not to include MCP by default. It changed course after MCP advanced and the company concluded that combining tools and discovering or loading them only when needed could be useful beyond MCP. Loading tool descriptions on demand, instead of sending them all to the model at the start, limits the amount of information handled. MCP, Codemode, and virtual-model support arrived in the preview release 0.99.0 on September 29. The October 1 full release also reduced the tokens used to describe Codemode, added calls to image-generation models, and standardized full-screen display.

Pi 1.0's virtual-model extension lets users keep one model selected while routing each request to a different underlying model. Routing conditions can include task type, cost, and conversation state. In an official demo, Claude Opus handled planning, GPT handled implementation, and the classifier Jev decided when to switch models. Users can design model assignments for different stages of work instead of relying only on manual switching.

Earendil also released Pi Durable, an experimental, separate development package for long-running agents. It records checkpoints—saved states used to resume processing—and can continue unfinished work after a process stops. It reruns an interrupted tool call only if that call was designated safe to repeat. Pi Durable supports concurrent conversations and setups where multiple users instruct the same agent. It serves a different purpose from Pi, which is operated by one person in a terminal, and does not replace Pi. Pi Durable is also MIT-licensed, but its API may change while it remains experimental.

Earendil released Pi 1.0, the official version of its open-source AI coding tool, under the MIT license on October 1, 2026 (local time). The release retains a minimal, extensible design while adding built-in support for the Model Context Protocol (MCP) and a way to switch AI models according to the task.

Pi accepts natural-language instructions in a terminal and can read and edit files and run commands. It can also be embedded in other applications as an agent harness, which manages model requests, tool execution, and conversation history. Subagents and planning-only mode are not built in; users can add functions through extensions and packages. Extensions, skills, prompt templates, and display themes can be shared through npm or Git.

Pi supports model providers including Anthropic, OpenAI, and Google, and users can switch models during a conversation or configure it to use local models. The release also adds Codemode, which combines multiple tool calls in JavaScript. Codemode can run tools sequentially or in parallel, filter large amounts of retrieved data in a script, and pass only the needed results to the model.

Earendil had previously chosen not to include MCP by default. It changed course after MCP advanced and the company concluded that combining tools and discovering or loading them only when needed could be useful beyond MCP. Loading tool descriptions on demand, instead of sending them all to the model at the start, limits the amount of information handled. MCP, Codemode, and virtual-model support arrived in the preview release 0.99.0 on September 29. The October 1 full release also reduced the tokens used to describe Codemode, added calls to image-generation models, and standardized full-screen display.

Pi 1.0's virtual-model extension lets users keep one model selected while routing each request to a different underlying model. Routing conditions can include task type, cost, and conversation state. In an official demo, Claude Opus handled planning, GPT handled implementation, and the classifier Jev decided when to switch models. Users can design model assignments for different stages of work instead of relying only on manual switching.

Earendil also released Pi Durable, an experimental, separate development package for long-running agents. It records checkpoints—saved states used to resume processing—and can continue unfinished work after a process stops. It reruns an interrupted tool call only if that call was designated safe to repeat. Pi Durable supports concurrent conversations and setups where multiple users instruct the same agent. It serves a different purpose from Pi, which is operated by one person in a terminal, and does not replace Pi. Pi Durable is also MIT-licensed, but its API may change while it remains experimental.