Earendil announced Pi 1.0 on October 1, 2026, describing it as a hardened, minimal, and extensible agent harness that hundreds of thousands of people use weekly. The release follows months of refinement driven by community feedback to stabilize the tool for personal and business use. Pi 1.0 adds deferred tool loading to improve start‑up speed, cache warming for Anthropic models to reduce latency, a new text‑based user interface theme, and sets full‑screen mode as the default. It also adds extension support for virtual models, allowing users to plug in additional AI capabilities beyond the built‑in model set.
The image is the header or featured image from the Earendil blog post about the Pi 1.0 release, which is the subject of the article.
The centerpiece of Pi 1.0 is Codemode, which provides native support for the Model Context Protocol (MCP) and makes it possible to run non‑LLM models such as Jev and image models alongside conventional large language models. Codemode lets Pi write a script that summarizes a week of commits into a short plan, then have Jev decide when to switch from planning with Claude Opus to implementation with GPT‑6 Luna, run the finished script, and finally break down cost per model and cache use. In addition, Pi 1.0 introduces mid‑conversation system messages that enable transcript‑aware changes to the prompt and available tools while a conversation is ongoing, so the agent can adapt its behavior as the dialogue progresses.
Earendil also shipped an experimental Pi Durable package, intended as a substrate for long‑running agentic applications. Pi Durable extends the same minimalism and supermalleability of Pi into new dimensions, giving users greater dexterity to wield and steer underlying intelligence over extended workflows. Both Pi 1.0 and Pi Durable are released under the MIT license; Pi 1.0 can be installed with a shell script (`curl -fsSL https://pi.dev/install.sh | sh`) or PowerShell command, while Pi Durable is available via npm as `@earendil-works/pi-durable` together with its peer packages `@earendil-works/pi-ai` and `@earendil-works/chord`. Documentation is hosted at pi.dev and the source code resides at github.com/earendil-works/pi.