Pi 1.0
🚀 Announcing Pi 1.0
Date: Thursday, October 1, 2026
From: Earendil (rfc@earendil.com)
Subject: The Launch of Pi 1.0
We are thrilled to officially release Pi 1.0: a minimal, hardened, and highly extensible agent harness designed for total user customization.
With a weekly user base numbering in the hundreds of thousands, we have spent several months refining Pi. By integrating real-world feedback, we have evolved it from a prototype into a stable software foundation that both individual developers and enterprises can rely upon.
🛠 Our Development Philosophy
In an industry where agentic tools shift every few days, we choose a path of intentionality. We chase every trend We prioritize stability.
"We wait until a technology has proven its worth before considering its adoption, carefully balancing actual utility against the burden of added complexity."
This disciplined approach—which we recently discussed in relation to MCP and Codemode—is exactly what led to the version of Pi we are shipping today.
🌟 What's New in Pi 1.0?
Pi has already established itself as a flexible substrate for agentic apps and a primary coding agent supporting all major model providers. Version 1.0 expands this capability significantly.
Feature Checklist
- Codemode: Native integration for MCP, image models, and non-LLM models (e.g.,
Jev). - Virtual Models: Full extension support.
- Optimized Loading: Implementation of deferred tool loading.
- Performance: Cache warming specifically for Anthropic models.
- Dynamic Context: Mid-conversation system messages (allowing for tool and prompt adjustments based on the transcript).
- UI Overhaul: A fresh TUI theme and full-screen mode enabled by default.
Technical Summary
| Feature | Benefit | Scope |
|---|---|---|
| Codemode | Multi-model orchestration | Native |
| Virtual Models | Custom agent logic | Extensible |
| Cache Warming | Reduced latency | Anthropic |
| TUI Theme | Improved aesthetics | Interface |
🌐 Introducing: Pi Durable
While refining Pi, we realized that some user needs didn't fit the "terminal-centric" or "coding-agent" mold. Users wanted AI that was reachable across various surfaces and capable of managing long-term tasks.
To avoid bloating the core Pi experience, we created an experimental companion: Pi Durable.
Pi Durable is a substrate for building long-running agentic applications. It allows creators and users to steer underlying intelligence with extreme precision while maintaining the core tenets of minimalism and supermalleability.
Architectural Relationship
📦 Installation & Licensing
Both Pi 1.0 and Pi Durable are released under the MIT License.
Install Pi 1.0
Linux/macOS:
curl -fsSL https://pi.dev/install.sh | sh
Windows (PowerShell):
powershell -c "irm https://pi.dev/install.ps1 | iex"
Install Pi Durable (Experimental)
npm install @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chord
🧪 Codemode in Action: A Use Case
To demonstrate the power of the new system, consider this workflow where Pi automates a complex commit summary:
- The Task: Pi writes a script to condense a week's worth of commits into a concise summary.
- The Extension: Pi creates a virtual model router:
- Planning: Handled by
Claude Opus. - Implementation: Handled by
GPT. - Routing:
Jevmonitors the flow and decides when to switch models.
- Planning: Handled by
- Execution:
- User triggers
/sessionPlanning starts on Opus. Jevdetects the shift to coding Hands over to GPT.- The script is executed via
6 Luna !.
- User triggers
- Analysis: The session concludes with a detailed breakdown of resource usage.
Cost Calculation Logic: The system calculates the total expenditure using a formula similar to:
Our Mission: Earendil exists to build open protocols and software that amplify human agency. We look forward to your feedback as we build the future of AI in the open.