Earendil has announced the general availability of Pi 1.0, a minimal coding agent harness designed for developers building and working with AI agents.
The company also introduced Pi Durable, an experimental package designed for long-running, fault-tolerant AI agents that can survive crashes, maintain conversation continuity and support collaboration between multiple users.
According to Earendil, Pi has already attracted hundreds of thousands of weekly users worldwide, with the latest releases reflecting months of development driven by community feedback and issue tracking.
Pi 1.0 Adds Native MCP Support
The stable Pi 1.0 release introduces several features aimed at making the agent more suitable for production use.
Among the key additions is native Model Context Protocol (MCP) support, allowing Pi to connect with external tools and services through the increasingly adopted protocol for AI agents.
The release also includes:
- Cache warming for Anthropic models
- Mid-conversation system message updates that retain awareness of the existing transcript
- A redesigned terminal user interface theme
- Full-screen mode enabled by default
The features are consistent with Earendil’s minimalist approach to software development. Rather than adopting every emerging AI technology immediately, the company says it prefers to integrate technologies once they have demonstrated their usefulness.
Pi Durable Targets Long-Running Agents
Alongside Pi 1.0, Earendil launched Pi Durable as an experimental npm package.
The package is designed for scenarios where AI agents need to operate continuously without losing their state when something goes wrong.
Unlike the standard Pi experience, which runs locally within a user's terminal session, Pi Durable is designed to withstand events such as process termination, container redeployment and memory exhaustion while preserving the agent’s conversation.
It also allows multiple people to steer the same AI agent simultaneously, potentially making it useful for collaborative workflows and production environments.
Another feature is support for infinitely long conversation histories, allowing agents to maintain context across extended interactions.
From Coding Assistant to Agent Infrastructure
The launch represents a potential expansion of Earendil’s focus beyond individual developers using AI coding assistants.
While Pi 1.0 provides a minimal and production-ready agent harness, Pi Durable is aimed at developers and organisations building persistent, collaborative AI agent applications.
The distinction could become increasingly important as businesses move from short AI interactions towards agents expected to perform tasks over hours, days or longer periods.
An agent that loses its state after a crash or deployment could require work to be restarted, while a durable agent can potentially resume from where it stopped.
Open-Source Approach
Both Pi and Pi Durable remain open source under the MIT licence, with their code hosted on GitHub and documentation available through Pi’s website.
The projects reflect Earendil’s broader focus on developing software and protocols intended to give developers greater control over how AI agents operate.
With Pi 1.0 now generally available and Pi Durable entering experimental use, Earendil is positioning its technology for a broader shift from single-user coding assistance towards persistent and collaborative AI agent infrastructure.
Source: Alternativeto