Earendil ships Pi 1.0, a minimal open-source agent harness, plus experimental Pi Durable

Pi 1.0 arrives after months of community hardening

Earendil has released version 1.0 of Pi, its open-source harness for AI coding agents. According to the company's announcement, which reached the Hacker News front page, hundreds of thousands of people use Pi every week, and the 1.0 milestone is largely the product of issues and pull requests submitted by that community over many months. Earendil presents the release as a stable base that individuals and businesses can rely on.

Before this release, Pi already ran the newest models from every major provider and served as an everyday coding agent for many of its users, the company says. Version 1.0 builds on that foundation with a set of features the team says survived a long internal vetting process.

What is new in 1.0

The release adds the following:

• Codemode, which brings native support for MCP, the Model Context Protocol, as well as support for non-LLM models such as Jev and image models
• Extension support for virtual models, which let builders compose multiple models into a single routed unit
• Deferred tool loading, so tools do not all have to be available upfront
• Cache warming for Anthropic models, aimed at Anthropic's prompt caching
• Mid-conversation system messages, enabling transcript-aware changes to prompts and tools during a session
• A new theme for the terminal interface
• Full-screen mode as the default

The announcement is accompanied by demonstrations of the new capabilities. In one, Pi uses Codemode to write a script that condenses a week of git commits into a short summary. In another, Pi writes an extension for itself: a virtual model that plans with Claude Opus, implements with GPT, and relies on Jev to decide when to hand off between the two. A session command in the demo breaks down cost per model alongside cache usage.

Minimalism as a deliberate constraint

A recurring theme in the post is restraint. The Earendil team says agentic tooling shifts every week, but that many of those changes do not stick. Rather than chasing each new development, Pi only adopts functionality once it has proven itself, with the team weighing what a feature genuinely offers against the complexity it introduces. According to the post, the list of candidate features that did not make the cut is far longer than the one that shipped.

Pi Durable takes the idea beyond the terminal

Alongside Pi 1.0, Earendil is shipping an experimental package called Pi Durable. The company says that during Pi's evolution it became clear some users wanted a shape Pi did not have: something reachable from surfaces beyond the terminal, outside the coding-agent use case, and able to sustain longer-running conversations and tasks.

Rather than stretching Pi itself, that work was split into Pi Durable, described as a substrate for building long-running agentic applications. It carries over Pi's core principles of minimalism and what the team calls "supermalleability" — the ability to reshape the underlying intelligence — but extends them to new kinds of applications. A separate post by a team member named Mario covers its design in more depth.

Availability

Both projects are released under the MIT license. Pi 1.0 installs through a shell script served from pi.dev, with a PowerShell equivalent for Windows. Pi Durable is distributed via npm as three packages: @earendil-works/pi-durable, @earendil-works/pi-ai and @earendil-works/chord. Documentation lives at pi.dev and the source code is published on GitHub under the earendil-works organisation. Earendil frames its broader mission as building software and open protocols that reinforce people's control over AI tools.

Why it matters

Agent tooling currently churns at a pace that makes framework choices feel risky: what is standard this month may be abandoned the next. Pi's counter-position — adopt only what has proven durable, keep the core lean — offers developers a hedge against that churn, and the 1.0 label signals the project now considers itself dependable enough to bet on. The reported scale of weekly users suggests this is not a niche experiment but a widely used piece of infrastructure. Meanwhile, Pi Durable indicates where Earendil believes the category is heading: away from single terminal sessions and toward long-running agents embedded in many surfaces. With permissive MIT licensing and open code, both lower the barrier for teams that want to build their own agent stack rather than rent one.