On the evening of October 1, 2026, the Vienna-based software company Earendil released Pi 1.0, a specialized coding agent designed for terminal environments. Pi acts as a bridge between artificial intelligence models and local file systems, allowing the AI to read files, execute commands, and suggest code corrections directly within a developer's workspace. The announcement quickly ascended to the front page of Hacker News, amassing 1,684 points and 602 comments by October 5.

The version number signifies a maturation of the project. Earendil reports that Pi is now used by hundreds of thousands of developers each week, with the 1.0 release incorporating months of stability improvements and feedback to serve as a production-ready foundation for both individual engineers and enterprises. Most notably, the release marks the company's decision to embrace the Model Context Protocol (MCP), a standard the project had explicitly avoided since its inception.

In this report:
  • How a developer frustrated with existing tools built an agent that reached 112,000 stars.
  • Why Pi reversed its stance on the MCP standard to include it in version 1.0.
  • The story behind the licensing agreements during the company's acquisition.
  • The capabilities and experimental nature of the new Pi Durable package.

How did a developer-led protest evolve into a software company?

Pi's creator, Mario Zechner, is a Graz-based software engineer widely recognized for authoring libGDX, a prominent game development framework for Android. In early 2025, Zechner began using Claude Code but grew disillusioned with the tool within months. In a manifesto published on November 30, 2025, he cited two primary grievances: the agent had become bloated with features that the vast majority of users never utilized, and the constant flux in system commands led to inconsistent model behavior across versions.

Zechner founded Pi on a philosophy of radical minimalism. The agent was restricted to four fundamental tools: reading files, writing files, performing line-by-line text modifications, and executing shell commands. By omitting complex features like auxiliary sub-agents or task lists, the project kept its system instructions under 1,000 tokens. Zechner also explained his initial rejection of the Model Context Protocol, noting that common MCP servers burdened the model with excessive tool definitions, consuming up to 9% of the context window before a single task was initiated.

What is the engine powering the Pi agent?

Pi gained widespread adoption after Peter Steinberger utilized it as the foundation for his personal AI assistant, OpenClaw. Following positive exposure from industry figures, Zechner entered a period of intense negotiation, eventually selling the project to Earendil—a firm founded by Colin Daymond Hanna and Armin Ronacher—on April 8, 2026. Zechner became a partner in the company, taking care to structure the project's licensing to prevent the repeat of historical industry failures, such as the total closure of the open-source RoboVM project after its acquisition by Microsoft. The current plan includes a perpetual MIT license for the core, Fair Source licensing for commercial add-ons, and proprietary cloud infrastructure.

Why did Pi eventually integrate the Model Context Protocol?

Version 1.0 introduces deferred tool loading and mid-session system updates, allowing the model to adapt its capabilities dynamically without overwhelming the context window. The integration of MCP was facilitated by "Codemode," a new feature where the model writes and executes JavaScript scripts inside an isolated WebAssembly interpreter rather than calling tools individually. By keeping these operations within the isolated environment, Pi avoids the overhead of loading massive MCP definitions at the start of every session. Earendil suggests that while MCP implementations are still maturing, this architecture allows for more efficient data processing and classification tasks.

What is the purpose of the experimental Pi Durable package?

Accompanying the 1.0 release is Pi Durable, an infrastructure package for long-running, multi-user agents. Pi Durable records checkpoints at every operational step, allowing a crashed process to resume exactly where it left off by accessing the existing storage state. While the initial GitHub repository has reached over 112,000 stars and millions of downloads, the project remains in an iterative phase. Earendil is currently balancing rapid development with the promise of long-term stability as it transitions from a solo developer project to an enterprise-grade toolkit.

Notes and Sources

  1. Earendil, "Pi 1.0," earendil.com, October 1, 2026.
  2. Hacker News, "Pi 1.0" discussion, post no. 49926069, October 1, 2026.
  3. Mario Zechner, "What I learned building an opinionated and minimal coding agent," mariozechner.at, November 30, 2025.
  4. Mario Zechner, "I've sold out," mariozechner.at, April 8, 2026.
  5. Earendil, "Announcing Pi & Lefos," earendil.com, April 8, 2026.
  6. Earendil Engineering, "You Said No MCP!," earendil.com, September 29, 2026.
  7. Earendil Engineering, "Pi Durable," earendil.com, October 1, 2026.
  8. GitHub, earendil-works/pi repository data and release logs; npm, @earendil-works/pi-coding-agent weekly download statistics, September 28-October 4, 2026.