The Pi coding agent reached version 1.0 on Thursday carrying a feature its own creator had publicly written off. Earendil's release post names native support for the Model Context Protocol as one of two headline additions β a reversal for a project whose author, Mario Zechner, had argued the standard was unnecessary.
Key takeaways
- Pi 1.0 adds Codemode, a harness-side sandbox that gives the agent native MCP support alongside decision models and image models.
- Earendil says hundreds of thousands of people use Pi every week, and ships both Pi and the new Pi Durable package under the MIT license.
- Pi Durable is roughly 15,000 lines of source excluding tests, which Earendil measures at about 150,000 tokens for GPT models and 250,000 for Claude.
Why Earendil changed its mind on MCP
Earendil was formed by Armin Ronacher and Colin Daymond Hanna and acquired Pi in April 2026. The company told The Register that things change: MCP itself has improved, and modifications Pi's developers made to it have made integrating other capabilities easier. It pointed specifically at Jev, a decision model, saying the MCP work made Jev easier to use inside Pi.
The company framed the underlying requirement as a shared one, arguing that what Pi needs closely resembles what MCP needs β a sandbox to play in, in the form of an interpreter. Earendil's stated policy is to wait until a piece of agentic tooling has proven itself, then weigh its real functionality against the complexity it adds. The release post describes most of the candidate features as having fallen off the wall rather than stuck.
What Codemode actually does
Codemode is the harness-side sandbox that carries the new support. Instead of threading protocol clients through the agent loop, it gives the model a place to issue tool calls and run scripts, covering MCP servers, non-LLM decision models such as Jev, and image models. Earendil's demonstrations show Pi writing a script that compresses a week of commits into a short summary, and Pi building itself an extension that plans with Claude Opus and implements with GPT while Jev decides when to hand over.
The release also adds mid-conversation system messages, which let prompts and tool sets change partway through a transcript. The Register lists further changes in the same build: extension support for virtual models, deferred tool loading, cache warming for Anthropic models, a new TUI theme, and full-screen mode enabled by default.
Pi Durable targets long-running agents
The second release is Pi Durable, an experimental package for agentic applications that outlive a terminal session. Earendil kept it deliberately separate so Pi itself stays minimal. Where Pi connects one model to tools for one person at a keyboard, Pi Durable orchestrates multiple conversations and owns the storage layer beneath them.
It ships memory, SQLite and JSONL backends plus a conformance suite and benchmarks for custom ones. Earendil notes that the SQLite and JSONL code uses no Node APIs, so a small adapter runs it on Bun or inside a Cloudflare Durable Object. Only the working set stays in memory β active transcripts, live tasks and pending submissions β while compaction summarizes older messages before they overflow the model's context window.
Code size is treated as a feature rather than a footnote. Earendil puts Pi Durable at about 15,000 lines without tests, roughly 150,000 tokens for GPT models and 250,000 for Claude, and notes the storage backends account for 3,000 lines an agent can usually skip. The stated goal is that an agent can read the harness it runs on.
How it lands against bigger harnesses
The release arrives while far larger vendors push harness layers of their own; our report on AWS open-sourcing its Strands harness covered one of them. Pi's pitch runs the other direction, and The Register's analogy was Flask rather than Django. The announcement drew more than 290 points on Hacker News on the day it posted.
Both packages carry MIT licenses. Pi installs through a single shell command on macOS and Linux and a PowerShell equivalent on Windows, with documentation at pi.dev, while Pi Durable is published on npm as three experimental packages.
FAQ
Is Pi open source?
Yes. Earendil ships both Pi and the experimental Pi Durable under the MIT license, with documentation hosted at pi.dev. Pi installs with one shell command on macOS and Linux or a PowerShell command on Windows.
Does Pi Durable replace the Pi coding agent?
No. Earendil states that Pi Durable does not replace Pi and is instead a framework for building any agentic application, coding agents included. It shares code such as pi-ai with Pi, and Earendil says lessons from it will flow back into the coding agent as they prove themselves.
What is Jev in Pi 1.0?
Jev is a decision model that Pi can call as a non-LLM component. In Earendil's demonstrations it chooses when to switch between a planning model and an implementation model, and the company says its MCP changes made Jev easier to wire into Pi.
