Pi 1.0 ships with native MCP and the Pi Durable harness

On October 16, 2026, the Pi project moved to version 1.0, with Code Mode and native MCP reaching stable release, and the team introduced a new Harness for long-running tasks called Pi Durable on the same day. According to InfoQ Chinese, these changes had been in the works for some time: over the past year the MCP protocol went through six or seven rounds of adjustments before reaching a genuinely usable stage, while the Pi Durable architecture went through eight iterations. The report notes that a protocol Pi had once publicly rejected has now entered the core, alongside a long-running Agent framework that was refined repeatedly, marking Pi's expansion from a terminal coding Agent toward full Agent infrastructure.

For developers, the most immediate value of Pi Durable is the amount of repetitive code it removes. Previously, building a personal assistant, a Slack Bot, or another Agent application on Pi required developers to design their own database for storing messages and to handle Agent startup, task scheduling, process crashes, and state recovery. One developer estimated that Pi Durable could replace about half of their surrounding code, since these details are now handled by the framework. It stores user messages, model replies, tool execution results, and task state; after a process crash, the Agent can continue from where it left off. Operations that can be safely replayed are re-executed, while operations that might produce duplicate side effects are not rerun carelessly.

Mario Zechner summarized it simply: an Agent starts, calls tools, and the process suddenly dies; after restarting, it should continue exactly from where it stopped, without a person typing "continue" again, and tool results and streaming output must not be lost. Armin Ronacher focused on the reliability behind it, arguing that models are already unpredictable enough, and when Agents begin to operate without real-time human oversight, the outer task orchestration and failure recovery should at least be stable and understandable. The report also notes that Pi, which had publicly criticized MCP as bloated, now loads tools on demand, while Code Mode lets the Agent write JavaScript in a QuickJS sandbox to orchestrate tool calls and reduce Token consumption.