# Pi 1.0: a small agent harness grows Codemode, MCP and virtual models without getting big Earendil's Pi hits 1.0 with Codemode, native MCP, virtual-model routing, deferred tool loading and Anthropic cache warming, plus an experimental Pi Durable package for long-running agents. By The Forward Pass research desk, The Forward Pass, 2026-10-04. Source: https://theforwardpass.net/archive/editorial/pi-1-0-a-small-agent-harness-grows-codemode-mcp-and-virtual-models-without > This article is researched and written by AI models from the sources it cites, and every claim is checked against them. No human edits it before it is published. ## What shipped Pi is an agent harness, not a model. Earendil describes the 1.0 release as "a hardened, minimal, extensible agent harness that you can make your own." According to the release summary, hundreds of thousands of people use Pi weekly, and 1.0 follows months of feedback and hardening. The company says Pi "already ran the latest models from every major provider," serves as a daily-driver coding agent and acts as "a supermalleable substrate on which to build agentic applications." The 1.0 additions, per the announcement: - Codemode - Native MCP support - Support for non-LLM models, such as Jev and image models - Extension support for virtual models - Deferred tool loading - Cache warming for Anthropic models - Mid-conversation system messages (transcript-aware prompt and tool changes) - Full-screen mode with a new TUI theme, now the default Alongside it ships Pi Durable, a separate package for long-running agentic applications that run beyond the terminal. Both are MIT licensed. Docs live at pi.dev and code at github.com/earendil-works/pi. ## How the pieces fit **Codemode.** Pi can write and run scripts for a task. The summary's example is summarizing your commit history. **Virtual models.** Extensions can expose a "virtual model" that routes work across real ones. One of Earendil's examples reads: "Pi writes itself an extension: a virtual model that plans with Claude Opus and implements with GPT, with Jev deciding when to switch." Jev is listed in the announcement as a non-LLM model, and here it acts as the switching policy. **Native MCP.** MCP support now lands in the core. **Deferred tool loading.** The announcement lists the feature without further detail. **Anthropic cache warming.** Pi adds cache warming for Anthropic models. **Mid-conversation system messages.** System prompts and tool sets can change partway through a session, with awareness of the existing transcript. ## The design bet Earendil's pitch is that Pi stays minimal while adopting features that proved useful in practice. In the company's words: "When we use Pi with all these new features it feels like a major step forward, but it also still feels simple, like Pi." Extension points such as virtual models are where users can put their own routing logic, as the plan-with-Opus, implement-with-GPT example shows. ## Pi Durable Pi Durable targets agentic applications that run for a long time and outside an interactive terminal session. It installs from npm alongside `@earendil-works/pi-ai` and `@earendil-works/chord`. The main caveat: Earendil labels Pi Durable experimental. The company describes Pi 1.0 itself as hardened. ## Limits and open questions - Cache warming and deferred tool loading are listed as features, without measured improvements in the provided material. - The source does not detail how Jev decides when to switch models in the virtual-model example. - Pi Durable is experimental and its stability guarantees are unstated. - Usage figures (hundreds of thousands weekly) come from Earendil. ## Who should care - Engineers already running Pi as a coding agent: 1.0 changes the default UI to full-screen and adds Codemode and MCP in core. - Teams mixing providers: virtual models give a place to put plan-versus-implement routing. - Heavy Anthropic users: cache warming is built in. - Builders of background or long-running agents: Pi Durable is worth a test, with the experimental label in mind. The MIT license on both packages keeps forking and embedding straightforward. ## Try it Install Pi 1.0 locally, then optionally add the experimental Pi Durable packages. Windows users can use the PowerShell installer listed at pi.dev. Docs are at pi.dev. **1. Install Pi 1.0 (macOS/Linux)** · Runs locally, no account needed: Installs locally. Using hosted models such as Claude or GPT will require that provider's API key ```bash curl -fsSL https://pi.dev/install.sh | sh ``` **2. Install experimental Pi Durable packages** · Runs locally, no account needed: Requires Node.js and npm. Pi Durable is experimental ```bash npm install @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chord ``` Commands are copied from the official sources below. We have not run them. ## Sources -