Earendil ships Pi 1.0 with native MCP support
Earendil has released Pi 1.0, a hardened agent harness that now includes native Model Context Protocol (MCP) support and broader model compatibility. The company said hundreds of thousands of users run Pi weekly, and months of feedback from GitHub issues and pull requests shaped the 1.0 release. For development teams building AI agents, the update signals a shift from experimental tooling toward a stable, community-driven foundation.
Pi 1.0 adds Codemode, which extends the harness beyond large language models to non-LLM systems such as Jev and image models. The release also introduces extensions for virtual models, giving developers more flexibility in how they compose agent workflows. Earendil positions the release as the result of sustained hardening rather than a feature sprint.
What's new in Pi 1.0
The headline addition is native MCP support. MCP, or Model Context Protocol, has become a common standard for connecting AI agents to external tools and data sources. Shipping it by default removes a setup step that previously required manual configuration. Codemode extends the same interface to non-LLM models, which matters for teams running hybrid pipelines that mix language models with computer vision or other model types.
Pi 1.0 remains minimal and extensible by design. The team has kept the core small while layering in compatibility improvements. The project is MIT-licensed, with installation via curl on Linux and macOS or npm on Windows. Documentation lives at pi.dev, and source code is available on GitHub.
Pi Durable targets long-running agentic apps
Alongside the 1.0 release, Earendil introduced Pi Durable, an experimental package for longer-running agentic applications. It keeps Pi's minimalist approach while adding durability features that let builders steer underlying intelligence with fine-grained control. The package is aimed at multi-step workflows that need to persist state or run across extended sessions.
Both Pi 1.0 and Pi Durable ship under the MIT license. The durable substrate is explicitly experimental, a contrast with the stable 1.0 line. Teams evaluating the two should treat Pi 1.0 as the production option and Pi Durable as a preview of where the platform is heading for complex orchestration.
Why this matters for developers and product teams
If your team is already building on MCP or evaluating agent frameworks, Pi 1.0 removes a meaningful integration hurdle. Native protocol support means less glue code between your agent and the tools it calls. The MIT license also removes procurement friction for commercial use. For product development leads, the Pi Durable package is worth tracking: long-running agent workflows are a hard problem, and Earendil's approach to durability could influence how you architect multi-step automations. Teams looking to build internal MCP skills can start with MCP Courses or broader AI Agent Courses to bring developers up to speed before committing to a framework.
Your membership also unlocks:
