Pi 1.0 is a minimal, MIT-licensed coding-agent harness for developers who want an open terminal agent they can reshape. Released by Earendil Works on October 1, 2026, it is used by hundreds of thousands of people each week.
Core Features
- Codemode: the model writes JavaScript in a QuickJS sandbox and calls tools in parallel, cutting prompt tokens about 40 percent (roughly 5,300 to 3,300 on GPT-5.6).
- Native MCP support, so agents connect to existing Model Context Protocol servers without a wrapper layer.
- Virtual models: a Jev classifier routes requests between providers, planning on Claude Opus and implementing on GPT-6 Luna.
- Deferred tool loading, Anthropic prompt-cache warming, and mid-conversation system messages.
- Full-screen TUI by default, about 111,000 GitHub stars, and installable from pi.dev or npm.
Use Cases
- Builders who want a provider-agnostic agent core instead of a locked-in commercial product.
- Teams writing extensions or custom tool orchestration directly in code.
- Long-running agent apps via the separate experimental Pi Durable package, which adds SQLite and JSONL checkpoints so a crashed process resumes unfinished work.
Pricing
Free and open-source under the MIT license. There is no subscription; you pay only the underlying model provider such as Anthropic, OpenAI, or Google. Pi Durable is published to npm as an experimental package.
Pros and Cons
Pros: open, extensible, vendor-neutral, with real token savings from Codemode.
Cons: minimalism pushes security responsibility onto the user, and Pi Durable remains experimental.
Our Take
Best for engineers who treat the agent harness as infrastructure they own. The trade-off is that Pi hands you a sharp, small core and expects you to build the guardrails. See related tools in our AI programming and development category.
