Pi
Pi is a minimal, extensible agent harness for developers. It pairs a keyboard-driven TUI with skills, extensions, and prompt templates, supports print/JSON, RPC, and SDK modes, switches among many model providers, and stores navigable, shareable session trees for reproducible automation and coding.
Overview
In practice, you start Pi in the TUI, load project skills and prompts, and iterate. You can steer a running job, queue follow-ups, or jump within a session tree to branch work. For automation, use print mode for scripts or RPC/SDK for programmatic control.
How Pi works
Pi suits engineers who want tight control over agent behavior and integration points. It’s a strong fit for platform teams, toolsmiths, researchers, and maintainers building CLI tools, scripted automations, or editor-integrated flows. Teams that need to branch experiments, compare models, and ship reproducible, auditable runs benefit from tree-structured history. Organizations preferring composable primitives over heavy, opinionated assistants can evolve Pi to match internal workflows without abandoning existing tooling or security practices.
- Switch models and providers mid-session without restarting or losing conversation context.
- Navigate, branch, and share tree-structured sessions for reproducible explorations and workflows.
- Generate shell scripts in print mode, or stream JSON events for automation.
- Load per-project AGENTS.md instructions, optional SYSTEM.md, skills, and prompt templates.
- Extend the harness with TypeScript modules that add tools, UI, and workflows.
Highlights
What users build with Pi
Install via a one-line script for Unix or PowerShell, or add the global package using npm, pnpm, or bun. Configure provider credentials with API keys or OAuth, then open the TUI to create a project. Add AGENTS.md and optional SYSTEM.md, and load skills or extensions as needed. Use /model to switch models, /tree to navigate history, and -p to print scripts. For automation, stream events in JSON mode, talk over the RPC protocol, or embed via the SDK. Reload after edits and share branches when ready.
A small, composable harness that adapts to your workflow instead of enforcing one.
Getting started
Pi’s differentiator is scope control: a small core with powerful primitives instead of heavyweight features. It omits sub-agents, plan mode, permission popups, and MCP by default, yet enables them via extensions when warranted. With a fast TUI, shareable session trees, and multi-provider switching, teams gain precise, portable workflows that fit existing tooling.
Open the tool and review its core product experience.
Create your account or access your existing workspace.
Use your own task to judge speed, quality, and fit.
Check similar AI tools before making a final decision.
Pi
Earendil Inc. leads development of Pi with an open community of contributors. The team focuses on pragmatic, developer-first tooling that favors composable primitives over sealed products. They maintain extensive examples, documentation, and active GitHub and Discord channels to help engineers adopt, customize, and share Pi-based workflows.
Pi is a minimal agent harness that lets developers tailor how an AI assistant behaves and integrates into their tools. A keyboard-first TUI orchestrates runs, tools, and messages, while extensions, skills, and prompt templates provide composable capabilities. Pi runs in four modes—interactive TUI, print/JSON for script generation and event streams, an RPC protocol for stdin/stdout integrations, and an embeddable SDK—so teams can automate tasks or embed the agent in existing systems. Sessions are stored as tree-structured histories that support branching, rewinding, export, and sharing. Pi connects to many providers and models via API keys or OAuth and allows model switching mid-session. Context is explicitly engineered using AGENTS.md, optional SYSTEM.md, dynamic injection, and compaction strategies, enabling per-project control without bloating the prompt.
Visit Pi
Comments (0)
No Comments Found