Pi
Open SourceThere are many agent harnesses — but this one is yours.
Published 27 September 2026
Scores
Popularity4/5
90K+ GitHub stars reflects strong adoption within its technical, extensibility-focused niche, ahead of most other independent (non single-vendor) agents, though it targets a narrower audience than default-recommendation tools like Claude Code or OpenCode.
Learning Curve4/5
The minimal four-tool core is quick to start, but getting real value requires writing or installing extensions, skills, and prompt templates — a steeper ramp than a fully-featured agent that works well out of the box.
Flexibility5/5
Built around extensibility primitives rather than fixed features — TypeScript extensions, skills, prompt templates, themes, RPC/SDK embedding, and 15+ interchangeable providers make it one of the most open-ended harnesses in the subcategory.
Performance4/5
The minimal system prompt and tool set keep token overhead low and session behavior predictable, and output quality scales directly with whichever model is configured, from frontier to cheap open-weight options.
Portability5/5
MIT-licensed, provider-agnostic across 15+ providers, and embeddable via SDK/RPC with no default vendor lock-in — among the most portable options in the subcategory.
About Pi
Pi (pi.dev) is an open-source, MIT-licensed terminal coding agent harness created by Armin Ronacher and Mario Zechner, distinguished by a deliberately minimal, token-efficient core: a ~200-token system prompt and exactly four built-in tools (read, write, edit, bash). Rather than growing a large built-in feature surface, Pi is designed around primitives: everything beyond the core loop is meant to be added through its extension system, which includes TypeScript extensions with tool and command access, "skills" (bundled capability packages with instructions and tools, loaded on demand to preserve prompt cache), reusable Markdown prompt templates, custom themes, and installable packages via npm or git. 50+ example extensions ship with the project as a starting point.
The agent runs as an interactive TUI, plus print/JSON, RPC, and SDK modes for embedding into other tools or automations. Sessions are stored as a tree structure supporting branching and sharing, models can be switched mid-session with a /model command, and the agent supports real-time steering, sending follow-up instructions while it's still working rather than waiting for a turn to end. Context engineering features include project-level instruction files (AGENTS.md, SYSTEM.md) and auto-compacting message history to manage long sessions.
Pi is fully bring-your-own-key and provider-agnostic, supporting 15+ providers including Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, Cerebras, xAI, Hugging Face, Kimi, MiniMax, NVIDIA, OpenRouter, and Ollama, authenticated via API keys or OAuth. Because it imposes no default model or markup, cost is entirely a function of the provider and model chosen, from frontier models down to cheap open-weight options.
Key Features
- Minimal core: ~200-token system prompt and just four built-in tools (read, write, edit, bash)
- Extension system: TypeScript extensions, skills, prompt templates, themes, and installable packages
- 50+ example extensions ship with the project
- Interactive TUI plus print/JSON, RPC, and SDK modes for embedding
- Tree-structured session history with branching and sharing
- Mid-session model switching via `/model`, plus real-time steering during execution
- Provider-agnostic across 15+ providers via API key or OAuth
- Auto-compacting message history and project-level instruction files (AGENTS.md, SYSTEM.md)
Pros
- Minimal, token-efficient core keeps prompt-cache costs low and behavior predictable
- Primitives-first extension model gives serious developers deep control over agent behavior rather than a fixed feature set
- Genuinely provider-agnostic across 15+ providers, from frontier to cheap open-weight models
- RPC and SDK modes make it straightforward to embed the agent into other tools and automations
- Real-time steering and mid-session model switching are more flexible than most single-shot agent loops
Cons
- The minimal-by-default philosophy means most useful behavior has to be added via extensions rather than working out of the box
- Narrower, more technical/extensibility-first audience than default-recommendation agents like Claude Code or OpenCode
- No hosted/managed billing option — cost management is entirely on the user across whichever providers they configure
- Smaller ecosystem of pre-built extensions than the plugin marketplaces of larger competitors
Pi Pricing
Open Source- · Free, MIT-licensed CLI/SDK agent harness — bring your own API key to any of 15+ providers
- · Local model support via Ollama
- · Usage-based cost depends entirely on the provider and model chosen
Tech Stacks with Pi
Herdr Dev Stack
DeveloperHerdr keeps several coding agents running in persistent terminal sessions on your laptop or a server, with your pick of agents and models and GitHub or GitLab for version control.
Paseo Dev Stack
DeveloperPaseo runs coding agents on a machine you control, a laptop, home server, or VPS, and lets you steer them from desktop, web, phone, or terminal, with your pick of agents, models, and git host.
Tools Related to Pi
Alternatives to Pi(12)
Pi is the deliberately minimal upstream project Oh My Pi forked from; Oh My Pi takes the opposite design philosophy, adding 31 built-in tools, LSP/DAP integration, and browser automation. The closest comparison in the subcategory.
Pi is a minimal, model-agnostic terminal harness with four core tools; Qwen Code is a fuller terminal agent tuned for Qwen3-Coder. Pi for building your own agent, Qwen Code for a ready Qwen setup.
Kilo Code is a full-featured agent with specialized modes across editors and cloud; Pi is a minimal terminal harness with four core tools and a tiny system prompt. Kilo Code works out of the box, Pi is a base to build on.
Grok Build ships parallel sub-agents in git worktrees out of the box; Pi is a minimal four-tool terminal harness you extend yourself in TypeScript. Grok Build is batteries-included, Pi deliberately bare.
GitHub Copilot is a broad, managed product from completions to cloud agents; Pi is a minimal open-source terminal harness built to be extended in TypeScript. Copilot suits teams on GitHub, Pi developers who want to shape their own agent.
Both are minimal, hackable terminal harnesses: Pi ships four core tools and a tiny system prompt extended in TypeScript, DeepSeek's harness is built from swappable Cordis plugins around DeepSeek models. Pi is model-agnostic, DeepSeek's harness vendor-tuned.
Learning Resources
No resources yet — check back soon.