Pi

Pi takes a different approach to AI agents. Where many agent products force you into their way of working, Pi is deliberately minimal and hackable — a harness you adapt to your workflows, not the other way around. In this guide we’ll cover what Pi is, who builds it, how it works, what you get, pricing, and who it’s for.

What is Pi?

Pi is a minimal AI agent harness from Earendil. It’s best described by its own tagline: “adapt Pi to your workflows, not the other way around.” Concretely, it packages four things into one open-source toolkit:

• A unified multi-provider LLM API (@earendil-works/pi-ai) — one interface across many AI providers.
• An agent runtime — the loop that lets an agent reason, use tools, and act.
• A TUI — a terminal user interface for working with agents.
• A coding agent CLI — a ready-made agent for coding tasks.

Because it’s open source and MIT-licensed, you can inspect, fork, and extend any part of it.

Who creates Pi?

Pi is built by Earendil Inc., an independent software company (earendil.com), with Mario Zechner (“badlogic”) as a lead maintainer. Mario is a well-known figure in the developer tools world, and Pi has grown rapidly since its repository was created in August 2025 — reaching roughly 82,000 GitHub stars by mid-2026, with multiple releases per week.

How Pi works

Pi’s design philosophy is modularity. Instead of one monolithic product, you get building blocks you combine:

1. Unified LLM API — one consistent interface across providers, so your agent code doesn’t lock you into a single model vendor.
2. Agent runtime — the core loop where an agent interprets instructions, calls tools, and produces results.
3. TUI and CLI — interfaces for driving agents from the terminal, including a ready-made coding agent.
4. Extensions — plug in capabilities, skills, prompt templates, and even themes to shape how Pi behaves.

Because it’s a harness rather than a fixed product, you configure and extend Pi to match your exact workflow.

Key features of Pi

• Minimal, hackable design — a harness you adapt, not an opinionated product.
• Unified multi-provider API — switch providers without rewriting agent code.
• Agent runtime — the loop, tools, and state for building agents.
• Terminal UI (TUI) — drive agents from your terminal.
• Coding agent CLI — a ready-to-use agent for coding tasks.
• Broad provider support — Anthropic, OpenAI, Google, Bedrock, Mistral, OpenRouter, xAI/Grok, Kimi, GitHub Copilot, and local llama.cpp.
• Extensible — extensions, skills, prompt templates, themes, RPC/SDK interfaces.
• MCP support — connect external tools via the Model Context Protocol.
• Supply-chain hardened — pinned dependencies, shrinkwrap, and npm audit in CI.

What can you use Pi for?

• Coding agents — the built-in CLI agent for editing code in your repos.
• Custom agent workflows — build agents adapted to your specific tasks.
• Multi-provider experimentation — try different models with one codebase.
• Local-first AI — connect local llama.cpp models alongside cloud providers.
• Learning agent frameworks — a clean, readable codebase to study and extend.

Pricing: what does Pi cost?

Pi is free and open source under the MIT license. There is no paid plan and no subscription. The only cost is what you pay your LLM provider directly — through your own API keys or subscriptions. This is a genuinely free toolkit.

Integrations

• LLM providers — OpenAI, Anthropic, Google Gemini, Amazon Bedrock, Mistral, OpenRouter, xAI, Kimi, GitHub Copilot, local llama.cpp.
• MCP — Model Context Protocol for external tools.
• SDK / RPC interfaces — build your own integrations.
• Extensions and skills — the ecosystem of Pi capabilities.

Who is Pi for?

Pi is aimed at developers who want control over their AI agents. It’s a strong fit for:

• Agent framework enthusiasts who like to build and adapt.
• Developers who want one codebase across multiple LLM providers.
• Terminal users who prefer minimal, hackable tools.
• Privacy-minded users who want local models as an option.

If you want a plug-and-play product with zero configuration, Pi isn’t that — it’s a toolkit that rewards developers who invest in it.

Advantages of Pi

• Free and open source — MIT, no paid plan.
• Minimal and adaptable — adapt it to your workflows.
• Multi-provider — one API across many LLMs.
• Fast-moving — active development with frequent releases.
• Local-first option — llama.cpp support for on-device AI.
• Strong supply-chain practices — hardened dependency management.

Limitations and honest considerations

• No built-in permission system — it runs with your user permissions, so you must sandbox it yourself for isolation.
• Requires technical comfort — configuration, code, and provider setup.
• Not a managed service — no hosting, support, or enterprise guarantees.
• Community scale — younger than some incumbent frameworks.

Security and privacy

Pi’s README explicitly documents that it has no built-in permission system and runs with the user’s permissions, recommending containerization for isolation. The project takes supply-chain security seriously — pinned dependencies, shrinkwrap, and npm audit in CI. For privacy, the multi-provider support including local llama.cpp gives you an on-device option. As with any agent framework, your API keys and prompts are your responsibility to protect.

Pi vs. other agent frameworks

	Pi	CrewAI	LangChain
Philosophy	Minimal, hackable harness	Role-based agent crews	Broad framework ecosystem
Style	Adapt to your workflows	Structured teams of agents	Component library
License	MIT	MIT (framework)	MIT
Best for	Tinkerers, terminal users	Team-based orchestration	General LLM app building

Each has its strengths; Pi’s niche is minimalism and adaptability.

Getting started with Pi

1. Visit pi.dev — read the docs and get the toolkit.
2. Install it — from npm or the GitHub repo (earendil-works/pi).
3. Configure a provider — add your LLM API key or local model.
4. Try the coding agent CLI — run it in a repo and give it a task.
5. Extend it — add skills, extensions, or your own agent definitions.

The bottom line

Pi is a refreshingly minimal take on AI agents: a free, MIT-licensed harness you adapt to your own workflows. With a unified multi-provider API, an agent runtime, a TUI, and a coding agent CLI, it gives developers the building blocks and the freedom to build what they want — at the cost of doing your own setup and sandboxing. If you’re a developer who’d rather adapt your tools than be adapted to them, Pi is worth a serious look.

To discover more developer tools and AI agents, browse our Developer Tools and AI Tools categories, or explore the full bla5k directory.

Official resources: the Pi website and the official GitHub repository.