# Pi

A minimal, extensible, MIT-licensed agent harness for coding and beyond

**Verdict: conditional.** Pi is a deliberately minimal, MIT-licensed agent harness with TUI, print/JSON, RPC and SDK modes, 15+ model providers plus local servers, and built-in MCP. The deciding caveat is safety: Pi's docs say it does not ask for approval before every tool call and that tools run with your account's permissions. Right for tinkerers who build their own harness inside a VM or container; wrong for anyone wanting guardrails out of the box. Last verified 2026-10-03.

Pi is described by its maintainers as "a hardened, minimal, extensible agent harness that you can make your own" (Earendil, Oct 1, 2026) and "a minimal agent harness" (pi.dev). It runs in interactive TUI, print/JSON, RPC and SDK modes. The monorepo (github.com/earendil-works/pi, MIT, 111.8k stars at fetch) holds packages for a multi-provider LLM API (pi-ai), an agent runtime (pi-agent-core), the coding agent CLI (pi-coding-agent), a TUI library (pi-tui) and an application-composition runtime (chord). pi.dev lists "15+ providers, hundreds of models", naming Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, Cerebras, xAI, Hugging Face, Kimi For Coding, MiniMax, NVIDIA, OpenRouter and Ollama; local models connect through the llama.cpp router or compatible endpoints (Ollama, LM Studio, vLLM, SGLang). pi.dev says "MCP is built in, and Codemode lets agents compose tool calls in a JavaScript sandbox." Design stance per pi.dev: sub-agents and plan mode are not built in; users can spawn Pi instances via tmux, build their own with extensions, or install third-party packages; 50+ example extensions are listed. Sessions are tree-structured (navigate to earlier points and branch, all branches kept in one file), can be exported to HTML with `/export` or uploaded to GitHub gists with `/share`, and older messages are auto-summarized when approaching the context limit; skills are capability packages loaded on demand. Pi 1.0 (October 1, 2026) adds Codemode (native MCP plus non-LLM models), virtual models via extensions, deferred tool loading, Anthropic cache warming, mid-conversation system messages, a new TUI theme and full-screen mode by default; the vendor says "Hundreds of thousands of people around the world use Pi every week" (vendor claim). Pi Durable, an experimental npm package, is a separate add-on and not part of the verdict. Install: curl script from pi.dev, or npm/pnpm/bun. Security, per pi.dev/docs/latest/security: Pi can read, change and execute files "with the permissions of the account that started it" and "does not ask for approval before every tool call"; "After Pi starts, enabled tools still use the operating-system permissions of the Pi process"; "Extensions execute inside the Pi process"; "Project trust is not a complete startup boundary"; running Pi entirely inside a container, virtual machine or sandbox is "usually the strongest practical option". The page does not state that Pi sandboxes tool calls.

## Strengths

- Developers who want to shape their own agent harness through extensions, skills and prompt templates rather than accept fixed features
- Embedding an agent in your own software through RPC or the TypeScript SDK
- Mixing providers, subscriptions and local models (llama.cpp, Ollama, LM Studio, vLLM, SGLang) behind one interface
- Anyone who wants sub-agents or plan mode built in (pi.dev says neither is built in; they come from tmux, extensions or third-party packages)
- Running untrusted repos, extensions or unattended automation without a container or VM (the docs say Pi does not ask for approval before every tool call and tools use the account's permissions)
- Buyers who need a vendor-backed paid plan or support (no paid Pi plan appears on the pages we read)

## Pricing

- Model: open-source
- Open source: Free (MIT)

Website: https://pi.dev

## Sources

- [Pi site](https://pi.dev)
- [Pi 1.0 announcement (Earendil)](https://earendil.com/posts/pi-1-0/)
- [Pi security doc](https://pi.dev/docs/latest/security)
- [Pi docs index](https://pi.dev/docs/latest)
- [Pi models doc](https://pi.dev/docs/latest/models)
- [Pi GitHub repository](https://github.com/earendil-works/pi)

Canonical: https://neomanex.com/directory/pi