Pi vs. Claude Code

Pi (@mariozechner/pi-coding-agent) is a minimal, aggressively-extensible terminal coding harness — an alternative to Claude Code. Here’s how they relate.

Similarities

Both are terminal-based coding agents you cd into a project and talk to.
Both load AGENTS.md / CLAUDE.md context files (pi reads either) from ~, parent dirs, and cwd.
Both can use Claude models — pi works with your Anthropic Claude Pro/Max subscription via /login, or ANTHROPIC_API_KEY.
Both ship with the same primitive toolset: read, write, edit, bash (pi adds grep, find, ls).
Both support Agent Skills (the standard from agentskills.io).

Key Differences (pi’s philosophy)

Pi deliberately omits features Claude Code bakes in, expecting you to add them via extensions or pi packages instead:

Feature	Claude Code	Pi
Sub-agents	Built-in	Build via extension, or use tmux
Plan mode	Built-in	Write to a file, or extension
Permission popups	Built-in	Extension, or run in container
Built-in to-dos	Yes	Use `TODO.md`
Background bash	Yes	Use tmux
MCP	Yes	Skipped — use CLI tools w/ READMEs (Skills), or extension
Provider lock-in	Anthropic-only	25+ providers (OpenAI, Gemini, Bedrock, OpenRouter, local, etc.)
Sessions	Linear	Tree-structured (`/tree`, `/fork`, `/clone`, branching)
Customization	Limited	TS extensions, themes, prompt templates, sharable npm/git packages
RPC / SDK	No	Yes (`--mode rpc`, full TS SDK)

Practical “switching from Claude Code” tips

Install: npm install -g @mariozechner/pi-coding-agent
Use your Claude subscription: run pi, then /login → Anthropic.
Your CLAUDE.md files just work — pi reads them automatically.
Pick the Claude model: Ctrl+L (or /model) to switch between Sonnet/Opus/Haiku, Shift+Tab cycles thinking levels.
Branching instead of “undo”: use /tree (Esc Esc) to jump back to any prior point — much more powerful than Claude Code’s linear history.
Add what you miss: if you want Claude Code-style plan mode, sub-agents, or permission gates, install a pi-package (pi install npm:...) or write a small TS extension. The README explicitly says you can “make pi look like Claude Code” via extensions.

When to pick which

Claude Code: want batteries-included, Anthropic-only, opinionated workflow.
Pi: want a minimal core, multi-provider freedom, and full control to shape the agent (extensions, skills, RPC/SDK embedding, OSS session sharing).

For deeper comparison, see the philosophy section in the pi README and the linked blog post.