Pi 1.0

Original Article ↗ Hacker News Discussion ↗

Pi 1.0, a minimal terminal-based “harness” for coding agents, is drawing strong interest as an alternative to heavyweight tools like Claude Code, Codex and OpenCode, especially among developers who want fine-grained control over models, context usage and tooling. Commenters describe using Pi to orchestrate both cloud and local models, wire it into CI, editors, messaging systems and homelab setups, and extend it via custom plugins and MCP integration—while others prefer “batteries-included” variants like oh-my-pi or GUI multiplexers. Much of the debate centers on whether Pi is staying true to its minimalist philosophy as features like fullscreen TUI, codemode and cache-warming arrive, versus the practical need to match how modern models are trained and used.

How people actually use Pi

  • Many run Pi as their primary coding agent in a terminal, often alongside or instead of Claude Code/Codex/OpenCode.
  • Common workflows:
    • Plain Pi with just a few extensions (subagents, web search, review/plan helpers).
    • “Batteries-included” setups like oh-my-pi or lazypi for easier onboarding.
    • Remote and multi-device orchestration via daemons (e.g., Paseo, Pi Web, custom proxies) to resume sessions from desktop, laptop, or phone.
    • CI/CD and headless usage (code review, automated refactors, classification tasks).
    • Integrations with editors (Zed, VS Code plugins, Obsidian) and chat systems (XMPP, Slack, WhatsApp, Telegram).
    • Local-model setups using llama.cpp/llama-server, Qwen/Gemma, especially where context size and prefill cost matter.

Minimalism, tokens, and extensibility

  • Pi is praised for a tiny default system prompt, simple tool set (read/write/edit/bash), and visible context/token usage—good for local models and cost control.
  • Several compare token burn:
    • Pi (and some alternatives like 3code) reportedly use far fewer tokens than Claude Code; OpenCode somewhere in between.
  • Minimalism is framed as “minimal agent loop and tools” plus everything else as extensions.
  • Some feel recent additions (MCP support, codemode, cache warming, fullscreen TUI) erode the minimal ethos; maintainers argue features remain optional and are driven by how models are trained.

Plugins, ecosystems, and quality

  • Strong positive sentiment around Pi’s extension model: agents can self-modify, users can quickly “vibe code” custom skills, and per-project .pi folders enable profiles.
  • Counterpoint: plugin ecosystem is described as noisy, “vibe-coded,” and hard to vet for reliability and safety; several users prefer small, personally curated sets or batteries-included distributions (oh-my-pi) instead of hunting plugins.

TUI vs GUI and alternative harnesses

  • Split preferences:
    • TUI fans value working over SSH, integration with tmux/zellij/screen, and staying inside terminal workflows.
    • GUI/“agent multiplexer” fans cite better navigation, multi-session management, and richer visualization; alternatives mentioned include Orca, Conductor, Juggler, Cursor, Zed, various desktop apps.
  • Some worry full-screen TUI as default breaks the “just a CLI in my terminal scrollback” feel; others say it fixes scrollback bugs and flicker, especially on Windows.

Implementation language and performance debates

  • Debate over TypeScript/Node vs Rust/Go/C for a “fundamental” CLI tool:
    • Critics cite startup time, RAM usage, dependency stacks, and Node fragility.
    • Supporters emphasize rapid iteration, hot-reloadable extensions, wide JS/TS familiarity, and that latency is dominated by model calls anyway.
  • Several Rust/Go/Nim-based minimalist harnesses (hax, VT Code, 3code, others) are discussed as leaner alternatives.